Skip to content

X
X2Paddle

PaddlePaddle / X2Paddle
大约 2 年 前同步成功

通知 329
Star 698
Fork 167
X
X2Paddle
提交 cf7f9b88 编写于 作者: S sunyanfang01
浏览文件
操作

for 0.8.4

上级 85e44863
隐藏空白更改
内联 并排
Showing with 26016 addition and 0 deletion
x2paddle/x2paddle/__init__.py 0 → 100644
浏览文件 @ cf7f9b88
__version__ = "0.8.4"
from .core.program import PaddleProgram
program = PaddleProgram()
name_counter = dict()
def gen_name(op_name, var_name):
name = "{}.{}".format(op_name, var_name)
if name not in name_counter:
name_counter[name] = 0
else:
name_counter[name] += 1
name = name + "." + str(name_counter[name])
return name
x2paddle/x2paddle/convert.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from six import text_type as _text_type
import argparse
import sys
def arg_parser():
parser = argparse.ArgumentParser()
parser.add_argument(
"--model",
"-m",
type=_text_type,
default=None,
help="define model file path for tensorflow or onnx")
parser.add_argument(
"--prototxt",
"-p",
type=_text_type,
default=None,
help="prototxt file of caffe model")
parser.add_argument(
"--weight",
"-w",
type=_text_type,
default=None,
help="weight file of caffe model")
parser.add_argument(
"--save_dir",
"-s",
type=_text_type,
default=None,
help="path to save translated model")
parser.add_argument(
"--framework",
"-f",
type=_text_type,
default=None,
help="define which deeplearning framework(tensorflow/caffe/onnx/paddle2onnx)"
)
parser.add_argument(
"--caffe_proto",
"-c",
type=_text_type,
default=None,
help="optional: the .py file compiled by caffe proto file of caffe model"
)
parser.add_argument(
"--version",
"-v",
action="store_true",
default=False,
help="get version of x2paddle")
parser.add_argument(
"--without_data_format_optimization",
"-wo",
type=_text_type,
default="True",
help="tf model conversion without data format optimization")
parser.add_argument(
"--define_input_shape",
"-d",
action="store_true",
default=False,
help="define input shape for tf model")
parser.add_argument(
"--onnx_opset",
"-oo",
type=int,
default=10,
help="when paddle2onnx set onnx opset version to export")
parser.add_argument(
"--params_merge",
"-pm",
action="store_true",
default=False,
help="define whether merge the params")
return parser
def tf2paddle(model_path,
save_dir,
without_data_format_optimization,
define_input_shape=False,
params_merge=False):
# check tensorflow installation and version
try:
import os
os.environ["TF_CPP_MIN_LOG_LEVEL"] = '3'
import tensorflow as tf
version = tf.__version__
if version >= '2.0.0' or version < '1.0.0':
print(
"[ERROR] 1.0.0<=tensorflow<2.0.0 is required, and v1.14.0 is recommended"
)
return
except:
print(
"[ERROR] Tensorflow is not installed, use \"pip install tensorflow\"."
)
return
from x2paddle.decoder.tf_decoder import TFDecoder
from x2paddle.op_mapper.tf_op_mapper import TFOpMapper
from x2paddle.op_mapper.tf_op_mapper_nhwc import TFOpMapperNHWC
from x2paddle.optimizer.tf_optimizer import TFOptimizer
print("Now translating model from tensorflow to paddle.")
model = TFDecoder(model_path, define_input_shape=define_input_shape)
if not without_data_format_optimization:
mapper = TFOpMapper(model)
optimizer = TFOptimizer(mapper)
# neccesary optimization
optimizer.delete_redundance_code()
# optimizer below is experimental
optimizer.optimize_elementwise_op()
optimizer.merge_activation()
optimizer.merge_bias()
optimizer.optimize_sub_graph()
# optimizer.merge_batch_norm()
# optimizer.merge_prelu()
else:
mapper = TFOpMapperNHWC(model)
optimizer = TFOptimizer(mapper)
optimizer.delete_redundance_code()
optimizer.strip_graph()
optimizer.merge_activation()
optimizer.merge_bias()
optimizer.make_nchw_input_output()
optimizer.remove_transpose()
mapper.save_inference_model(save_dir, params_merge)
def caffe2paddle(proto, weight, save_dir, caffe_proto, params_merge=False):
from x2paddle.decoder.caffe_decoder import CaffeDecoder
from x2paddle.op_mapper.caffe_op_mapper import CaffeOpMapper
from x2paddle.optimizer.caffe_optimizer import CaffeOptimizer
import google.protobuf as gpb
ver_part = gpb.__version__.split('.')
version_satisfy = False
if (int(ver_part[0]) == 3 and int(ver_part[1]) >= 6) \
or (int(ver_part[0]) > 3):
version_satisfy = True
assert version_satisfy, '[ERROR] google.protobuf >= 3.6.0 is required'
print("Now translating model from caffe to paddle.")
model = CaffeDecoder(proto, weight, caffe_proto)
mapper = CaffeOpMapper(model)
optimizer = CaffeOptimizer(mapper)
optimizer.merge_bn_scale()
optimizer.merge_op_activation()
mapper.save_inference_model(save_dir, params_merge)
def onnx2paddle(model_path, save_dir, params_merge=False):
# check onnx installation and version
try:
import onnx
version = onnx.version.version
if version < '1.6.0':
print("[ERROR] onnx>=1.6.0 is required")
return
except:
print("[ERROR] onnx is not installed, use \"pip install onnx==1.6.0\".")
return
print("Now translating model from onnx to paddle.")
from x2paddle.op_mapper.onnx2paddle.onnx_op_mapper import ONNXOpMapper
from x2paddle.decoder.onnx_decoder import ONNXDecoder
from x2paddle.optimizer.onnx_optimizer import ONNXOptimizer
model = ONNXDecoder(model_path)
mapper = ONNXOpMapper(model)
print("Model optimizing ...")
optimizer = ONNXOptimizer(mapper)
print("Model optimized.")
print("Paddle model and code generating ...")
mapper.save_inference_model(save_dir, params_merge)
print("Paddle model and code generated.")
def paddle2onnx(model_path, save_dir, opset_version=10):
from x2paddle.decoder.paddle_decoder import PaddleDecoder
from x2paddle.op_mapper.paddle2onnx.paddle_op_mapper import PaddleOpMapper
import paddle.fluid as fluid
model = PaddleDecoder(model_path, '__model__', '__params__')
mapper = PaddleOpMapper()
mapper.convert(
model.program,
save_dir,
scope=fluid.global_scope(),
opset_version=opset_version)
def main():
if len(sys.argv) < 2:
print("Use \"x2paddle -h\" to print the help information")
print("For more information, please follow our github repo below:)")
print("\nGithub: https://github.com/PaddlePaddle/X2Paddle.git\n")
return
parser = arg_parser()
args = parser.parse_args()
if args.version:
import x2paddle
print("x2paddle-{} with python>=3.5, paddlepaddle>=1.6.0\n".format(
x2paddle.__version__))
return
assert args.framework is not None, "--framework is not defined(support tensorflow/caffe/onnx)"
assert args.save_dir is not None, "--save_dir is not defined"
try:
import paddle
v0, v1, v2 = paddle.__version__.split('.')
print("paddle.__version__ = {}".format(paddle.__version__))
if v0 == '0' and v1 == '0' and v2 == '0':
print("[WARNING] You are use develop version of paddlepaddle")
elif int(v0) != 1 or int(v1) < 6:
print("[ERROR] paddlepaddle>=1.6.0 is required")
return
except:
print(
"[ERROR] paddlepaddle not installed, use \"pip install paddlepaddle\""
)
if args.framework == "tensorflow":
assert args.model is not None, "--model should be defined while translating tensorflow model"
assert args.without_data_format_optimization in [
"True", "False"
], "--the param without_data_format_optimization should be defined True or False"
define_input_shape = False
params_merge = False
without_data_format_optimization = True if args.without_data_format_optimization == "True" else False
if args.define_input_shape:
define_input_shape = True
if args.params_merge:
params_merge = True
tf2paddle(args.model, args.save_dir, without_data_format_optimization,
define_input_shape, params_merge)
elif args.framework == "caffe":
assert args.prototxt is not None and args.weight is not None, "--prototxt and --weight should be defined while translating caffe model"
params_merge = False
if args.params_merge:
params_merge = True
caffe2paddle(args.prototxt, args.weight, args.save_dir,
args.caffe_proto, params_merge)
elif args.framework == "onnx":
assert args.model is not None, "--model should be defined while translating onnx model"
params_merge = False
if args.params_merge:
params_merge = True
onnx2paddle(args.model, args.save_dir, params_merge)
elif args.framework == "paddle2onnx":
assert args.model is not None, "--model should be defined while translating paddle model to onnx"
paddle2onnx(args.model, args.save_dir, opset_version=args.onnx_opset)
else:
raise Exception(
"--framework only support tensorflow/caffe/onnx/paddle2onnx now")
if __name__ == "__main__":
main()
x2paddle/x2paddle/core/fluid_code.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from x2paddle.core.graph import GraphNode
from x2paddle.core.util import *
import collections
import six
class Layer(object):
def __init__(self):
self.op = None
self.param_attr = dict()
self.inputs = dict()
self.output = None
self.is_custom_layer = False
self.use_fluid = False
def get_code(self):
layer_code = ""
if self.output is not None:
if isinstance(self.output, six.string_types):
layer_code = self.output + " = "
else:
layer_code = self.output.layer_name + " = "
if self.is_custom_layer:
layer_code = layer_code + self.op + "("
elif self.op == "=":
layer_code = layer_code
elif self.use_fluid:
layer_code = layer_code + "fluid." + self.op + "("
else:
layer_code = layer_code + "fluid.layers." + self.op + "("
if isinstance(self.inputs, list):
in_list = "["
for input in self.inputs:
if isinstance(input, GraphNode):
if hasattr(input, "index"):
in_list += (
input.layer_name + "[{}]".format(input.index) + ", "
)
else:
in_list += (input.layer_name + ", ")
elif isinstance(input, six.string_types):
in_list += (input + ", ")
else:
raise Exception(
"Element of inputs should GraphNode or String")
in_list = in_list.strip(", ") + "], "
layer_code += in_list
elif isinstance(self.inputs, dict):
inputs = collections.OrderedDict(self.inputs)
for key, input in inputs.items():
if isinstance(input, GraphNode):
if hasattr(input, "index"):
layer_code = layer_code + key + "={}, ".format(
input.layer_name + "[{}]".format(input.index))
else:
layer_code = layer_code + key + "={}, ".format(
input.layer_name)
else:
layer_code = layer_code + key + "={}, ".format(input)
elif isinstance(self.inputs, GraphNode):
if hasattr(self.inputs, "index"):
layer_code += (
self.inputs.layer_name + "[{}]".format(self.inputs.index))
else:
layer_code += (self.inputs.layer_name)
if self.op != "=":
layer_code += ", "
elif isinstance(self.inputs, six.string_types):
layer_code += (self.inputs)
if self.op != "=":
layer_code += ", "
else:
raise Exception("Unknown type of inputs.")
param_attr = collections.OrderedDict(self.param_attr)
for key, value in param_attr.items():
if '\n' in str(value):
value = string(str(value).replace('\n', ','))
if str(key) == 'attr':
value = 'ParamAttr(' + str(value) + ')'
layer_code = layer_code + key + "={}, ".format(value)
layer_code = layer_code.strip(", ")
if self.op != "=":
layer_code += ")"
return layer_code
class FluidCode(object):
def __init__(self):
self.layers = list()
def add_layer(self,
op,
inputs,
output,
param_attr=None,
use_fluid=False,
is_custom_layer=False):
layer = Layer()
layer.op = op
layer.use_fluid = use_fluid
layer.is_custom_layer = is_custom_layer
if inputs is not None:
layer.inputs = inputs
layer.output = output
if param_attr is not None:
layer.param_attr = param_attr
self.layers.append(layer)
def add_note(self, note):
# note should be string
self.layers.append(note)
def clear(self):
self.layers = list()
def gen_codes(self):
codes = list()
for layer in self.layers:
if isinstance(layer, Layer):
codes.append(layer.get_code())
elif isinstance(layer, six.string_types):
codes.append(layer)
return codes
x2paddle/x2paddle/core/graph.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import print_function
from __future__ import division
import collections
import copy as cp
class GraphNode(object):
def __init__(self, layer, layer_name=None):
self.inputs = list()
self.outputs = list()
self.layer = layer
assert layer_name is not None, "layer_name for GraphNode should not be None"
self.layer_name = layer_name
def __hash__(self):
return hash(self.layer.name)
def __eq__(self, other):
if self.layer.name == other.layer.name:
return True
return False
class Graph(object):
def __init__(self, model):
self.node_map = collections.OrderedDict()
self.input_nodes = list()
self.output_nodes = list()
self.topo_sort = list()
self.model = model
def build(self):
self.get_input_nodes()
self.get_output_nodes()
self.get_topo_sort()
def get_input_nodes(self):
for name, node in self.node_map.items():
name = name.replace('/', '_').replace('-', '_')
if len(node.inputs) == 0:
self.input_nodes.append(name)
def get_output_nodes(self):
for name, node in self.node_map.items():
name = name.replace('/', '_').replace('-', '_')
if len(node.outputs) == 0:
self.output_nodes.append(name)
def get_topo_sort(self):
num_inputs = dict()
for name, node in self.node_map.items():
num_inputs[name] = len(node.inputs)
self.topo_sort = self.input_nodes[:]
idx = 0
while idx < len(self.topo_sort):
current_node = self.node_map[self.topo_sort[idx]]
for node in current_node.outputs:
num_inputs[node] -= 1
if num_inputs[node] == 0:
self.topo_sort.append(node)
idx += 1
def get_node(self, name, copy=False):
if name not in self.node_map:
if name.split(':')[0] in self.node_map:
name_prefix, idx = name.split(':')
if copy:
node = cp.copy(self.node_map[name_prefix])
else:
node = self.node_map[name_prefix]
node.index = int(idx)
return node
else:
return None
else:
if copy:
node = cp.copy(self.node_map[name])
else:
node = self.node_map[name]
return node
def connect(self, src, dst):
if dst not in self.node_map:
raise Exception("node[{}] not in graph".format(dst))
self.node_map[dst].inputs.append(src)
self.node_map[src].outputs.append(dst)
x2paddle/x2paddle/core/op_mapper.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import paddle.fluid as fluid
from paddle.fluid.proto import framework_pb2
from x2paddle.core.util import *
import inspect
import os
def export_paddle_param(param, param_name, dir):
dtype_map = {
"int16": [framework_pb2.VarType.INT16, 'h'],
"int32": [framework_pb2.VarType.INT32, 'i'],
"int64": [framework_pb2.VarType.INT64, 'q'],
"float16": [framework_pb2.VarType.FP16, 'e'],
"float32": [framework_pb2.VarType.FP32, 'f'],
"float64": [framework_pb2.VarType.FP64, 'd'],
"bool": [framework_pb2.VarType.BOOL, None]
}
shape = param.shape
if str(param.dtype) in ['uint8', 'uint_8', 'bool']:
param = param.astype('int64')
if len(shape) == 0:
assert param.size == 1, "Unexpected situation happend!"
shape = [1]
assert str(
param.dtype) in dtype_map, "Unknown dtype {} of params: {}.".format(
str(param.dtype), param_name)
fp = open(os.path.join(dir, param_name), 'wb')
numpy.array([0], dtype='int32').tofile(fp)
numpy.array([0], dtype='int64').tofile(fp)
numpy.array([0], dtype='int32').tofile(fp)
tensor_desc = framework_pb2.VarType.TensorDesc()
tensor_desc.data_type = dtype_map[str(param.dtype)][0]
tensor_desc.dims.extend(shape)
desc_size = tensor_desc.ByteSize()
numpy.array([desc_size], dtype='int32').tofile(fp)
fp.write(tensor_desc.SerializeToString())
param.tofile(fp)
fp.close()
# This func will copy to generate code file
def run_net(param_dir="./"):
import os
inputs, outputs = x2paddle_net()
ops = fluid.default_main_program().global_block().ops
used_vars = list()
for op in ops:
used_vars += op.input_arg_names
tmp = list()
for input in inputs:
if isinstance(input, list):
for ipt in input:
if ipt.name not in used_vars:
continue
tmp.append(ipt)
else:
if input.name not in used_vars:
continue
tmp.append(input)
inputs = tmp
for i, out in enumerate(outputs):
if isinstance(out, list):
for out_part in out:
outputs.append(out_part)
del outputs[i]
exe = fluid.Executor(fluid.CPUPlace())
exe.run(fluid.default_startup_program())
def if_exist(var):
b = os.path.exists(os.path.join(param_dir, var.name))
return b
fluid.io.load_vars(
exe, param_dir, fluid.default_main_program(), predicate=if_exist)
class OpMapper(object):
def __init__(self):
self.paddle_codes = ""
self.tab = " "
self.net_code = list()
self.weights = dict()
self.inputs = list()
self.outputs = list()
def op_checker(self):
unsupported_ops = set()
for node_name in self.graph.topo_sort:
node = self.graph.get_node(node_name)
op = node.layer_type
if not hasattr(self, op):
unsupported_ops.add(op)
if len(unsupported_ops) == 0:
return True
else:
print("There are {} ops not supported yet, list as below".format(
len(unsupported_ops)))
for op in unsupported_ops:
print(op)
return False
def add_codes(self, codes, indent=0):
if isinstance(codes, list):
for code in codes:
self.paddle_codes += (
self.tab * indent + code.strip('\n') + '\n')
elif isinstance(codes, str):
self.paddle_codes += (self.tab * indent + codes.strip('\n') + '\n')
else:
raise Exception("Unknown type of codes")
def add_heads(self):
self.add_codes("from paddle.fluid.initializer import Constant")
self.add_codes("from paddle.fluid.param_attr import ParamAttr")
self.add_codes("import paddle.fluid as fluid")
self.add_codes("")
def save_inference_model(self, save_dir, params_merge):
self.save_python_model(save_dir)
import sys
import paddle.fluid as fluid
py_code_dir = os.path.join(save_dir, "model_with_code")
sys.path.append(py_code_dir)
import model
try:
inputs, outputs = model.x2paddle_net()
ops = fluid.default_main_program().global_block().ops
used_vars = list()
for op in ops:
used_vars += op.input_arg_names
for i, out in enumerate(outputs):
if isinstance(out, list):
for out_part in out:
outputs.append(out_part)
del outputs[i]
input_names = list()
for input in inputs:
if isinstance(input, list):
for ipt in input:
if ipt.name not in used_vars:
continue
input_names.append(ipt.name)
else:
if input.name not in used_vars:
continue
input_names.append(input.name)
exe = fluid.Executor(fluid.CPUPlace())
exe.run(fluid.default_startup_program())
def if_exist(var):
b = os.path.exists(
os.path.join(os.path.join(py_code_dir, var.name)))
return b
fluid.io.load_vars(
exe,
py_code_dir,
fluid.default_main_program(),
predicate=if_exist)
if params_merge:
fluid.io.save_inference_model(
dirname=os.path.join(save_dir, "inference_model"),
feeded_var_names=input_names,
target_vars=outputs,
executor=exe,
params_filename="__params__")
else:
fluid.io.save_inference_model(
dirname=os.path.join(save_dir, "inference_model"),
feeded_var_names=input_names,
target_vars=outputs,
executor=exe,
params_filename=None)
except:
raise Exception(
"Paddle code was saved in {}/model.py, but seems there's wrong exist, please check model.py manually."
.format(py_code_dir))
def save_python_model(self, save_dir):
if not os.path.exists(save_dir):
os.makedirs(save_dir)
py_code_dir = os.path.join(save_dir, "model_with_code")
if not os.path.exists(py_code_dir):
os.makedirs(py_code_dir)
for name, param in self.weights.items():
export_paddle_param(param, name, py_code_dir)
self.add_heads()
if hasattr(self, "used_custom_layers"):
for _, layer_code in self.used_custom_layers.items():
self.add_codes(layer_code, 0)
self.add_codes("", 0)
self.add_codes("\ndef x2paddle_net():", 0)
for i in range(len(self.graph.topo_sort)):
node_name = self.graph.topo_sort[i]
node = self.graph.get_node(node_name)
if node is None:
continue
if len(node.fluid_code.layers) == 0:
continue
self.add_codes(node.fluid_code.gen_codes(), 1)
self.add_codes("", 0)
input_str = "["
for name in self.graph.input_nodes:
input_str += (name + ", ")
input_str = input_str.strip(", ") + "]"
output_str = "["
for name in self.graph.output_nodes:
output_str += (name + ", ")
output_str = output_str.strip(", ") + "]"
return_code = "return {}, {}".format(input_str, output_str)
self.add_codes(return_code, 1)
self.add_codes("", 0)
self.add_codes(inspect.getsourcelines(run_net)[0])
fp = open(os.path.join(py_code_dir, "model.py"), 'w')
fp.write(self.paddle_codes)
fp.close()
x2paddle/x2paddle/core/program.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import print_function
from __future__ import division
import collections
import os
class PaddleLayer(object):
def __init__(self, kernel, inputs, outputs, **kwargs):
assert isinstance(
inputs,
dict), "parameter 'inputs' for PaddleLayer should be type of dict"
assert isinstance(
outputs,
list), "parameter, 'outputs' for PaddleLayer should be type of list"
self.kernel = kernel
self.inputs = inputs
self.outputs = outputs
self.attrs = kwargs
class PaddleProgram(object):
def __init__(self):
self.layers = list()
self.edges_out = dict()
self.edges_in = dict()
self.inputs = list()
self.outputs = list()
self.parameters = dict()
def add_layer(self, kernel, inputs, outputs, **kwargs):
layer = PaddleLayer(kernel, inputs, outputs, **kwargs)
self.layers.append(layer)
def build(self):
outputs = dict()
for i in range(len(self.layers)):
layer = self.layers[i]
for output in layer.outputs:
outputs[output] = i
for k, v in layer.inputs.items():
assert v in outputs, "Couldn't find {} in previous layers, the layers should be make by topological sort".format(
v)
in_layer_index = outputs[v]
if in_layer_index not in self.edges_out:
self.edges_out[in_layer_index] = list()
self.edges_out[in_layer_index].append(i)
if i not in self.edges_in:
self.edges_in[i] = list()
self.edges_in[i].append(in_layer_index)
def get_layer_outputs(self, i):
return self.edges_out[i]
def get_layer_inputs(self, i):
return self.edges_in[i]
def gen_code(self, code_dir):
def write_code(f, code_list, indent=0):
indent_blank = " " * indent
for code_line in code_list:
if code_line.strip() == "":
f.write('\n')
else:
f.write(indent_blank + code_line + '\n')
f = open(os.path.join(code_dir, 'model.py'), 'w')
write_code(
f, [
"from paddle.fluid.initializer import Constant",
"from paddle.fluid.param_attr import ParamAttr",
"import paddle.fluid as fluid"
"", "def x2paddle_net():"
],
indent=0)
for i, layer in enumerate(self.layers):
if self.edges_in.get(i, 0) == 0 and self.edges_out.get(i, 0) == 0:
continue
line = ""
if len(layer.outputs) == 1:
line = layer.outputs[0]
else:
for output in layer.outputs:
line += "{}, ".format(output)
line = line.strip(", ")
line += " = {}(".format(layer.kernel)
for k, v in layer.inputs.items():
line += "{}={}, ".format(k, v)
for k, v in layer.attrs.items():
line += "{}={}, ".format(k, v)
line = line.strip(", ")
line += ")"
write_code(f, [line], indent=1)
f.close()
def gen_parameters(self, code_dir):
pass
def gen_inference_model(self, model_dir):
pass
x2paddle/x2paddle/core/util.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy
import math
import os
def string(param):
return "\'{}\'".format(param)
x2paddle/x2paddle/decoder/caffe.proto 0 → 100644
浏览文件 @ cf7f9b88
syntax = "proto2";
package caffe;
// Specifies the shape (dimensions) of a Blob.
message BlobShape {
repeated int64 dim = 1 [packed = true];
}
message BlobProto {
optional BlobShape shape = 7;
repeated float data = 5 [packed = true];
repeated float diff = 6 [packed = true];
repeated double double_data = 8 [packed = true];
repeated double double_diff = 9 [packed = true];
// 4D dimensions -- deprecated. Use "shape" instead.
optional int32 num = 1 [default = 0];
optional int32 channels = 2 [default = 0];
optional int32 height = 3 [default = 0];
optional int32 width = 4 [default = 0];
}
// The BlobProtoVector is simply a way to pass multiple blobproto instances
// around.
message BlobProtoVector {
repeated BlobProto blobs = 1;
}
message Datum {
optional int32 channels = 1;
optional int32 height = 2;
optional int32 width = 3;
// the actual image data, in bytes
optional bytes data = 4;
optional int32 label = 5;
// Optionally, the datum could also hold float data.
repeated float float_data = 6;
// If true data contains an encoded image that need to be decoded
optional bool encoded = 7 [default = false];
}
// The label (display) name and label id.
message LabelMapItem {
// Both name and label are required.
optional string name = 1;
optional int32 label = 2;
// display_name is optional.
optional string display_name = 3;
}
message LabelMap {
repeated LabelMapItem item = 1;
}
// Sample a bbox in the normalized space [0, 1] with provided constraints.
message Sampler {
// Minimum scale of the sampled bbox.
optional float min_scale = 1 [default = 1.];
// Maximum scale of the sampled bbox.
optional float max_scale = 2 [default = 1.];
// Minimum aspect ratio of the sampled bbox.
optional float min_aspect_ratio = 3 [default = 1.];
// Maximum aspect ratio of the sampled bbox.
optional float max_aspect_ratio = 4 [default = 1.];
}
// Constraints for selecting sampled bbox.
message SampleConstraint {
// Minimum Jaccard overlap between sampled bbox and all bboxes in
// AnnotationGroup.
optional float min_jaccard_overlap = 1;
// Maximum Jaccard overlap between sampled bbox and all bboxes in
// AnnotationGroup.
optional float max_jaccard_overlap = 2;
// Minimum coverage of sampled bbox by all bboxes in AnnotationGroup.
optional float min_sample_coverage = 3;
// Maximum coverage of sampled bbox by all bboxes in AnnotationGroup.
optional float max_sample_coverage = 4;
// Minimum coverage of all bboxes in AnnotationGroup by sampled bbox.
optional float min_object_coverage = 5;
// Maximum coverage of all bboxes in AnnotationGroup by sampled bbox.
optional float max_object_coverage = 6;
}
// Sample a batch of bboxes with provided constraints.
message BatchSampler {
// Use original image as the source for sampling.
optional bool use_original_image = 1 [default = true];
// Constraints for sampling bbox.
optional Sampler sampler = 2;
// Constraints for determining if a sampled bbox is positive or negative.
optional SampleConstraint sample_constraint = 3;
// If provided, break when found certain number of samples satisfing the
// sample_constraint.
optional uint32 max_sample = 4;
// Maximum number of trials for sampling to avoid infinite loop.
optional uint32 max_trials = 5 [default = 100];
}
// Condition for emitting annotations.
message EmitConstraint {
enum EmitType {
CENTER = 0;
MIN_OVERLAP = 1;
}
optional EmitType emit_type = 1 [default = CENTER];
// If emit_type is MIN_OVERLAP, provide the emit_overlap.
optional float emit_overlap = 2;
}
// The normalized bounding box [0, 1] w.r.t. the input image size.
message NormalizedBBox {
optional float xmin = 1;
optional float ymin = 2;
optional float xmax = 3;
optional float ymax = 4;
optional int32 label = 5;
optional bool difficult = 6;
optional float score = 7;
optional float size = 8;
}
// Annotation for each object instance.
message Annotation {
optional int32 instance_id = 1 [default = 0];
optional NormalizedBBox bbox = 2;
}
// Group of annotations for a particular label.
message AnnotationGroup {
optional int32 group_label = 1;
repeated Annotation annotation = 2;
}
// An extension of Datum which contains "rich" annotations.
message AnnotatedDatum {
enum AnnotationType {
BBOX = 0;
}
optional Datum datum = 1;
// If there are "rich" annotations, specify the type of annotation.
// Currently it only supports bounding box.
// If there are no "rich" annotations, use label in datum instead.
optional AnnotationType type = 2;
// Each group contains annotation for a particular class.
repeated AnnotationGroup annotation_group = 3;
}
message FillerParameter {
// The filler type.
optional string type = 1 [default = 'constant'];
optional float value = 2 [default = 0]; // the value in constant filler
optional float min = 3 [default = 0]; // the min value in uniform filler
optional float max = 4 [default = 1]; // the max value in uniform filler
optional float mean = 5 [default = 0]; // the mean value in Gaussian filler
optional float std = 6 [default = 1]; // the std value in Gaussian filler
// The expected number of non-zero output weights for a given input in
// Gaussian filler -- the default -1 means don't perform sparsification.
optional int32 sparse = 7 [default = -1];
// Normalize the filler variance by fan_in, fan_out, or their average.
// Applies to 'xavier' and 'msra' fillers.
enum VarianceNorm {
FAN_IN = 0;
FAN_OUT = 1;
AVERAGE = 2;
}
optional VarianceNorm variance_norm = 8 [default = FAN_IN];
}
message NetParameter {
optional string name = 1; // consider giving the network a name
// DEPRECATED. See InputParameter. The input blobs to the network.
repeated string input = 3;
// DEPRECATED. See InputParameter. The shape of the input blobs.
repeated BlobShape input_shape = 8;
// 4D input dimensions -- deprecated. Use "input_shape" instead.
// If specified, for each input blob there should be four
// values specifying the num, channels, height and width of the input blob.
// Thus, there should be a total of (4 * #input) numbers.
repeated int32 input_dim = 4;
// Whether the network will force every layer to carry out backward operation.
// If set False, then whether to carry out backward is determined
// automatically according to the net structure and learning rates.
optional bool force_backward = 5 [default = false];
// The current "state" of the network, including the phase, level, and stage.
// Some layers may be included/excluded depending on this state and the states
// specified in the layers' include and exclude fields.
optional NetState state = 6;
// Print debugging information about results while running Net::Forward,
// Net::Backward, and Net::Update.
optional bool debug_info = 7 [default = false];
// The layers that make up the net. Each of their configurations, including
// connectivity and behavior, is specified as a LayerParameter.
repeated LayerParameter layer = 100; // ID 100 so layers are printed last.
// DEPRECATED: use 'layer' instead.
repeated V1LayerParameter layers = 2;
}
// NOTE
// Update the next available ID when you add a new SolverParameter field.
//
// SolverParameter next available ID: 44 (last added: plateau_winsize)
message SolverParameter {
//////////////////////////////////////////////////////////////////////////////
// Specifying the train and test networks
//
// Exactly one train net must be specified using one of the following fields:
// train_net_param, train_net, net_param, net
// One or more test nets may be specified using any of the following fields:
// test_net_param, test_net, net_param, net
// If more than one test net field is specified (e.g., both net and
// test_net are specified), they will be evaluated in the field order given
// above: (1) test_net_param, (2) test_net, (3) net_param/net.
// A test_iter must be specified for each test_net.
// A test_level and/or a test_stage may also be specified for each test_net.
//////////////////////////////////////////////////////////////////////////////
// Proto filename for the train net, possibly combined with one or more
// test nets.
optional string net = 24;
// Inline train net param, possibly combined with one or more test nets.
optional NetParameter net_param = 25;
optional string train_net = 1; // Proto filename for the train net.
repeated string test_net = 2; // Proto filenames for the test nets.
optional NetParameter train_net_param = 21; // Inline train net params.
repeated NetParameter test_net_param = 22; // Inline test net params.
// The states for the train/test nets. Must be unspecified or
// specified once per net.
//
// By default, all states will have solver = true;
// train_state will have phase = TRAIN,
// and all test_state's will have phase = TEST.
// Other defaults are set according to the NetState defaults.
optional NetState train_state = 26;
repeated NetState test_state = 27;
// Evaluation type.
optional string eval_type = 41 [default = "classification"];
// ap_version: different ways of computing Average Precision.
// Check https://sanchom.wordpress.com/tag/average-precision/ for details.
// 11point: the 11-point interpolated average precision. Used in VOC2007.
// MaxIntegral: maximally interpolated AP. Used in VOC2012/ILSVRC.
// Integral: the natural integral of the precision-recall curve.
optional string ap_version = 42 [default = "Integral"];
// If true, display per class result.
optional bool show_per_class_result = 44 [default = false];
// The number of iterations for each test net.
repeated int32 test_iter = 3;
// The number of iterations between two testing phases.
optional int32 test_interval = 4 [default = 0];
optional bool test_compute_loss = 19 [default = false];
// If true, run an initial test pass before the first iteration,
// ensuring memory availability and printing the starting value of the loss.
optional bool test_initialization = 32 [default = true];
optional float base_lr = 5; // The base learning rate
// the number of iterations between displaying info. If display = 0, no info
// will be displayed.
optional int32 display = 6;
// Display the loss averaged over the last average_loss iterations
optional int32 average_loss = 33 [default = 1];
optional int32 max_iter = 7; // the maximum number of iterations
// accumulate gradients over `iter_size` x `batch_size` instances
optional int32 iter_size = 36 [default = 1];
// The learning rate decay policy. The currently implemented learning rate
// policies are as follows:
// - fixed: always return base_lr.
// - step: return base_lr * gamma ^ (floor(iter / step))
// - exp: return base_lr * gamma ^ iter
// - inv: return base_lr * (1 + gamma * iter) ^ (- power)
// - multistep: similar to step but it allows non uniform steps defined by
// stepvalue
// - poly: the effective learning rate follows a polynomial decay, to be
// zero by the max_iter. return base_lr (1 - iter/max_iter) ^ (power)
// - sigmoid: the effective learning rate follows a sigmod decay
// return base_lr ( 1/(1 + exp(-gamma * (iter - stepsize))))
// - plateau: decreases lr
// if the minimum loss isn't updated for 'plateau_winsize' iters
//
// where base_lr, max_iter, gamma, step, stepvalue and power are defined
// in the solver parameter protocol buffer, and iter is the current iteration.
optional string lr_policy = 8;
optional float gamma = 9; // The parameter to compute the learning rate.
optional float power = 10; // The parameter to compute the learning rate.
optional float momentum = 11; // The momentum value.
optional float weight_decay = 12; // The weight decay.
// regularization types supported: L1 and L2
// controlled by weight_decay
optional string regularization_type = 29 [default = "L2"];
// the stepsize for learning rate policy "step"
optional int32 stepsize = 13;
// the stepsize for learning rate policy "multistep"
repeated int32 stepvalue = 34;
// the stepsize for learning rate policy "plateau"
repeated int32 plateau_winsize = 43;
// Set clip_gradients to >= 0 to clip parameter gradients to that L2 norm,
// whenever their actual L2 norm is larger.
optional float clip_gradients = 35 [default = -1];
optional int32 snapshot = 14 [default = 0]; // The snapshot interval
optional string snapshot_prefix = 15; // The prefix for the snapshot.
// whether to snapshot diff in the results or not. Snapshotting diff will help
// debugging but the final protocol buffer size will be much larger.
optional bool snapshot_diff = 16 [default = false];
enum SnapshotFormat {
HDF5 = 0;
BINARYPROTO = 1;
}
optional SnapshotFormat snapshot_format = 37 [default = BINARYPROTO];
// the mode solver will use: 0 for CPU and 1 for GPU. Use GPU in default.
enum SolverMode {
CPU = 0;
GPU = 1;
}
optional SolverMode solver_mode = 17 [default = GPU];
// the device_id will that be used in GPU mode. Use device_id = 0 in default.
optional int32 device_id = 18 [default = 0];
// If non-negative, the seed with which the Solver will initialize the Caffe
// random number generator -- useful for reproducible results. Otherwise,
// (and by default) initialize using a seed derived from the system clock.
optional int64 random_seed = 20 [default = -1];
// type of the solver
optional string type = 40 [default = "SGD"];
// numerical stability for RMSProp, AdaGrad and AdaDelta and Adam
optional float delta = 31 [default = 1e-8];
// parameters for the Adam solver
optional float momentum2 = 39 [default = 0.999];
// RMSProp decay value
// MeanSquare(t) = rms_decay*MeanSquare(t-1) + (1-rms_decay)*SquareGradient(t)
optional float rms_decay = 38 [default = 0.99];
// If true, print information about the state of the net that may help with
// debugging learning problems.
optional bool debug_info = 23 [default = false];
// If false, don't save a snapshot after training finishes.
optional bool snapshot_after_train = 28 [default = true];
// DEPRECATED: old solver enum types, use string instead
enum SolverType {
SGD = 0;
NESTEROV = 1;
ADAGRAD = 2;
RMSPROP = 3;
ADADELTA = 4;
ADAM = 5;
}
// DEPRECATED: use type instead of solver_type
optional SolverType solver_type = 30 [default = SGD];
}
// A message that stores the solver snapshots
message SolverState {
optional int32 iter = 1; // The current iteration
optional string learned_net = 2; // The file that stores the learned net.
repeated BlobProto history = 3; // The history for sgd solvers
optional int32 current_step = 4 [default = 0]; // The current step for learning rate
optional float minimum_loss = 5 [default = 1E38]; // Historical minimum loss
optional int32 iter_last_event = 6 [default = 0]; // The iteration when last lr-update or min_loss-update happend
}
enum Phase {
TRAIN = 0;
TEST = 1;
}
message NetState {
optional Phase phase = 1 [default = TEST];
optional int32 level = 2 [default = 0];
repeated string stage = 3;
}
message NetStateRule {
// Set phase to require the NetState have a particular phase (TRAIN or TEST)
// to meet this rule.
optional Phase phase = 1;
// Set the minimum and/or maximum levels in which the layer should be used.
// Leave undefined to meet the rule regardless of level.
optional int32 min_level = 2;
optional int32 max_level = 3;
// Customizable sets of stages to include or exclude.
// The net must have ALL of the specified stages and NONE of the specified
// "not_stage"s to meet the rule.
// (Use multiple NetStateRules to specify conjunctions of stages.)
repeated string stage = 4;
repeated string not_stage = 5;
}
// Specifies training parameters (multipliers on global learning constants,
// and the name and other settings used for weight sharing).
message ParamSpec {
// The names of the parameter blobs -- useful for sharing parameters among
// layers, but never required otherwise. To share a parameter between two
// layers, give it a (non-empty) name.
optional string name = 1;
// Whether to require shared weights to have the same shape, or just the same
// count -- defaults to STRICT if unspecified.
optional DimCheckMode share_mode = 2;
enum DimCheckMode {
// STRICT (default) requires that num, channels, height, width each match.
STRICT = 0;
// PERMISSIVE requires only the count (num*channels*height*width) to match.
PERMISSIVE = 1;
}
// The multiplier on the global learning rate for this parameter.
optional float lr_mult = 3 [default = 1.0];
// The multiplier on the global weight decay for this parameter.
optional float decay_mult = 4 [default = 1.0];
}
// NOTE
// Update the next available ID when you add a new LayerParameter field.
//
// LayerParameter next available layer-specific ID: 147 (last added: recurrent_param)
message LayerParameter {
optional string name = 1; // the layer name
optional string type = 2; // the layer type
repeated string bottom = 3; // the name of each bottom blob
repeated string top = 4; // the name of each top blob
// The train / test phase for computation.
optional Phase phase = 10;
// The amount of weight to assign each top blob in the objective.
// Each layer assigns a default value, usually of either 0 or 1,
// to each top blob.
repeated float loss_weight = 5;
// Specifies training parameters (multipliers on global learning constants,
// and the name and other settings used for weight sharing).
repeated ParamSpec param = 6;
// The blobs containing the numeric parameters of the layer.
repeated BlobProto blobs = 7;
// Specifies whether to backpropagate to each bottom. If unspecified,
// Caffe will automatically infer whether each input needs backpropagation
// to compute parameter gradients. If set to true for some inputs,
// backpropagation to those inputs is forced; if set false for some inputs,
// backpropagation to those inputs is skipped.
//
// The size must be either 0 or equal to the number of bottoms.
repeated bool propagate_down = 11;
// Rules controlling whether and when a layer is included in the network,
// based on the current NetState. You may specify a non-zero number of rules
// to include OR exclude, but not both. If no include or exclude rules are
// specified, the layer is always included. If the current NetState meets
// ANY (i.e., one or more) of the specified rules, the layer is
// included/excluded.
repeated NetStateRule include = 8;
repeated NetStateRule exclude = 9;
// Parameters for data pre-processing.
optional TransformationParameter transform_param = 100;
// Parameters shared by loss layers.
optional LossParameter loss_param = 101;
// Layer type-specific parameters.
//
// Note: certain layers may have more than one computational engine
// for their implementation. These layers include an Engine type and
// engine parameter for selecting the implementation.
// The default for the engine is set by the ENGINE switch at compile-time.
optional AccuracyParameter accuracy_param = 102;
optional AnnotatedDataParameter annotated_data_param = 200;
optional ArgMaxParameter argmax_param = 103;
optional BatchNormParameter batch_norm_param = 139;
optional BiasParameter bias_param = 141;
optional ConcatParameter concat_param = 104;
optional ContrastiveLossParameter contrastive_loss_param = 105;
optional ConvolutionParameter convolution_param = 106;
optional CropParameter crop_param = 144;
optional DataParameter data_param = 107;
optional DetectionEvaluateParameter detection_evaluate_param = 205;
optional DetectionOutputParameter detection_output_param = 204;
optional DropoutParameter dropout_param = 108;
optional DummyDataParameter dummy_data_param = 109;
optional EltwiseParameter eltwise_param = 110;
optional ELUParameter elu_param = 140;
optional EmbedParameter embed_param = 137;
optional ExpParameter exp_param = 111;
optional FlattenParameter flatten_param = 135;
optional HDF5DataParameter hdf5_data_param = 112;
optional HDF5OutputParameter hdf5_output_param = 113;
optional HingeLossParameter hinge_loss_param = 114;
optional ImageDataParameter image_data_param = 115;
optional InfogainLossParameter infogain_loss_param = 116;
optional InnerProductParameter inner_product_param = 117;
optional InputParameter input_param = 143;
optional LogParameter log_param = 134;
optional LRNParameter lrn_param = 118;
optional MemoryDataParameter memory_data_param = 119;
optional MultiBoxLossParameter multibox_loss_param = 201;
optional MVNParameter mvn_param = 120;
optional NormalizeParameter norm_param = 206;
optional ParameterParameter parameter_param = 145;
optional PermuteParameter permute_param = 202;
optional PoolingParameter pooling_param = 121;
optional PowerParameter power_param = 122;
optional PReLUParameter prelu_param = 131;
optional PriorBoxParameter prior_box_param = 203;
optional PythonParameter python_param = 130;
optional RecurrentParameter recurrent_param = 146;
optional ReductionParameter reduction_param = 136;
optional ReLUParameter relu_param = 123;
optional ReshapeParameter reshape_param = 133;
optional ScaleParameter scale_param = 142;
optional SigmoidParameter sigmoid_param = 124;
optional SoftmaxParameter softmax_param = 125;
optional SPPParameter spp_param = 132;
optional SliceParameter slice_param = 126;
optional TanHParameter tanh_param = 127;
optional ThresholdParameter threshold_param = 128;
optional TileParameter tile_param = 138;
optional VideoDataParameter video_data_param = 207;
optional WindowDataParameter window_data_param = 129;
optional AxpyParameter axpy_param = 210;
optional UpsampleParameter upsample_param = 211;
optional ROIPoolingParameter roi_pooling_param = 212;
optional ShuffleChannelParameter shuffle_channel_param = 213;
}
// Message that stores parameters used to apply transformation
// to the data layer's data
message TransformationParameter {
// For data pre-processing, we can do simple scaling and subtracting the
// data mean, if provided. Note that the mean subtraction is always carried
// out before scaling.
optional float scale = 1 [default = 1];
// Specify if we want to randomly mirror data.
optional bool mirror = 2 [default = false];
// Specify if we would like to randomly crop an image.
optional uint32 crop_size = 3 [default = 0];
optional uint32 crop_h = 11 [default = 0];
optional uint32 crop_w = 12 [default = 0];
// mean_file and mean_value cannot be specified at the same time
optional string mean_file = 4;
// if specified can be repeated once (would substract it from all the channels)
// or can be repeated the same number of times as channels
// (would subtract them from the corresponding channel)
repeated float mean_value = 5;
// Force the decoded image to have 3 color channels.
optional bool force_color = 6 [default = false];
// Force the decoded image to have 1 color channels.
optional bool force_gray = 7 [default = false];
// Resize policy
optional ResizeParameter resize_param = 8;
// Noise policy
optional NoiseParameter noise_param = 9;
// Distortion policy
optional DistortionParameter distort_param = 13;
// Expand policy
optional ExpansionParameter expand_param = 14;
// Constraint for emitting the annotation after transformation.
optional EmitConstraint emit_constraint = 10;
}
// Message that stores parameters used by data transformer for resize policy
message ResizeParameter {
//Probability of using this resize policy
optional float prob = 1 [default = 1];
enum Resize_mode {
WARP = 1;
FIT_SMALL_SIZE = 2;
FIT_LARGE_SIZE_AND_PAD = 3;
}
optional Resize_mode resize_mode = 2 [default = WARP];
optional uint32 height = 3 [default = 0];
optional uint32 width = 4 [default = 0];
// A parameter used to update bbox in FIT_SMALL_SIZE mode.
optional uint32 height_scale = 8 [default = 0];
optional uint32 width_scale = 9 [default = 0];
enum Pad_mode {
CONSTANT = 1;
MIRRORED = 2;
REPEAT_NEAREST = 3;
}
// Padding mode for BE_SMALL_SIZE_AND_PAD mode and object centering
optional Pad_mode pad_mode = 5 [default = CONSTANT];
// if specified can be repeated once (would fill all the channels)
// or can be repeated the same number of times as channels
// (would use it them to the corresponding channel)
repeated float pad_value = 6;
enum Interp_mode { //Same as in OpenCV
LINEAR = 1;
AREA = 2;
NEAREST = 3;
CUBIC = 4;
LANCZOS4 = 5;
}
//interpolation for for resizing
repeated Interp_mode interp_mode = 7;
}
message SaltPepperParameter {
//Percentage of pixels
optional float fraction = 1 [default = 0];
repeated float value = 2;
}
// Message that stores parameters used by data transformer for transformation
// policy
message NoiseParameter {
//Probability of using this resize policy
optional float prob = 1 [default = 0];
// Histogram equalized
optional bool hist_eq = 2 [default = false];
// Color inversion
optional bool inverse = 3 [default = false];
// Grayscale
optional bool decolorize = 4 [default = false];
// Gaussian blur
optional bool gauss_blur = 5 [default = false];
// JPEG compression quality (-1 = no compression)
optional float jpeg = 6 [default = -1];
// Posterization
optional bool posterize = 7 [default = false];
// Erosion
optional bool erode = 8 [default = false];
// Salt-and-pepper noise
optional bool saltpepper = 9 [default = false];
optional SaltPepperParameter saltpepper_param = 10;
// Local histogram equalization
optional bool clahe = 11 [default = false];
// Color space conversion
optional bool convert_to_hsv = 12 [default = false];
// Color space conversion
optional bool convert_to_lab = 13 [default = false];
}
// Message that stores parameters used by data transformer for distortion policy
message DistortionParameter {
// The probability of adjusting brightness.
optional float brightness_prob = 1 [default = 0.0];
// Amount to add to the pixel values within [-delta, delta].
// The possible value is within [0, 255]. Recommend 32.
optional float brightness_delta = 2 [default = 0.0];
// The probability of adjusting contrast.
optional float contrast_prob = 3 [default = 0.0];
// Lower bound for random contrast factor. Recommend 0.5.
optional float contrast_lower = 4 [default = 0.0];
// Upper bound for random contrast factor. Recommend 1.5.
optional float contrast_upper = 5 [default = 0.0];
// The probability of adjusting hue.
optional float hue_prob = 6 [default = 0.0];
// Amount to add to the hue channel within [-delta, delta].
// The possible value is within [0, 180]. Recommend 36.
optional float hue_delta = 7 [default = 0.0];
// The probability of adjusting saturation.
optional float saturation_prob = 8 [default = 0.0];
// Lower bound for the random saturation factor. Recommend 0.5.
optional float saturation_lower = 9 [default = 0.0];
// Upper bound for the random saturation factor. Recommend 1.5.
optional float saturation_upper = 10 [default = 0.0];
// The probability of randomly order the image channels.
optional float random_order_prob = 11 [default = 0.0];
}
// Message that stores parameters used by data transformer for expansion policy
message ExpansionParameter {
//Probability of using this expansion policy
optional float prob = 1 [default = 1];
// The ratio to expand the image.
optional float max_expand_ratio = 2 [default = 1.];
}
// Message that stores parameters shared by loss layers
message LossParameter {
// If specified, ignore instances with the given label.
optional int32 ignore_label = 1;
// How to normalize the loss for loss layers that aggregate across batches,
// spatial dimensions, or other dimensions. Currently only implemented in
// SoftmaxWithLoss and SigmoidCrossEntropyLoss layers.
enum NormalizationMode {
// Divide by the number of examples in the batch times spatial dimensions.
// Outputs that receive the ignore label will NOT be ignored in computing
// the normalization factor.
FULL = 0;
// Divide by the total number of output locations that do not take the
// ignore_label. If ignore_label is not set, this behaves like FULL.
VALID = 1;
// Divide by the batch size.
BATCH_SIZE = 2;
// Do not normalize the loss.
NONE = 3;
}
// For historical reasons, the default normalization for
// SigmoidCrossEntropyLoss is BATCH_SIZE and *not* VALID.
optional NormalizationMode normalization = 3 [default = VALID];
// Deprecated. Ignored if normalization is specified. If normalization
// is not specified, then setting this to false will be equivalent to
// normalization = BATCH_SIZE to be consistent with previous behavior.
optional bool normalize = 2;
}
// Messages that store parameters used by individual layer types follow, in
// alphabetical order.
message AccuracyParameter {
// When computing accuracy, count as correct by comparing the true label to
// the top k scoring classes. By default, only compare to the top scoring
// class (i.e. argmax).
optional uint32 top_k = 1 [default = 1];
// The "label" axis of the prediction blob, whose argmax corresponds to the
// predicted label -- may be negative to index from the end (e.g., -1 for the
// last axis). For example, if axis == 1 and the predictions are
// (N x C x H x W), the label blob is expected to contain N*H*W ground truth
// labels with integer values in {0, 1, ..., C-1}.
optional int32 axis = 2 [default = 1];
// If specified, ignore instances with the given label.
optional int32 ignore_label = 3;
}
message AnnotatedDataParameter {
// Define the sampler.
repeated BatchSampler batch_sampler = 1;
// Store label name and label id in LabelMap format.
optional string label_map_file = 2;
// If provided, it will replace the AnnotationType stored in each
// AnnotatedDatum.
optional AnnotatedDatum.AnnotationType anno_type = 3;
}
message ArgMaxParameter {
// If true produce pairs (argmax, maxval)
optional bool out_max_val = 1 [default = false];
optional uint32 top_k = 2 [default = 1];
// The axis along which to maximise -- may be negative to index from the
// end (e.g., -1 for the last axis).
// By default ArgMaxLayer maximizes over the flattened trailing dimensions
// for each index of the first / num dimension.
optional int32 axis = 3;
}
message ConcatParameter {
// The axis along which to concatenate -- may be negative to index from the
// end (e.g., -1 for the last axis). Other axes must have the
// same dimension for all the bottom blobs.
// By default, ConcatLayer concatenates blobs along the "channels" axis (1).
optional int32 axis = 2 [default = 1];
// DEPRECATED: alias for "axis" -- does not support negative indexing.
optional uint32 concat_dim = 1 [default = 1];
}
message BatchNormParameter {
// If false, accumulate global mean/variance values via a moving average. If
// true, use those accumulated values instead of computing mean/variance
// across the batch.
optional bool use_global_stats = 1;
// How much does the moving average decay each iteration?
optional float moving_average_fraction = 2 [default = .999];
// Small value to add to the variance estimate so that we don't divide by
// zero.
optional float eps = 3 [default = 1e-5];
}
message BiasParameter {
// The first axis of bottom[0] (the first input Blob) along which to apply
// bottom[1] (the second input Blob). May be negative to index from the end
// (e.g., -1 for the last axis).
//
// For example, if bottom[0] is 4D with shape 100x3x40x60, the output
// top[0] will have the same shape, and bottom[1] may have any of the
// following shapes (for the given value of axis):
// (axis == 0 == -4) 100; 100x3; 100x3x40; 100x3x40x60
// (axis == 1 == -3) 3; 3x40; 3x40x60
// (axis == 2 == -2) 40; 40x60
// (axis == 3 == -1) 60
// Furthermore, bottom[1] may have the empty shape (regardless of the value of
// "axis") -- a scalar bias.
optional int32 axis = 1 [default = 1];
// (num_axes is ignored unless just one bottom is given and the bias is
// a learned parameter of the layer. Otherwise, num_axes is determined by the
// number of axes by the second bottom.)
// The number of axes of the input (bottom[0]) covered by the bias
// parameter, or -1 to cover all axes of bottom[0] starting from `axis`.
// Set num_axes := 0, to add a zero-axis Blob: a scalar.
optional int32 num_axes = 2 [default = 1];
// (filler is ignored unless just one bottom is given and the bias is
// a learned parameter of the layer.)
// The initialization for the learned bias parameter.
// Default is the zero (0) initialization, resulting in the BiasLayer
// initially performing the identity operation.
optional FillerParameter filler = 3;
}
message ContrastiveLossParameter {
// margin for dissimilar pair
optional float margin = 1 [default = 1.0];
// The first implementation of this cost did not exactly match the cost of
// Hadsell et al 2006 -- using (margin - d^2) instead of (margin - d)^2.
// legacy_version = false (the default) uses (margin - d)^2 as proposed in the
// Hadsell paper. New models should probably use this version.
// legacy_version = true uses (margin - d^2). This is kept to support /
// reproduce existing models and results
optional bool legacy_version = 2 [default = false];
}
message ConvolutionParameter {
optional uint32 num_output = 1; // The number of outputs for the layer
optional bool bias_term = 2 [default = true]; // whether to have bias terms
// Pad, kernel size, and stride are all given as a single value for equal
// dimensions in all spatial dimensions, or once per spatial dimension.
repeated uint32 pad = 3; // The padding size; defaults to 0
repeated uint32 kernel_size = 4; // The kernel size
repeated uint32 stride = 6; // The stride; defaults to 1
// Factor used to dilate the kernel, (implicitly) zero-filling the resulting
// holes. (Kernel dilation is sometimes referred to by its use in the
// algorithme à trous from Holschneider et al. 1987.)
repeated uint32 dilation = 18; // The dilation; defaults to 1
// For 2D convolution only, the *_h and *_w versions may also be used to
// specify both spatial dimensions.
optional uint32 pad_h = 9 [default = 0]; // The padding height (2D only)
optional uint32 pad_w = 10 [default = 0]; // The padding width (2D only)
optional uint32 kernel_h = 11; // The kernel height (2D only)
optional uint32 kernel_w = 12; // The kernel width (2D only)
optional uint32 stride_h = 13; // The stride height (2D only)
optional uint32 stride_w = 14; // The stride width (2D only)
optional uint32 group = 5 [default = 1]; // The group size for group conv
optional FillerParameter weight_filler = 7; // The filler for the weight
optional FillerParameter bias_filler = 8; // The filler for the bias
enum Engine {
DEFAULT = 0;
CAFFE = 1;
CUDNN = 2;
}
optional Engine engine = 15 [default = DEFAULT];
// The axis to interpret as "channels" when performing convolution.
// Preceding dimensions are treated as independent inputs;
// succeeding dimensions are treated as "spatial".
// With (N, C, H, W) inputs, and axis == 1 (the default), we perform
// N independent 2D convolutions, sliding C-channel (or (C/g)-channels, for
// groups g>1) filters across the spatial axes (H, W) of the input.
// With (N, C, D, H, W) inputs, and axis == 1, we perform
// N independent 3D convolutions, sliding (C/g)-channels
// filters across the spatial axes (D, H, W) of the input.
optional int32 axis = 16 [default = 1];
// Whether to force use of the general ND convolution, even if a specific
// implementation for blobs of the appropriate number of spatial dimensions
// is available. (Currently, there is only a 2D-specific convolution
// implementation; for input blobs with num_axes != 2, this option is
// ignored and the ND implementation will be used.)
optional bool force_nd_im2col = 17 [default = false];
}
message CropParameter {
// To crop, elements of the first bottom are selected to fit the dimensions
// of the second, reference bottom. The crop is configured by
// - the crop `axis` to pick the dimensions for cropping
// - the crop `offset` to set the shift for all/each dimension
// to align the cropped bottom with the reference bottom.
// All dimensions up to but excluding `axis` are preserved, while
// the dimensions including and trailing `axis` are cropped.
// If only one `offset` is set, then all dimensions are offset by this amount.
// Otherwise, the number of offsets must equal the number of cropped axes to
// shift the crop in each dimension accordingly.
// Note: standard dimensions are N,C,H,W so the default is a spatial crop,
// and `axis` may be negative to index from the end (e.g., -1 for the last
// axis).
optional int32 axis = 1 [default = 2];
repeated uint32 offset = 2;
}
message DataParameter {
enum DB {
LEVELDB = 0;
LMDB = 1;
}
// Specify the data source.
optional string source = 1;
// Specify the batch size.
optional uint32 batch_size = 4;
// The rand_skip variable is for the data layer to skip a few data points
// to avoid all asynchronous sgd clients to start at the same point. The skip
// point would be set as rand_skip * rand(0,1). Note that rand_skip should not
// be larger than the number of keys in the database.
// DEPRECATED. Each solver accesses a different subset of the database.
optional uint32 rand_skip = 7 [default = 0];
optional DB backend = 8 [default = LEVELDB];
// DEPRECATED. See TransformationParameter. For data pre-processing, we can do
// simple scaling and subtracting the data mean, if provided. Note that the
// mean subtraction is always carried out before scaling.
optional float scale = 2 [default = 1];
optional string mean_file = 3;
// DEPRECATED. See TransformationParameter. Specify if we would like to randomly
// crop an image.
optional uint32 crop_size = 5 [default = 0];
// DEPRECATED. See TransformationParameter. Specify if we want to randomly mirror
// data.
optional bool mirror = 6 [default = false];
// Force the encoded image to have 3 color channels
optional bool force_encoded_color = 9 [default = false];
// Prefetch queue (Number of batches to prefetch to host memory, increase if
// data access bandwidth varies).
optional uint32 prefetch = 10 [default = 4];
}
// Message that store parameters used by DetectionEvaluateLayer
message DetectionEvaluateParameter {
// Number of classes that are actually predicted. Required!
optional uint32 num_classes = 1;
// Label id for background class. Needed for sanity check so that
// background class is neither in the ground truth nor the detections.
optional uint32 background_label_id = 2 [default = 0];
// Threshold for deciding true/false positive.
optional float overlap_threshold = 3 [default = 0.5];
// If true, also consider difficult ground truth for evaluation.
optional bool evaluate_difficult_gt = 4 [default = true];
// A file which contains a list of names and sizes with same order
// of the input DB. The file is in the following format:
// name height width
// ...
// If provided, we will scale the prediction and ground truth NormalizedBBox
// for evaluation.
optional string name_size_file = 5;
// The resize parameter used in converting NormalizedBBox to original image.
optional ResizeParameter resize_param = 6;
}
message NonMaximumSuppressionParameter {
// Threshold to be used in nms.
optional float nms_threshold = 1 [default = 0.3];
// Maximum number of results to be kept.
optional int32 top_k = 2;
// Parameter for adaptive nms.
optional float eta = 3 [default = 1.0];
}
message SaveOutputParameter {
// Output directory. If not empty, we will save the results.
optional string output_directory = 1;
// Output name prefix.
optional string output_name_prefix = 2;
// Output format.
// VOC - PASCAL VOC output format.
// COCO - MS COCO output format.
optional string output_format = 3;
// If you want to output results, must also provide the following two files.
// Otherwise, we will ignore saving results.
// label map file.
optional string label_map_file = 4;
// A file which contains a list of names and sizes with same order
// of the input DB. The file is in the following format:
// name height width
// ...
optional string name_size_file = 5;
// Number of test images. It can be less than the lines specified in
// name_size_file. For example, when we only want to evaluate on part
// of the test images.
optional uint32 num_test_image = 6;
// The resize parameter used in saving the data.
optional ResizeParameter resize_param = 7;
}
// Message that store parameters used by DetectionOutputLayer
message DetectionOutputParameter {
// Number of classes to be predicted. Required!
optional uint32 num_classes = 1;
// If true, bounding box are shared among different classes.
optional bool share_location = 2 [default = true];
// Background label id. If there is no background class,
// set it as -1.
optional int32 background_label_id = 3 [default = 0];
// Parameters used for non maximum suppression.
optional NonMaximumSuppressionParameter nms_param = 4;
// Parameters used for saving detection results.
optional SaveOutputParameter save_output_param = 5;
// Type of coding method for bbox.
optional PriorBoxParameter.CodeType code_type = 6 [default = CORNER];
// If true, variance is encoded in target; otherwise we need to adjust the
// predicted offset accordingly.
optional bool variance_encoded_in_target = 8 [default = false];
// Number of total bboxes to be kept per image after nms step.
// -1 means keeping all bboxes after nms step.
optional int32 keep_top_k = 7 [default = -1];
// Only consider detections whose confidences are larger than a threshold.
// If not provided, consider all boxes.
optional float confidence_threshold = 9;
// If true, visualize the detection results.
optional bool visualize = 10 [default = false];
// The threshold used to visualize the detection results.
optional float visualize_threshold = 11;
// If provided, save outputs to video file.
optional string save_file = 12;
}
message DropoutParameter {
optional float dropout_ratio = 1 [default = 0.5]; // dropout ratio
}
// DummyDataLayer fills any number of arbitrarily shaped blobs with random
// (or constant) data generated by "Fillers" (see "message FillerParameter").
message DummyDataParameter {
// This layer produces N >= 1 top blobs. DummyDataParameter must specify 1 or N
// shape fields, and 0, 1 or N data_fillers.
//
// If 0 data_fillers are specified, ConstantFiller with a value of 0 is used.
// If 1 data_filler is specified, it is applied to all top blobs. If N are
// specified, the ith is applied to the ith top blob.
repeated FillerParameter data_filler = 1;
repeated BlobShape shape = 6;
// 4D dimensions -- deprecated. Use "shape" instead.
repeated uint32 num = 2;
repeated uint32 channels = 3;
repeated uint32 height = 4;
repeated uint32 width = 5;
}
message EltwiseParameter {
enum EltwiseOp {
PROD = 0;
SUM = 1;
MAX = 2;
}
optional EltwiseOp operation = 1 [default = SUM]; // element-wise operation
repeated float coeff = 2; // blob-wise coefficient for SUM operation
// Whether to use an asymptotically slower (for >2 inputs) but stabler method
// of computing the gradient for the PROD operation. (No effect for SUM op.)
optional bool stable_prod_grad = 3 [default = true];
}
// Message that stores parameters used by ELULayer
message ELUParameter {
// Described in:
// Clevert, D.-A., Unterthiner, T., & Hochreiter, S. (2015). Fast and Accurate
// Deep Network Learning by Exponential Linear Units (ELUs). arXiv
optional float alpha = 1 [default = 1];
}
// Message that stores parameters used by EmbedLayer
message EmbedParameter {
optional uint32 num_output = 1; // The number of outputs for the layer
// The input is given as integers to be interpreted as one-hot
// vector indices with dimension num_input. Hence num_input should be
// 1 greater than the maximum possible input value.
optional uint32 input_dim = 2;
optional bool bias_term = 3 [default = true]; // Whether to use a bias term
optional FillerParameter weight_filler = 4; // The filler for the weight
optional FillerParameter bias_filler = 5; // The filler for the bias
}
// Message that stores parameters used by ExpLayer
message ExpParameter {
// ExpLayer computes outputs y = base ^ (shift + scale * x), for base > 0.
// Or if base is set to the default (-1), base is set to e,
// so y = exp(shift + scale * x).
optional float base = 1 [default = -1.0];
optional float scale = 2 [default = 1.0];
optional float shift = 3 [default = 0.0];
}
/// Message that stores parameters used by FlattenLayer
message FlattenParameter {
// The first axis to flatten: all preceding axes are retained in the output.
// May be negative to index from the end (e.g., -1 for the last axis).
optional int32 axis = 1 [default = 1];
// The last axis to flatten: all following axes are retained in the output.
// May be negative to index from the end (e.g., the default -1 for the last
// axis).
optional int32 end_axis = 2 [default = -1];
}
// Message that stores parameters used by HDF5DataLayer
message HDF5DataParameter {
// Specify the data source.
optional string source = 1;
// Specify the batch size.
optional uint32 batch_size = 2;
// Specify whether to shuffle the data.
// If shuffle == true, the ordering of the HDF5 files is shuffled,
// and the ordering of data within any given HDF5 file is shuffled,
// but data between different files are not interleaved; all of a file's
// data are output (in a random order) before moving onto another file.
optional bool shuffle = 3 [default = false];
}
message HDF5OutputParameter {
optional string file_name = 1;
}
message HingeLossParameter {
enum Norm {
L1 = 1;
L2 = 2;
}
// Specify the Norm to use L1 or L2
optional Norm norm = 1 [default = L1];
}
message ImageDataParameter {
// Specify the data source.
optional string source = 1;
// Specify the batch size.
optional uint32 batch_size = 4 [default = 1];
// The rand_skip variable is for the data layer to skip a few data points
// to avoid all asynchronous sgd clients to start at the same point. The skip
// point would be set as rand_skip * rand(0,1). Note that rand_skip should not
// be larger than the number of keys in the database.
optional uint32 rand_skip = 7 [default = 0];
// Whether or not ImageLayer should shuffle the list of files at every epoch.
optional bool shuffle = 8 [default = false];
// It will also resize images if new_height or new_width are not zero.
optional uint32 new_height = 9 [default = 0];
optional uint32 new_width = 10 [default = 0];
// Specify if the images are color or gray
optional bool is_color = 11 [default = true];
// DEPRECATED. See TransformationParameter. For data pre-processing, we can do
// simple scaling and subtracting the data mean, if provided. Note that the
// mean subtraction is always carried out before scaling.
optional float scale = 2 [default = 1];
optional string mean_file = 3;
// DEPRECATED. See TransformationParameter. Specify if we would like to randomly
// crop an image.
optional uint32 crop_size = 5 [default = 0];
// DEPRECATED. See TransformationParameter. Specify if we want to randomly mirror
// data.
optional bool mirror = 6 [default = false];
optional string root_folder = 12 [default = ""];
}
message InfogainLossParameter {
// Specify the infogain matrix source.
optional string source = 1;
}
message InnerProductParameter {
optional uint32 num_output = 1; // The number of outputs for the layer
optional bool bias_term = 2 [default = true]; // whether to have bias terms
optional FillerParameter weight_filler = 3; // The filler for the weight
optional FillerParameter bias_filler = 4; // The filler for the bias
// The first axis to be lumped into a single inner product computation;
// all preceding axes are retained in the output.
// May be negative to index from the end (e.g., -1 for the last axis).
optional int32 axis = 5 [default = 1];
// Specify whether to transpose the weight matrix or not.
// If transpose == true, any operations will be performed on the transpose
// of the weight matrix. The weight matrix itself is not going to be transposed
// but rather the transfer flag of operations will be toggled accordingly.
optional bool transpose = 6 [default = false];
}
message InputParameter {
// This layer produces N >= 1 top blob(s) to be assigned manually.
// Define N shapes to set a shape for each top.
// Define 1 shape to set the same shape for every top.
// Define no shape to defer to reshaping manually.
repeated BlobShape shape = 1;
}
// Message that stores parameters used by LogLayer
message LogParameter {
// LogLayer computes outputs y = log_base(shift + scale * x), for base > 0.
// Or if base is set to the default (-1), base is set to e,
// so y = ln(shift + scale * x) = log_e(shift + scale * x)
optional float base = 1 [default = -1.0];
optional float scale = 2 [default = 1.0];
optional float shift = 3 [default = 0.0];
}
// Message that stores parameters used by LRNLayer
message LRNParameter {
optional uint32 local_size = 1 [default = 5];
optional float alpha = 2 [default = 1.];
optional float beta = 3 [default = 0.75];
enum NormRegion {
ACROSS_CHANNELS = 0;
WITHIN_CHANNEL = 1;
}
optional NormRegion norm_region = 4 [default = ACROSS_CHANNELS];
optional float k = 5 [default = 1.];
enum Engine {
DEFAULT = 0;
CAFFE = 1;
CUDNN = 2;
}
optional Engine engine = 6 [default = DEFAULT];
}
message MemoryDataParameter {
optional uint32 batch_size = 1;
optional uint32 channels = 2;
optional uint32 height = 3;
optional uint32 width = 4;
}
// Message that store parameters used by MultiBoxLossLayer
message MultiBoxLossParameter {
// Localization loss type.
enum LocLossType {
L2 = 0;
SMOOTH_L1 = 1;
}
optional LocLossType loc_loss_type = 1 [default = SMOOTH_L1];
// Confidence loss type.
enum ConfLossType {
SOFTMAX = 0;
LOGISTIC = 1;
}
optional ConfLossType conf_loss_type = 2 [default = SOFTMAX];
// Weight for localization loss.
optional float loc_weight = 3 [default = 1.0];
// Number of classes to be predicted. Required!
optional uint32 num_classes = 4;
// If true, bounding box are shared among different classes.
optional bool share_location = 5 [default = true];
// Matching method during training.
enum MatchType {
BIPARTITE = 0;
PER_PREDICTION = 1;
}
optional MatchType match_type = 6 [default = PER_PREDICTION];
// If match_type is PER_PREDICTION, use overlap_threshold to
// determine the extra matching bboxes.
optional float overlap_threshold = 7 [default = 0.5];
// Use prior for matching.
optional bool use_prior_for_matching = 8 [default = true];
// Background label id.
optional uint32 background_label_id = 9 [default = 0];
// If true, also consider difficult ground truth.
optional bool use_difficult_gt = 10 [default = true];
// If true, perform negative mining.
// DEPRECATED: use mining_type instead.
optional bool do_neg_mining = 11;
// The negative/positive ratio.
optional float neg_pos_ratio = 12 [default = 3.0];
// The negative overlap upperbound for the unmatched predictions.
optional float neg_overlap = 13 [default = 0.5];
// Type of coding method for bbox.
optional PriorBoxParameter.CodeType code_type = 14 [default = CORNER];
// If true, encode the variance of prior box in the loc loss target instead of
// in bbox.
optional bool encode_variance_in_target = 16 [default = false];
// If true, map all object classes to agnostic class. It is useful for learning
// objectness detector.
optional bool map_object_to_agnostic = 17 [default = false];
// If true, ignore cross boundary bbox during matching.
// Cross boundary bbox is a bbox who is outside of the image region.
optional bool ignore_cross_boundary_bbox = 18 [default = false];
// If true, only backpropagate on corners which are inside of the image
// region when encode_type is CORNER or CORNER_SIZE.
optional bool bp_inside = 19 [default = false];
// Mining type during training.
// NONE : use all negatives.
// MAX_NEGATIVE : select negatives based on the score.
// HARD_EXAMPLE : select hard examples based on "Training Region-based Object Detectors with Online Hard Example Mining", Shrivastava et.al.
enum MiningType {
NONE = 0;
MAX_NEGATIVE = 1;
HARD_EXAMPLE = 2;
}
optional MiningType mining_type = 20 [default = MAX_NEGATIVE];
// Parameters used for non maximum suppression durig hard example mining.
optional NonMaximumSuppressionParameter nms_param = 21;
optional int32 sample_size = 22 [default = 64];
optional bool use_prior_for_nms = 23 [default = false];
}
message MVNParameter {
// This parameter can be set to false to normalize mean only
optional bool normalize_variance = 1 [default = true];
// This parameter can be set to true to perform DNN-like MVN
optional bool across_channels = 2 [default = false];
// Epsilon for not dividing by zero while normalizing variance
optional float eps = 3 [default = 1e-9];
}
// Message that stores parameters used by NormalizeLayer
message NormalizeParameter {
optional bool across_spatial = 1 [default = true];
// Initial value of scale. Default is 1.0 for all
optional FillerParameter scale_filler = 2;
// Whether or not scale parameters are shared across channels.
optional bool channel_shared = 3 [default = true];
// Epsilon for not dividing by zero while normalizing variance
optional float eps = 4 [default = 1e-10];
}
message ParameterParameter {
optional BlobShape shape = 1;
}
message PermuteParameter {
// The new orders of the axes of data. Notice it should be with
// in the same range as the input data, and it starts from 0.
// Do not provide repeated order.
repeated uint32 order = 1;
}
message PoolingParameter {
enum PoolMethod {
MAX = 0;
AVE = 1;
STOCHASTIC = 2;
}
optional PoolMethod pool = 1 [default = MAX]; // The pooling method
// Pad, kernel size, and stride are all given as a single value for equal
// dimensions in height and width or as Y, X pairs.
optional uint32 pad = 4 [default = 0]; // The padding size (equal in Y, X)
optional uint32 pad_h = 9 [default = 0]; // The padding height
optional uint32 pad_w = 10 [default = 0]; // The padding width
optional uint32 kernel_size = 2; // The kernel size (square)
optional uint32 kernel_h = 5; // The kernel height
optional uint32 kernel_w = 6; // The kernel width
optional uint32 stride = 3 [default = 1]; // The stride (equal in Y, X)
optional uint32 stride_h = 7; // The stride height
optional uint32 stride_w = 8; // The stride width
enum Engine {
DEFAULT = 0;
CAFFE = 1;
CUDNN = 2;
}
optional Engine engine = 11 [default = DEFAULT];
// If global_pooling then it will pool over the size of the bottom by doing
// kernel_h = bottom->height and kernel_w = bottom->width
optional bool global_pooling = 12 [default = false];
}
message PowerParameter {
// PowerLayer computes outputs y = (shift + scale * x) ^ power.
optional float power = 1 [default = 1.0];
optional float scale = 2 [default = 1.0];
optional float shift = 3 [default = 0.0];
}
// Message that store parameters used by PriorBoxLayer
message PriorBoxParameter {
// Encode/decode type.
enum CodeType {
CORNER = 1;
CENTER_SIZE = 2;
CORNER_SIZE = 3;
}
// Minimum box size (in pixels). Required!
repeated float min_size = 1;
// Maximum box size (in pixels). Required!
repeated float max_size = 2;
// Various of aspect ratios. Duplicate ratios will be ignored.
// If none is provided, we use default ratio 1.
repeated float aspect_ratio = 3;
// If true, will flip each aspect ratio.
// For example, if there is aspect ratio "r",
// we will generate aspect ratio "1.0/r" as well.
optional bool flip = 4 [default = true];
// If true, will clip the prior so that it is within [0, 1]
optional bool clip = 5 [default = false];
// Variance for adjusting the prior bboxes.
repeated float variance = 6;
// By default, we calculate img_height, img_width, step_x, step_y based on
// bottom[0] (feat) and bottom[1] (img). Unless these values are explicitely
// provided.
// Explicitly provide the img_size.
optional uint32 img_size = 7;
// Either img_size or img_h/img_w should be specified; not both.
optional uint32 img_h = 8;
optional uint32 img_w = 9;
// Explicitly provide the step size.
optional float step = 10;
// Either step or step_h/step_w should be specified; not both.
optional float step_h = 11;
optional float step_w = 12;
// Offset to the top left corner of each cell.
optional float offset = 13 [default = 0.5];
}
message PythonParameter {
optional string module = 1;
optional string layer = 2;
// This value is set to the attribute `param_str` of the `PythonLayer` object
// in Python before calling the `setup()` method. This could be a number,
// string, dictionary in Python dict format, JSON, etc. You may parse this
// string in `setup` method and use it in `forward` and `backward`.
optional string param_str = 3 [default = ''];
// Whether this PythonLayer is shared among worker solvers during data parallelism.
// If true, each worker solver sequentially run forward from this layer.
// This value should be set true if you are using it as a data layer.
optional bool share_in_parallel = 4 [default = false];
}
// Message that stores parameters used by RecurrentLayer
message RecurrentParameter {
// The dimension of the output (and usually hidden state) representation --
// must be explicitly set to non-zero.
optional uint32 num_output = 1 [default = 0];
optional FillerParameter weight_filler = 2; // The filler for the weight
optional FillerParameter bias_filler = 3; // The filler for the bias
// Whether to enable displaying debug_info in the unrolled recurrent net.
optional bool debug_info = 4 [default = false];
// Whether to add as additional inputs (bottoms) the initial hidden state
// blobs, and add as additional outputs (tops) the final timestep hidden state
// blobs. The number of additional bottom/top blobs required depends on the
// recurrent architecture -- e.g., 1 for RNNs, 2 for LSTMs.
optional bool expose_hidden = 5 [default = false];
}
// Message that stores parameters used by ReductionLayer
message ReductionParameter {
enum ReductionOp {
SUM = 1;
ASUM = 2;
SUMSQ = 3;
MEAN = 4;
}
optional ReductionOp operation = 1 [default = SUM]; // reduction operation
// The first axis to reduce to a scalar -- may be negative to index from the
// end (e.g., -1 for the last axis).
// (Currently, only reduction along ALL "tail" axes is supported; reduction
// of axis M through N, where N < num_axes - 1, is unsupported.)
// Suppose we have an n-axis bottom Blob with shape:
// (d0, d1, d2, ..., d(m-1), dm, d(m+1), ..., d(n-1)).
// If axis == m, the output Blob will have shape
// (d0, d1, d2, ..., d(m-1)),
// and the ReductionOp operation is performed (d0 * d1 * d2 * ... * d(m-1))
// times, each including (dm * d(m+1) * ... * d(n-1)) individual data.
// If axis == 0 (the default), the output Blob always has the empty shape
// (count 1), performing reduction across the entire input --
// often useful for creating new loss functions.
optional int32 axis = 2 [default = 0];
optional float coeff = 3 [default = 1.0]; // coefficient for output
}
// Message that stores parameters used by ReLULayer
message ReLUParameter {
// Allow non-zero slope for negative inputs to speed up optimization
// Described in:
// Maas, A. L., Hannun, A. Y., & Ng, A. Y. (2013). Rectifier nonlinearities
// improve neural network acoustic models. In ICML Workshop on Deep Learning
// for Audio, Speech, and Language Processing.
optional float negative_slope = 1 [default = 0];
enum Engine {
DEFAULT = 0;
CAFFE = 1;
CUDNN = 2;
}
optional Engine engine = 2 [default = DEFAULT];
}
message ReshapeParameter {
// Specify the output dimensions. If some of the dimensions are set to 0,
// the corresponding dimension from the bottom layer is used (unchanged).
// Exactly one dimension may be set to -1, in which case its value is
// inferred from the count of the bottom blob and the remaining dimensions.
// For example, suppose we want to reshape a 2D blob "input" with shape 2 x 8:
//
// layer {
// type: "Reshape" bottom: "input" top: "output"
// reshape_param { ... }
// }
//
// If "input" is 2D with shape 2 x 8, then the following reshape_param
// specifications are all equivalent, producing a 3D blob "output" with shape
// 2 x 2 x 4:
//
// reshape_param { shape { dim: 2 dim: 2 dim: 4 } }
// reshape_param { shape { dim: 0 dim: 2 dim: 4 } }
// reshape_param { shape { dim: 0 dim: 2 dim: -1 } }
// reshape_param { shape { dim: 0 dim:-1 dim: 4 } }
//
optional BlobShape shape = 1;
// axis and num_axes control the portion of the bottom blob's shape that are
// replaced by (included in) the reshape. By default (axis == 0 and
// num_axes == -1), the entire bottom blob shape is included in the reshape,
// and hence the shape field must specify the entire output shape.
//
// axis may be non-zero to retain some portion of the beginning of the input
// shape (and may be negative to index from the end; e.g., -1 to begin the
// reshape after the last axis, including nothing in the reshape,
// -2 to include only the last axis, etc.).
//
// For example, suppose "input" is a 2D blob with shape 2 x 8.
// Then the following ReshapeLayer specifications are all equivalent,
// producing a blob "output" with shape 2 x 2 x 4:
//
// reshape_param { shape { dim: 2 dim: 2 dim: 4 } }
// reshape_param { shape { dim: 2 dim: 4 } axis: 1 }
// reshape_param { shape { dim: 2 dim: 4 } axis: -3 }
//
// num_axes specifies the extent of the reshape.
// If num_axes >= 0 (and axis >= 0), the reshape will be performed only on
// input axes in the range [axis, axis+num_axes].
// num_axes may also be -1, the default, to include all remaining axes
// (starting from axis).
//
// For example, suppose "input" is a 2D blob with shape 2 x 8.
// Then the following ReshapeLayer specifications are equivalent,
// producing a blob "output" with shape 1 x 2 x 8.
//
// reshape_param { shape { dim: 1 dim: 2 dim: 8 } }
// reshape_param { shape { dim: 1 dim: 2 } num_axes: 1 }
// reshape_param { shape { dim: 1 } num_axes: 0 }
//
// On the other hand, these would produce output blob shape 2 x 1 x 8:
//
// reshape_param { shape { dim: 2 dim: 1 dim: 8 } }
// reshape_param { shape { dim: 1 } axis: 1 num_axes: 0 }
//
optional int32 axis = 2 [default = 0];
optional int32 num_axes = 3 [default = -1];
}
message ScaleParameter {
// The first axis of bottom[0] (the first input Blob) along which to apply
// bottom[1] (the second input Blob). May be negative to index from the end
// (e.g., -1 for the last axis).
//
// For example, if bottom[0] is 4D with shape 100x3x40x60, the output
// top[0] will have the same shape, and bottom[1] may have any of the
// following shapes (for the given value of axis):
// (axis == 0 == -4) 100; 100x3; 100x3x40; 100x3x40x60
// (axis == 1 == -3) 3; 3x40; 3x40x60
// (axis == 2 == -2) 40; 40x60
// (axis == 3 == -1) 60
// Furthermore, bottom[1] may have the empty shape (regardless of the value of
// "axis") -- a scalar multiplier.
optional int32 axis = 1 [default = 1];
// (num_axes is ignored unless just one bottom is given and the scale is
// a learned parameter of the layer. Otherwise, num_axes is determined by the
// number of axes by the second bottom.)
// The number of axes of the input (bottom[0]) covered by the scale
// parameter, or -1 to cover all axes of bottom[0] starting from `axis`.
// Set num_axes := 0, to multiply with a zero-axis Blob: a scalar.
optional int32 num_axes = 2 [default = 1];
// (filler is ignored unless just one bottom is given and the scale is
// a learned parameter of the layer.)
// The initialization for the learned scale parameter.
// Default is the unit (1) initialization, resulting in the ScaleLayer
// initially performing the identity operation.
optional FillerParameter filler = 3;
// Whether to also learn a bias (equivalent to a ScaleLayer+BiasLayer, but
// may be more efficient). Initialized with bias_filler (defaults to 0).
optional bool bias_term = 4 [default = false];
optional FillerParameter bias_filler = 5;
}
message SigmoidParameter {
enum Engine {
DEFAULT = 0;
CAFFE = 1;
CUDNN = 2;
}
optional Engine engine = 1 [default = DEFAULT];
}
message SliceParameter {
// The axis along which to slice -- may be negative to index from the end
// (e.g., -1 for the last axis).
// By default, SliceLayer concatenates blobs along the "channels" axis (1).
optional int32 axis = 3 [default = 1];
repeated uint32 slice_point = 2;
// DEPRECATED: alias for "axis" -- does not support negative indexing.
optional uint32 slice_dim = 1 [default = 1];
}
// Message that stores parameters used by SoftmaxLayer, SoftmaxWithLossLayer
message SoftmaxParameter {
enum Engine {
DEFAULT = 0;
CAFFE = 1;
CUDNN = 2;
}
optional Engine engine = 1 [default = DEFAULT];
// The axis along which to perform the softmax -- may be negative to index
// from the end (e.g., -1 for the last axis).
// Any other axes will be evaluated as independent softmaxes.
optional int32 axis = 2 [default = 1];
}
message TanHParameter {
enum Engine {
DEFAULT = 0;
CAFFE = 1;
CUDNN = 2;
}
optional Engine engine = 1 [default = DEFAULT];
}
// Message that stores parameters used by TileLayer
message TileParameter {
// The index of the axis to tile.
optional int32 axis = 1 [default = 1];
// The number of copies (tiles) of the blob to output.
optional int32 tiles = 2;
}
// Message that stores parameters used by ThresholdLayer
message ThresholdParameter {
optional float threshold = 1 [default = 0]; // Strictly positive values
}
message VideoDataParameter{
enum VideoType {
WEBCAM = 0;
VIDEO = 1;
}
optional VideoType video_type = 1 [default = WEBCAM];
optional int32 device_id = 2 [default = 0];
optional string video_file = 3;
// Number of frames to be skipped before processing a frame.
optional uint32 skip_frames = 4 [default = 0];
}
message WindowDataParameter {
// Specify the data source.
optional string source = 1;
// For data pre-processing, we can do simple scaling and subtracting the
// data mean, if provided. Note that the mean subtraction is always carried
// out before scaling.
optional float scale = 2 [default = 1];
optional string mean_file = 3;
// Specify the batch size.
optional uint32 batch_size = 4;
// Specify if we would like to randomly crop an image.
optional uint32 crop_size = 5 [default = 0];
// Specify if we want to randomly mirror data.
optional bool mirror = 6 [default = false];
// Foreground (object) overlap threshold
optional float fg_threshold = 7 [default = 0.5];
// Background (non-object) overlap threshold
optional float bg_threshold = 8 [default = 0.5];
// Fraction of batch that should be foreground objects
optional float fg_fraction = 9 [default = 0.25];
// Amount of contextual padding to add around a window
// (used only by the window_data_layer)
optional uint32 context_pad = 10 [default = 0];
// Mode for cropping out a detection window
// warp: cropped window is warped to a fixed size and aspect ratio
// square: the tightest square around the window is cropped
optional string crop_mode = 11 [default = "warp"];
// cache_images: will load all images in memory for faster access
optional bool cache_images = 12 [default = false];
// append root_folder to locate images
optional string root_folder = 13 [default = ""];
}
message SPPParameter {
enum PoolMethod {
MAX = 0;
AVE = 1;
STOCHASTIC = 2;
}
optional uint32 pyramid_height = 1;
optional PoolMethod pool = 2 [default = MAX]; // The pooling method
enum Engine {
DEFAULT = 0;
CAFFE = 1;
CUDNN = 2;
}
optional Engine engine = 6 [default = DEFAULT];
}
// DEPRECATED: use LayerParameter.
message V1LayerParameter {
repeated string bottom = 2;
repeated string top = 3;
optional string name = 4;
repeated NetStateRule include = 32;
repeated NetStateRule exclude = 33;
enum LayerType {
NONE = 0;
ABSVAL = 35;
ACCURACY = 1;
ARGMAX = 30;
BNLL = 2;
CONCAT = 3;
CONTRASTIVE_LOSS = 37;
CONVOLUTION = 4;
DATA = 5;
DECONVOLUTION = 39;
DROPOUT = 6;
DUMMY_DATA = 32;
EUCLIDEAN_LOSS = 7;
ELTWISE = 25;
EXP = 38;
FLATTEN = 8;
HDF5_DATA = 9;
HDF5_OUTPUT = 10;
HINGE_LOSS = 28;
IM2COL = 11;
IMAGE_DATA = 12;
INFOGAIN_LOSS = 13;
INNER_PRODUCT = 14;
LRN = 15;
MEMORY_DATA = 29;
MULTINOMIAL_LOGISTIC_LOSS = 16;
MVN = 34;
POOLING = 17;
POWER = 26;
RELU = 18;
SIGMOID = 19;
SIGMOID_CROSS_ENTROPY_LOSS = 27;
SILENCE = 36;
SOFTMAX = 20;
SOFTMAX_LOSS = 21;
SPLIT = 22;
SLICE = 33;
TANH = 23;
WINDOW_DATA = 24;
THRESHOLD = 31;
}
optional LayerType type = 5;
repeated BlobProto blobs = 6;
repeated string param = 1001;
repeated DimCheckMode blob_share_mode = 1002;
enum DimCheckMode {
STRICT = 0;
PERMISSIVE = 1;
}
repeated float blobs_lr = 7;
repeated float weight_decay = 8;
repeated float loss_weight = 35;
optional AccuracyParameter accuracy_param = 27;
optional ArgMaxParameter argmax_param = 23;
optional ConcatParameter concat_param = 9;
optional ContrastiveLossParameter contrastive_loss_param = 40;
optional ConvolutionParameter convolution_param = 10;
optional DataParameter data_param = 11;
optional DropoutParameter dropout_param = 12;
optional DummyDataParameter dummy_data_param = 26;
optional EltwiseParameter eltwise_param = 24;
optional ExpParameter exp_param = 41;
optional HDF5DataParameter hdf5_data_param = 13;
optional HDF5OutputParameter hdf5_output_param = 14;
optional HingeLossParameter hinge_loss_param = 29;
optional ImageDataParameter image_data_param = 15;
optional InfogainLossParameter infogain_loss_param = 16;
optional InnerProductParameter inner_product_param = 17;
optional LRNParameter lrn_param = 18;
optional MemoryDataParameter memory_data_param = 22;
optional MVNParameter mvn_param = 34;
optional PoolingParameter pooling_param = 19;
optional PowerParameter power_param = 21;
optional ReLUParameter relu_param = 30;
optional SigmoidParameter sigmoid_param = 38;
optional SoftmaxParameter softmax_param = 39;
optional SliceParameter slice_param = 31;
optional TanHParameter tanh_param = 37;
optional ThresholdParameter threshold_param = 25;
optional WindowDataParameter window_data_param = 20;
optional TransformationParameter transform_param = 36;
optional LossParameter loss_param = 42;
optional V0LayerParameter layer = 1;
}
// DEPRECATED: V0LayerParameter is the old way of specifying layer parameters
// in Caffe. We keep this message type around for legacy support.
message V0LayerParameter {
optional string name = 1; // the layer name
optional string type = 2; // the string to specify the layer type
// Parameters to specify layers with inner products.
optional uint32 num_output = 3; // The number of outputs for the layer
optional bool biasterm = 4 [default = true]; // whether to have bias terms
optional FillerParameter weight_filler = 5; // The filler for the weight
optional FillerParameter bias_filler = 6; // The filler for the bias
optional uint32 pad = 7 [default = 0]; // The padding size
optional uint32 kernelsize = 8; // The kernel size
optional uint32 group = 9 [default = 1]; // The group size for group conv
optional uint32 stride = 10 [default = 1]; // The stride
enum PoolMethod {
MAX = 0;
AVE = 1;
STOCHASTIC = 2;
}
optional PoolMethod pool = 11 [default = MAX]; // The pooling method
optional float dropout_ratio = 12 [default = 0.5]; // dropout ratio
optional uint32 local_size = 13 [default = 5]; // for local response norm
optional float alpha = 14 [default = 1.]; // for local response norm
optional float beta = 15 [default = 0.75]; // for local response norm
optional float k = 22 [default = 1.];
// For data layers, specify the data source
optional string source = 16;
// For data pre-processing, we can do simple scaling and subtracting the
// data mean, if provided. Note that the mean subtraction is always carried
// out before scaling.
optional float scale = 17 [default = 1];
optional string meanfile = 18;
// For data layers, specify the batch size.
optional uint32 batchsize = 19;
// For data layers, specify if we would like to randomly crop an image.
optional uint32 cropsize = 20 [default = 0];
// For data layers, specify if we want to randomly mirror data.
optional bool mirror = 21 [default = false];
// The blobs containing the numeric parameters of the layer
repeated BlobProto blobs = 50;
// The ratio that is multiplied on the global learning rate. If you want to
// set the learning ratio for one blob, you need to set it for all blobs.
repeated float blobs_lr = 51;
// The weight decay that is multiplied on the global weight decay.
repeated float weight_decay = 52;
// The rand_skip variable is for the data layer to skip a few data points
// to avoid all asynchronous sgd clients to start at the same point. The skip
// point would be set as rand_skip * rand(0,1). Note that rand_skip should not
// be larger than the number of keys in the database.
optional uint32 rand_skip = 53 [default = 0];
// Fields related to detection (det_*)
// foreground (object) overlap threshold
optional float det_fg_threshold = 54 [default = 0.5];
// background (non-object) overlap threshold
optional float det_bg_threshold = 55 [default = 0.5];
// Fraction of batch that should be foreground objects
optional float det_fg_fraction = 56 [default = 0.25];
// optional bool OBSOLETE_can_clobber = 57 [default = true];
// Amount of contextual padding to add around a window
// (used only by the window_data_layer)
optional uint32 det_context_pad = 58 [default = 0];
// Mode for cropping out a detection window
// warp: cropped window is warped to a fixed size and aspect ratio
// square: the tightest square around the window is cropped
optional string det_crop_mode = 59 [default = "warp"];
// For ReshapeLayer, one needs to specify the new dimensions.
optional int32 new_num = 60 [default = 0];
optional int32 new_channels = 61 [default = 0];
optional int32 new_height = 62 [default = 0];
optional int32 new_width = 63 [default = 0];
// Whether or not ImageLayer should shuffle the list of files at every epoch.
// It will also resize images if new_height or new_width are not zero.
optional bool shuffle_images = 64 [default = false];
// For ConcatLayer, one needs to specify the dimension for concatenation, and
// the other dimensions must be the same for all the bottom blobs.
// By default it will concatenate blobs along the channels dimension.
optional uint32 concat_dim = 65 [default = 1];
optional HDF5OutputParameter hdf5_output_param = 1001;
}
message PReLUParameter {
// Parametric ReLU described in K. He et al, Delving Deep into Rectifiers:
// Surpassing Human-Level Performance on ImageNet Classification, 2015.
// Initial value of a_i. Default is a_i=0.25 for all i.
optional FillerParameter filler = 1;
// Whether or not slope paramters are shared across channels.
optional bool channel_shared = 2 [default = false];
}
message AxpyParameter{
}
message UpsampleParameter{
optional int32 scale = 1 [default = 1];
}
message ROIPoolingParameter {
// Pad, kernel size, and stride are all given as a single value for equal
// dimensions in height and width or as Y, X pairs.
optional uint32 pooled_h = 1 [default = 0]; // The pooled output height
optional uint32 pooled_w = 2 [default = 0]; // The pooled output width
// Multiplicative spatial scale factor to translate ROI coords from their
// input scale to the scale used when pooling
optional float spatial_scale = 3 [default = 1];
}
message ShuffleChannelParameter {
optional uint32 group = 1[default = 1]; // The number of group
}
x2paddle/x2paddle/decoder/caffe_decoder.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import sys
from google.protobuf import text_format
import numpy as np
from x2paddle.core.graph import GraphNode, Graph
from x2paddle.core.fluid_code import FluidCode
from x2paddle.op_mapper import caffe_shape
class CaffeResolver(object):
def __init__(self, caffe_proto):
self.caffe_proto = caffe_proto
self.import_caffe()
def import_caffepb(self):
if self.caffe_proto is None:
from x2paddle.decoder import caffe_pb2
out = caffe_pb2
else:
if not os.path.isfile(self.caffe_proto):
raise Exception(
"The .py file compiled by caffe.proto is not exist.")
(filepath,
tempfilename) = os.path.split(os.path.abspath(self.caffe_proto))
(filename, extension) = os.path.splitext(tempfilename)
sys.path.append(filepath)
out = __import__(filename)
return out
def import_caffe(self):
self.caffepb = self.import_caffepb()
self.NetParameter = self.caffepb.NetParameter
class CaffeGraphNode(GraphNode):
def __init__(self, layer, type_str, layer_name=None):
if layer_name is None:
super(CaffeGraphNode, self).__init__(
layer, layer.name.replace('/', '_').replace('-', '_'))
else:
super(CaffeGraphNode, self).__init__(
layer, layer_name.replace('/', '_').replace('-', '_'))
self.layer_type = type_str
self.fluid_code = FluidCode()
self.data = None
def set_params(self, params):
self.data = params
class CaffeGraph(Graph):
def __init__(self, model, params, caffe_pb):
self.params = params
self.caffe_pb = caffe_pb
super(CaffeGraph, self).__init__(model)
def filter_layers(self, layers):
'''Filter out layers based on the current phase.'''
phase_map = {0: 'train', 1: 'test'}
filtered_layer_names = set()
filtered_layers = []
for layer in layers:
if hasattr(layer, 'input'):
continue
type_str = self.get_layer_type(layer)
phase = 'test'
if len(layer.include):
phase = phase_map[layer.include[0].phase]
if len(layer.exclude):
phase = phase_map[1 - layer.include[0].phase]
exclude = (phase != 'test')
# Dropout layers appear in a fair number of Caffe
# test-time networks. These are just ignored. We'll
# filter them out here.
if (not exclude) and (phase == 'test'):
exclude = (type_str == 'Dropout')
if layer.type == 'Dropout':
drop_layer_top = layer.top[0]
drop_layer_bottom = layer.bottom[0]
if drop_layer_top != drop_layer_bottom:
for next_layer in layers:
for next_layer_bottom_idx, next_layer_bottom in enumerate(
next_layer.bottom):
if drop_layer_top == next_layer_bottom:
next_layer.bottom.remove(drop_layer_top)
next_layer.bottom.insert(
next_layer_bottom_idx,
drop_layer_bottom)
if not exclude:
filtered_layers.append(layer)
# Guard against dupes.
assert layer.name not in filtered_layer_names
filtered_layer_names.add(layer.name)
else:
print('The filter layer:' + layer.name)
return filtered_layers
def generate_input_layer(self, dims, index):
dim_str = ''
for dim in dims:
dim_str += 'dim: {}\n'.format(str(dim))
input_str = 'layer {\n'
input_str += 'name: \"{}\"\n '.format(str(self.model.input[index]))
input_str += 'type: "Input"\n'
input_str += 'top: \"{}\"\n'.format(str(self.model.input[index]))
input_str += 'input_param {\n'
input_str += 'shape {\n'
input_str += dim_str
input_str += '}}}'
input_str = str.encode(input_str)
net = self.caffe_pb.NetParameter()
text_format.Merge(input_str, net)
return net.layers or net.layer
def input2layers(self, input_layers=[]):
inputs_num = len(self.model.input)
if inputs_num != 0:
input_dims_num = len(self.model.input_dim)
if input_dims_num != 0:
if input_dims_num > 0 and input_dims_num != inputs_num * 4:
raise Error('invalid input_dim[%d] param in prototxt' %
(input_dims_num))
for i in range(inputs_num):
dims = self.model.input_dim[i * 4:(i + 1) * 4]
l = self.generate_input_layer(dims, i)
input_layers.append(l[0])
else:
for i in range(inputs_num):
dims = self.model.input_shape[i].dim[0:4]
l = self.generate_input_layer(dims, i)
input_layers.append(l[0])
def transform_input_layers(self, layers, input_layers=[]):
for layer in layers:
if hasattr(layer, 'input'):
input_dims_num = len(layers.input_dim)
if input_dims_num > 0 and input_dims_num != 4:
raise Error('invalid input_dim[%d] param in prototxt' %
(input_dims_num))
dims = self.model.input_dim[0:4]
l = self.generate_input_layer(dims, i)
input_layers.append(l[0])
def get_layer_type(self, layer):
if isinstance(layer.type, int):
enum_values = self.caffe_pb._V1LAYERPARAMETER_LAYERTYPE.values
vals = [val for val in enum_values if val.number == layer.type]
part = vals[0].name.split('_')
part = [s.capitalize() for s in part]
type_str = ''
type_str = type_str.join(part)
if 'relu' in type_str.lower():
type_str = type_str.replace('elu', 'eLU')
elif type_str.lower() == 'lrn':
type_str = 'LRN'
return type_str
else:
return layer.type
def build(self):
layers = self.model.layers or self.model.layer
layers = self.filter_layers(layers)
input_layers = []
self.input2layers(input_layers)
self.transform_input_layers(layers, input_layers)
layers = input_layers + layers
for layer in layers:
if hasattr(layer, 'name'):
name = getattr(layer, 'name')
setattr(layer, 'name', name.replace('/', '_').replace('-', '_'))
for i, name in enumerate(layer.bottom):
layer.bottom[i] = name.replace('/', '_').replace('-', '_')
for i, name in enumerate(layer.top):
layer.top[i] = name.replace('/', '_').replace('-', '_')
top_layer = {}
for layer in layers:
if hasattr(layer, 'input'):
continue
type_str = self.get_layer_type(layer)
self.node_map[layer.name] = CaffeGraphNode(layer, type_str)
for in_name in layer.bottom:
if in_name in top_layer:
self.connect(top_layer[in_name][-1], layer.name)
else:
raise Exception(
'input[{}] of node[{}] does not exist in node_map'.
format(in_name, layer.name))
for out_name in layer.top:
if out_name not in top_layer:
top_layer[out_name] = [layer.name]
else:
top_layer[out_name].append(layer.name)
for layer_name, data in self.params:
if layer_name in self.node_map:
node = self.node_map[layer_name]
node.set_params(data)
else:
print('Ignoring parameters for non-existent layer: %s' % \
layer_name)
super(CaffeGraph, self).build()
def get_bottom_node(self, node, idx=0, copy=False):
input_node_name = node.inputs[idx]
assert input_node_name in self.node_map, 'The {} isn\'t a valid node'.format(
name)
input_node = self.node_map[input_node_name]
if len(input_node.layer.top) > 1:
need_idx = list(input_node.layer.top).index(node.layer.bottom[idx])
name = input_node_name + ':' + str(need_idx)
else:
name = input_node_name
return self.get_node(name, copy=copy)
class CaffeDecoder(object):
def __init__(self, proto_path, model_path, caffe_proto):
self.proto_path = proto_path
self.model_path = model_path
self.resolver = CaffeResolver(caffe_proto=caffe_proto)
self.net = self.resolver.NetParameter()
with open(proto_path, 'rb') as proto_file:
proto_str = proto_file.read()
text_format.Merge(proto_str, self.net)
self.load_using_pb()
self.caffe_graph = CaffeGraph(self.net, self.params,
self.resolver.caffepb)
self.caffe_graph.build()
def load_using_pb(self):
data = self.resolver.NetParameter()
data.MergeFromString(open(self.model_path, 'rb').read())
layers = data.layers or data.layer
for layer in layers:
setattr(layer, 'name',
layer.name.replace('/', '_').replace('-', '_'))
pair = lambda layer: (layer.name, self.normalize_pb_data(layer))
self.params = [pair(layer) for layer in layers if layer.blobs]
def normalize_pb_data(self, layer):
transformed = []
for blob in layer.blobs:
if len(blob.shape.dim):
dims = blob.shape.dim
if layer.type == 'PReLU':
c_o, c_i, h, w = map(int, [1] + \
list(dims) + [1]* (3 - len(dims)))
elif layer.type == 'Normalize' and len(dims) == 4:
data = np.asarray(list(blob.data), dtype=np.float32)
transformed.append(data)
continue
else:
c_o, c_i, h, w = map(int,
[1] * (4 - len(dims)) + list(dims))
else:
c_o = blob.num
c_i = blob.channels
h = blob.height
w = blob.width
data = np.asarray(
list(blob.data), dtype=np.float32).reshape(c_o, c_i, h, w)
transformed.append(data)
return transformed
x2paddle/x2paddle/decoder/caffe_pb2.py 0 → 100644
浏览文件 @ cf7f9b88
因为 它太大了无法显示 source diff 。你可以改为 查看blob。
x2paddle/x2paddle/decoder/onnx_decoder.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from x2paddle.core.graph import GraphNode, Graph
from x2paddle.core.fluid_code import FluidCode
from x2paddle.decoder.onnx_shape_inference import SymbolicShapeInference
from onnx.checker import ValidationError
from onnx.checker import check_model
from onnx.utils import polish_model
from onnx import helper
from onnx.helper import get_attribute_value, make_attribute
from onnx.shape_inference import infer_shapes
from onnx.mapping import TENSOR_TYPE_TO_NP_TYPE
from onnx.numpy_helper import to_array
from onnx import AttributeProto, TensorProto, GraphProto
from collections import OrderedDict as Dict
import onnx
from onnx.helper import ValueInfoProto
import numpy as np
from copy import deepcopy
import logging as _logging
import os
default_op_domain = 'ai.onnx'
_logger = _logging.getLogger(__name__)
class ONNXGraphNode(GraphNode):
def __init__(self, layer, layer_name=None):
if layer_name is None:
super(ONNXGraphNode, self).__init__(layer, layer.name)
else:
super(ONNXGraphNode, self).__init__(layer, layer_name)
self.layer_type = layer.op_type
self.fluid_code = FluidCode()
self.attr_map = self.get_attr_map()
self.out_shapes = list()
self.dtype = None
self.which_child = {}
def get_attr_map(self):
"""
convert ONNX node attributes to dict
"""
return {
attr.name: self.get_attribute_value(attr)
for attr in self.layer.attribute
}
@property
def value(self):
assert 'Constant' in self.layer_type, "Only Constant | ConstantOfShape node has value."
if 'value' not in self.attr_map:
return None
return self.attr_map['value']
def get_attribute_value(self, attr):
"""
get_attribute_value enhanced
"""
if attr.type == onnx.AttributeProto.TENSOR:
dtype = np.dtype(TENSOR_TYPE_TO_NP_TYPE[attr.t.data_type])
data = attr.t.raw_data
value = np.frombuffer(
data, dtype=dtype, count=(len(data) // dtype.itemsize))
elif attr.type == onnx.AttributeProto.STRING:
value = attr.s
value = value.decode() if isinstance(value, bytes) else value
else:
value = get_attribute_value(attr)
return value
def get_attr(self, name, default=None):
"""
get_attribute_value from attr_map
"""
if name not in self.attr_map:
return default
return self.attr_map[name]
class ONNXGraphDataNode(GraphNode):
def __init__(self, layer, layer_name=None, is_global_input=False):
if layer_name is None:
super(ONNXGraphDataNode, self).__init__(layer, layer.name)
else:
super(ONNXGraphDataNode, self).__init__(layer, layer_name)
if is_global_input:
self.layer_type = 'place_holder'
else:
self.layer_type = 'create_parameter'
self.layer_name = layer_name
self.fluid_code = FluidCode()
self.weight = None
self.embeded_as = None
self.which_child = {}
@property
def out_shapes(self):
if isinstance(self.layer, ValueInfoProto):
values = self.layer.type.tensor_type.shape.dim
out_shapes = list()
out_shapes.append([dim.dim_value for dim in values])
return out_shapes
else:
values = self.layer.dims
out_shapes = list()
out_shapes.append(values)
return out_shapes
@property
def dtype(self):
if isinstance(self.layer, ValueInfoProto):
dtype = self.layer.type.tensor_type.elem_type
return TENSOR_TYPE_TO_NP_TYPE[dtype]
else:
dtype = self.layer.data_type
return TENSOR_TYPE_TO_NP_TYPE[dtype]
class ONNXGraph(Graph):
def __init__(self, onnx_model):
super(ONNXGraph, self).__init__(onnx_model)
self.fixed_input_shape = {}
self.initializer = {}
self.place_holder_nodes = list()
self.value_infos = {}
self.graph = onnx_model.graph
self.get_place_holder_nodes()
print("shape inferencing ...")
self.graph = SymbolicShapeInference.infer_shapes(
onnx_model, fixed_input_shape=self.fixed_input_shape)
print("shape inferenced.")
self.build()
self.collect_value_infos()
self.allocate_shapes()
def get_inner_nodes(self):
"""
generate inner node of ONNX model
"""
inner_nodes = []
if not isinstance(self.graph, onnx.GraphProto):
logger.error('graph is not a GraphProto instance')
return
for initializer in self.graph.initializer:
name = initializer.name
inner_nodes.append(name)
return inner_nodes
def get_symbolic_shape(self, dims):
shape = []
for dim in dims:
if dim.HasField('dim_param'):
shape.append(dim.dim_param)
else:
shape.append(dim.dim_value)
return shape
def check_input_shape(self, vi):
if vi.type.HasField('tensor_type'):
for dim in vi.type.tensor_type.shape.dim:
if dim.HasField(
'dim_param') and vi.name not in self.fixed_input_shape:
shape = self.get_symbolic_shape(
vi.type.tensor_type.shape.dim)
print(
"Unknown shape for input tensor[tensor name: '{}'] -> shape: {}, Please define shape of input here,\nNote:you can use visualization tools like Netron to check input shape."
.format(vi.name, shape))
right_shape_been_input = False
while not right_shape_been_input:
try:
shape = raw_input(
"Shape of Input(e.g. -1,3,224,224), enter 'N' to skip: "
)
except:
shape = input(
"Shape of Input(e.g. -1,3,224,224), enter 'N' to skip: "
)
if shape.count("-1") > 1:
print("Only 1 dimension can be -1, type again:)")
else:
right_shape_been_input = True
if shape == 'N':
break
shape = [int(dim) for dim in shape.strip().split(',')]
assert shape.count(-1) <= 1, "Only one dimension can be -1"
self.fixed_input_shape[vi.name] = shape
break
def get_place_holder_nodes(self):
"""
generate place_holder node of ONNX model
"""
inner_nodes = self.get_inner_nodes()
for ipt_vi in self.graph.input:
if ipt_vi.name not in inner_nodes:
self.check_input_shape(ipt_vi)
self.place_holder_nodes.append(ipt_vi.name)
def get_output_nodes(self):
"""
generate output_nodes node of ONNX model
"""
inner_nodes = self.get_inner_nodes()
output_nodes = [value.name for value in self.graph.output]
for opt_data in output_nodes:
if opt_data not in inner_nodes:
self.output_nodes.append(opt_data)
def is_place_holder_nodes(self, layer):
"""
return layer is or not place_holder node
"""
if layer in self.place_holder_nodes:
return True
return False
def build(self):
"""
build topo_sort of ONNX model
"""
for layer in self.graph.node:
node = ONNXGraphNode(layer)
self.node_map[layer.name] = node
for layer in self.graph.input:
if layer.name not in self.node_map:
is_place_holder = self.is_place_holder_nodes(layer.name)
self.node_map[layer.name] = ONNXGraphDataNode(
layer,
layer_name=layer.name,
is_global_input=is_place_holder)
#set data node's weight
for initializer in self.graph.initializer:
name = initializer.name
weight = to_array(initializer)
if name in self.node_map:
if isinstance(self.node_map[name], ONNXGraphDataNode):
self.node_map[name].weight = weight
self.node_map[name].embeded_as = []
else:
self.node_map[name] = ONNXGraphDataNode(
initializer, layer_name=name, is_global_input=False)
self.node_map[name].weight = weight
self.node_map[name].embeded_as = []
#generate connection between nodes for topo
for layer_name, node in self.node_map.items():
if isinstance(node, ONNXGraphNode):
self.build_connection(layer_name, node)
#generate topo
super(ONNXGraph, self).build()
self.input_nodes = self.place_holder_nodes
def build_connection(self, layer_name, node):
"""
find connection for nodes
"""
for idx, in_node in enumerate(node.layer.input):
if in_node == '':
continue
if in_node not in self.node_map:
flag = 0
for nd in self.graph.node:
for idx, opt in enumerate(nd.output):
if opt == in_node:
self.connect(nd.name, layer_name)
flag = 1
node.which_child[nd.name] = idx
self.node_map[nd.name].index = 0
break
if flag == 1:
break
if flag == 0:
raise Exception(
'input[{}] of node[{}] does not exist in node_map'.
format(in_node, layer_name))
else:
self.connect(in_node, layer_name)
def get_input_node(self, node, idx=0, copy=False):
if len(node.which_child) == 0:
ipt_node = super(ONNXGraph, self).get_node(node.inputs[idx], copy)
return ipt_node
else:
ipt_node = super(ONNXGraph, self).get_node(node.inputs[idx], copy)
if ipt_node.layer_name in node.which_child:
ipt_node.index = node.which_child[ipt_node.layer_name]
return ipt_node
def graph_weights(self):
"""
generator for weights
"""
if not isinstance(self.graph, onnx.GraphProto):
logger.error('graph is not a GraphProto instance')
return
for initializer in self.graph.initializer:
name = initializer.name
weight = to_array(initializer)
yield name, weight
def collect_value_infos(self):
"""
collect value/type info for an ONNX model
"""
assert isinstance(self.graph,
onnx.GraphProto), 'model is not a ModelProto instance'
for item in self.graph.value_info:
self.value_infos[item.name] = {
'dtype':
TENSOR_TYPE_TO_NP_TYPE[item.type.tensor_type.elem_type],
'shape':
[dim.dim_value for dim in item.type.tensor_type.shape.dim],
'external': False
}
def allocate_shapes(self):
"""
run shape inference
"""
for layer in self.graph.node:
node = self.node_map[layer.name]
for opt in layer.output:
if opt in self.value_infos:
value_info = self.value_infos[opt]
#if len(value_info['shape']) == 0 or value_info[
# 'dtype'] is None or 0 in value_info['shape']:
# #TODO add node shape inference
node.dtype = value_info['dtype']
node.out_shapes.append(value_info['shape'])
else:
node.out_shapes.append([])
class ONNXDecoder(object):
def __init__(self, onnx_model):
onnx_model = onnx.load(onnx_model)
print('model ir_version: {}, op version: {}'.format(
onnx_model.ir_version, onnx_model.opset_import[0].version))
self.op_set = onnx_model.opset_import[0].version
check_model(onnx_model)
onnx_model = self.optimize_model_skip_op(onnx_model)
onnx_model = self.optimize_model_strip_initializer(onnx_model)
onnx_model = self.optimize_node_name(onnx_model)
self.graph = ONNXGraph(onnx_model)
#self.onnx_model = onnx_model
def build_value_refs(self, nodes):
"""
build op reference of inputs and outputs
"""
input_refs = Dict()
output_refs = Dict()
for idx, node in enumerate(nodes):
for val_name in node.input:
input_refs.setdefault(val_name, set()).add(idx)
for val_name in node.output:
output_refs.setdefault(val_name, set()).add(idx)
return input_refs, output_refs
def skip_node_forward(self, nodes, src_output_name, dst_input_name,
input_refs):
"""
skip nodes between src_output_name -> dst_input_name and connect this pair
"""
processed = 0
for next_idx in input_refs[src_output_name]:
next_node = nodes[next_idx]
for val_idx, next_input_name in enumerate(next_node.input):
if next_input_name == src_output_name:
next_node.input[val_idx] = dst_input_name
processed += 1
return processed
def skip_node_backward(self, nodes, src_input_name, dst_output_name,
output_refs):
"""
skip nodes between dst_output_name -> src_input_name and connect this pair
"""
processed = 0
for prev_idx in output_refs[src_input_name]:
prev_node = nodes[prev_idx]
for val_idx, prev_output_name in enumerate(prev_node.output):
if prev_output_name == src_input_name:
prev_node.output[val_idx] = dst_output_name
processed += 1
return processed
def optimize_model_skip_op(self, model, op_list=None):
"""
skip ops can be bypassed for inference
"""
nodes = model.graph.node
if op_list is None:
op_list = ['Dropout']
input_refs, output_refs = self.build_value_refs(nodes)
ret = type(model)()
ret.CopyFrom(model)
ret_nodes = ret.graph.node
nodes_to_remove = []
for node_idx, node in enumerate(nodes):
if not (node.domain == default_op_domain or node.domain == ''):
continue
op_type = node.op_type
if not (op_type in op_list):
continue
if op_type in ['Dropout']:
input_name = node.input[0]
output_name = node.output[0]
elif not (len(node.input) == 1 and len(node.output) == 1):
print(
'currently only 1-input-1-output op supported, skip required %d: %s',
node_idx, node.op_type)
continue
else:
input_name = node.input[0]
output_name = node.output[0]
if output_name in input_refs:
processed = self.skip_node_forward(ret_nodes, output_name,
input_name, input_refs)
elif input_name in output_refs:
processed = self.skip_node_backward(ret_nodes, input_name,
output_name, output_refs)
else:
processed = -1
if processed > 0:
nodes_to_remove.append(node_idx)
for value_info in ret.graph.value_info:
for output in node.output:
if value_info.name == output:
ret.graph.value_info.remove(value_info)
print('skip op {}: {} -> {} -> {}'.format(
node_idx, input_name, node.op_type, output_name))
elif processed == 0:
print('weird, no node processed')
else:
print('standalone op {}: {} -> {} -> {} not skipped'.format(
node_idx, input_name, node.op_type, output_name))
nodes_to_remove.sort(reverse=True)
for node_idx in nodes_to_remove:
ret_nodes.pop(node_idx)
return ret
def optimize_model_strip_initializer(self, model, keep_input_only=True):
"""
strip weights for inference
"""
nodes = model.graph.node
input_refs, output_refs = self.build_value_refs(nodes)
out_names = [val.name for val in model.graph.output]
ret = type(model)()
ret.CopyFrom(model)
# strip initializers
ret.graph.ClearField('initializer')
ret_initializers = ret.graph.initializer
for initializer in model.graph.initializer:
name = initializer.name
if name in input_refs:
ret_initializers.add().CopyFrom(initializer)
elif not keep_input_only and name in output_refs:
ret_initializers.add().CopyFrom(initializer)
else:
dtype = TENSOR_TYPE_TO_NP_TYPE[initializer.data_type]
# strip inputs
ret.graph.ClearField('input')
ret_inputs = ret.graph.input
for item in model.graph.input:
name = item.name
if name in input_refs or name in out_names:
ret_inputs.add().CopyFrom(item)
return ret
def make_variable_name(self, name):
"""
make a valid code name for ParamAttr
"""
if name == '':
raise ValueError('name should not be empty')
for s in ' .*?\\/-:':
name = name.replace(s, '_')
return 'x2paddle_' + name
def optimize_node_name(self, model):
"""
standardize variable name for paddle's code
"""
graph = model.graph
for initializer in graph.initializer:
initializer.name = self.make_variable_name(initializer.name)
for ipt in graph.input:
ipt.name = self.make_variable_name(ipt.name)
for output in graph.output:
output.name = self.make_variable_name(output.name)
for item in graph.value_info:
item.name = self.make_variable_name(item.name)
for node in graph.node:
node.name = node.output[0]
node.name = self.make_variable_name(node.name)
for i in range(len(node.input)):
if node.input[i] == '':
continue
else:
node.input[i] = self.make_variable_name(node.input[i])
for i in range(len(node.output)):
node.output[i] = self.make_variable_name(node.output[i])
return model
x2paddle/x2paddle/decoder/onnx_shape_inference.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Reference Code from https://github.com/microsoft/onnxruntime, Licensed under the MIT License.
# -*- coding: UTF-8 -*-
import argparse
import numpy as np
import onnx
import sys
from onnx import helper, numpy_helper, shape_inference
import sympy
from packaging import version
assert version.parse(onnx.__version__) >= version.parse("1.5.0")
def get_attribute(node, attr_name, default_value=None):
found = [attr for attr in node.attribute if attr.name == attr_name]
if found:
return helper.get_attribute_value(found[0])
return default_value
def get_dim_from_type_proto(dim):
return getattr(dim, dim.WhichOneof('value')) if type(
dim.WhichOneof('value')) == str else None
def get_shape_from_type_proto(type_proto):
return [
get_dim_from_type_proto(d) for d in type_proto.tensor_type.shape.dim
]
def get_shape_from_sympy_shape(sympy_shape):
sympy_shape = [
None if i is None else (int(i) if is_literal(i) else str(i))
for i in sympy_shape
]
return sympy_shape
def is_literal(dim):
return type(dim) in [int, np.int64, np.int32, sympy.Integer] or (
hasattr(dim, 'is_number') and
dim.is_number) # or (hasattr(dim, 'is_integer') and dim.is_integer)
def handle_negative_axis(axis, rank):
assert axis < rank and axis >= -rank
return axis if axis >= 0 else rank + axis
def get_opset(mp, domain=['', 'onnx', 'ai.onnx']):
if type(domain) != list:
domain = [domain]
for opset in mp.opset_import:
if opset.domain in domain:
return opset.version
return None
def as_scalar(x):
if type(x) == list:
assert len(x) == 1
return x[0]
elif type(x) == np.ndarray:
return np.asscalar(x)
else:
return x
def as_list(x, keep_none):
if type(x) == list:
return x
elif type(x) == np.ndarray:
return list(x)
elif keep_none and x is None:
return None
else:
return [x]
def sympy_reduce_product(x):
if type(x) == list:
value = sympy.Integer(1)
for v in x:
value = value * v
else:
value = x
return value
class SymbolicShapeInference:
def __init__(self, int_max, auto_merge, guess_output_rank, verbose):
self.dispatcher_ = {
'Add': self._infer_symbolic_compute_ops,
'ArrayFeatureExtractor': self._infer_ArrayFeatureExtractor,
'AveragePool': self._infer_Pool,
'Cast': self._infer_Cast,
'CategoryMapper': self._infer_CategoryMapper,
'Compress': self._infer_Compress,
'Concat': self._infer_Concat,
'ConstantOfShape': self._infer_ConstantOfShape,
'Conv': self._infer_Conv,
'CumSum': self._pass_on_shape_and_type,
'Div': self._infer_symbolic_compute_ops,
'Expand': self._infer_Expand,
'Equal': self._infer_symbolic_compute_ops,
'Gather': self._infer_Gather,
'GatherElements': self._infer_GatherElements,
'GatherND': self._infer_GatherND,
'If': self._infer_If,
'Loop': self._infer_Loop,
'MatMul': self._infer_MatMul,
'MatMulInteger16': self._infer_MatMulInteger,
'MaxPool': self._infer_Pool,
'Max': self._infer_symbolic_compute_ops,
'Min': self._infer_symbolic_compute_ops,
'Mul': self._infer_symbolic_compute_ops,
'NonMaxSuppression': self._infer_NonMaxSuppression,
'NonZero': self._infer_NonZero,
'OneHot': self._infer_OneHot,
'Pad': self._infer_Pad,
'Range': self._infer_Range,
'ReduceProd': self._infer_ReduceProd,
'Reshape': self._infer_Reshape,
'Resize': self._infer_Resize,
'Round': self._pass_on_shape_and_type,
'Scan': self._infer_Scan,
'ScatterElements': self._infer_ScatterElements,
'Shape': self._infer_Shape,
'Size': self._infer_Size,
'Slice': self._infer_Slice,
'Split': self._infer_Split,
'Squeeze': self._infer_Squeeze,
'Sub': self._infer_symbolic_compute_ops,
'Tile': self._infer_Tile,
'TopK': self._infer_TopK,
'Unsqueeze': self._infer_Unsqueeze,
'Where': self._infer_symbolic_compute_ops,
'Transpose': self._infer_Transpose,
'ZipMap': self._infer_ZipMap
}
self.run_ = True
self.suggested_merge_ = {}
self.symbolic_dims_ = {}
self.input_symbols_ = {}
self.auto_merge_ = auto_merge
self.guess_output_rank_ = guess_output_rank
self.verbose_ = verbose
self.int_max_ = int_max
def _add_suggested_merge(self, symbols, apply=False):
assert all([(type(s) == str and s in self.symbolic_dims_) or
is_literal(s) for s in symbols])
symbols = set(symbols)
for k, v in self.suggested_merge_.items():
if k in symbols:
symbols.remove(k)
symbols.add(v)
map_to = None
# if there is literal, map to it first
for s in symbols:
if is_literal(s):
map_to = s
break
# when no literals, map to input symbolic dims, then existing symbolic dims
if map_to is None:
for s in symbols:
if s in self.input_symbols_:
map_to = s
break
if map_to is None:
for s in symbols:
if type(self.symbolic_dims_[s]) == sympy.Symbol:
map_to = s
break
# when nothing to map to, use the shorter one
if map_to is None:
if self.verbose_ > 0:
print(
'Potential unsafe merge between symbolic expressions: ({})'.
format(','.join(symbols)))
symbols_list = list(symbols)
lens = [len(s) for s in symbols_list]
map_to = symbols_list[lens.index(min(lens))]
symbols.remove(map_to)
for s in symbols:
if s == map_to:
continue
if is_literal(map_to) and is_literal(s):
assert int(map_to) == int(s)
self.suggested_merge_[s] = int(map_to) if is_literal(
map_to) else map_to
for k, v in self.suggested_merge_.items():
if v == s:
self.suggested_merge_[k] = map_to
if apply and self.auto_merge_:
self._apply_suggested_merge()
def _apply_suggested_merge(self, graph_input_only=False):
if not self.suggested_merge_:
return
for i in list(self.out_mp_.graph.input) + (
[] if graph_input_only else list(self.out_mp_.graph.value_info)):
for d in i.type.tensor_type.shape.dim:
if d.dim_param in self.suggested_merge_:
v = self.suggested_merge_[d.dim_param]
if is_literal(v):
d.dim_value = int(v)
else:
d.dim_param = v
def _preprocess(self, in_mp, input_shapes=None):
out_mp = onnx.ModelProto()
out_mp.CopyFrom(in_mp)
out_mp.graph.ClearField('node')
self.out_mp_ = out_mp
defined = set([
i.name
for i in list(in_mp.graph.input) + list(in_mp.graph.initializer)
])
pending_nodes = []
# returns True if no more ready nodes
def _insert_ready_nodes():
ready_nodes = [
pn for pn in pending_nodes
if all([i in defined for i in pn.input if i])
]
for rn in ready_nodes:
self.out_mp_.graph.node.add().CopyFrom(rn)
for o in rn.output:
defined.add(o)
pending_nodes.remove(rn)
return not ready_nodes
# constant op -> initializer, topological sort
for in_n in in_mp.graph.node:
if in_n.op_type == 'Constant':
t = get_attribute(in_n, 'value')
t.name = in_n.output[0]
self.out_mp_.graph.initializer.add().CopyFrom(t)
defined.add(t.name)
else:
pending_nodes.append(in_n)
_insert_ready_nodes()
while pending_nodes:
if _insert_ready_nodes():
break
if pending_nodes and self.verbose_ > 0:
print('SymbolicShapeInference: orphaned nodes discarded: ')
for n in pending_nodes:
print(n.op_type + ': ' + n.output[0])
if input_shapes is not None:
for input_name, shape in input_shapes.items():
for idx in range(len(self.out_mp_.graph.input)):
if self.out_mp_.graph.input[idx].name == input_name:
value_info = self.out_mp_.graph.input[idx]
del self.out_mp_.graph.input[idx]
self.out_mp_.graph.input.append(
helper.make_tensor_value_info(
value_info.name,
value_info.type.tensor_type.elem_type, shape))
self.initializers_ = dict(
[(i.name, i) for i in self.out_mp_.graph.initializer])
self.known_vi_ = dict(
[(i.name, i) for i in list(self.out_mp_.graph.input)])
self.known_vi_.update(
dict([(i.name, helper.make_tensor_value_info(i.name, i.data_type,
list(i.dims)))
for i in self.out_mp_.graph.initializer]))
def _merge_symbols(self, dims):
if not all([type(d) == str for d in dims]):
if self.auto_merge_:
assert len(
dims
) == 2 # only allow symbol->int merge in binary ops for now
is_int = [is_literal(d) for d in dims]
if sum(is_int) == 1:
int_dim = is_int.index(1)
if self.verbose_ > 0:
print('dim {} has been merged with value {}'.format(
dims[1 - int_dim], dims[int_dim]))
self._check_merged_dims(dims, allow_broadcast=False)
return dims[int_dim]
else:
if self.verbose_ > 0:
print('dim {} has been mergd with dim {}'.format(dims[
0], dims[1]))
return dims[0]
else:
return None
if all([d == dims[0] for d in dims]):
return dims[0]
merged = [
self.suggested_merge_[d] if d in self.suggested_merge_ else d
for d in dims
]
if all([d == merged[0] for d in merged]):
assert merged[0] in self.symbolic_dims_
return merged[0]
else:
return None
# broadcast from right to left, and merge symbolic dims if needed
def _broadcast_shapes(self, shape1, shape2):
new_shape = []
rank1 = len(shape1)
rank2 = len(shape2)
new_rank = max(rank1, rank2)
for i in range(new_rank):
dim1 = shape1[rank1 - 1 - i] if i < rank1 else 1
dim2 = shape2[rank2 - 1 - i] if i < rank2 else 1
if dim1 == 1 or dim1 == dim2:
new_dim = dim2
elif dim2 == 1:
new_dim = dim1
else:
new_dim = self._merge_symbols([dim1, dim2])
if not new_dim:
# warning about unsupported broadcast when not auto merge
# note that auto merge has the risk of incorrectly merge symbols while one of them being 1
# for example, 'a' = 1, 'b' = 5 at runtime is valid broadcasting, but with auto merge 'a' == 'b'
if self.auto_merge_:
self._add_suggested_merge([dim1, dim2], apply=True)
else:
print('unsupported broadcast between ' + str(dim1) + ' '
+ str(dim2))
new_shape = [new_dim] + new_shape
return new_shape
def _get_shape(self, node, idx):
name = node.input[idx]
shape = []
if name in self.known_vi_:
shape = get_shape_from_type_proto(self.known_vi_[name].type)
elif name in self.initializers_:
assert name in self.initializers_
shape = list(self.initializers_[name].dims)
return shape
def _get_initializer_value(self, node, idx):
name = node.input[idx]
if name in self.initializers_:
value = numpy_helper.to_array(self.initializers_[name])
return value
else:
return False
def _get_shape_rank(self, node, idx):
return len(self._get_shape(node, idx))
def _get_sympy_shape(self, node, idx):
sympy_shape = []
for d in self._get_shape(node, idx):
if type(d) is str:
sympy_shape.append(self.symbolic_dims_[d] if d in
self.symbolic_dims_ else sympy.Symbol(
d, integer=True))
else:
assert None != d
sympy_shape.append(d)
return sympy_shape
def _get_value(self, node, idx):
name = node.input[idx]
assert name in self.sympy_data_ or name in self.initializers_
return self.sympy_data_[
name] if name in self.sympy_data_ else numpy_helper.to_array(
self.initializers_[name])
def _try_get_value(self, node, idx):
if idx >= len(node.input):
return None
name = node.input[idx]
if name in self.sympy_data_ or name in self.initializers_:
return self._get_value(node, idx)
return None
def _update_computed_dims(self, new_sympy_shape):
for i, new_dim in enumerate(new_sympy_shape):
if not is_literal(new_dim) and not type(new_dim) == str:
str_dim = str(new_dim)
if str_dim in self.suggested_merge_:
new_sympy_shape[i] = self.symbolic_dims_[
self.suggested_merge_[str_dim]]
else:
# add new_dim if it's a computational expression
if not str(new_dim) in self.symbolic_dims_:
self.symbolic_dims_[str(new_dim)] = new_dim
def _onnx_infer_single_node(self, node):
# skip onnx shape inference for Scan/Loop
skip_infer = node.op_type in ['Scan', 'Loop']
if not skip_infer:
# run single node inference with self.known_vi_ shapes
# note that inference rely on initializer values is not handled
# as we don't copy initializer weights to tmp_graph for inference speed purpose
tmp_graph = helper.make_graph(
[node], 'tmp', [self.known_vi_[i] for i in node.input if i], [
helper.make_tensor_value_info(i, onnx.TensorProto.UNDEFINED,
None) for i in node.output
])
self.tmp_mp_.graph.CopyFrom(tmp_graph)
self.tmp_mp_ = shape_inference.infer_shapes(self.tmp_mp_)
for i_o in range(len(node.output)):
o = node.output[i_o]
vi = self.out_mp_.graph.value_info.add()
if not skip_infer:
vi.CopyFrom(self.tmp_mp_.graph.output[i_o])
self.known_vi_[o] = vi
def _onnx_infer_subgraph(self, node, subgraph, use_node_input=True):
if self.verbose_ > 2:
print('Inferencing subgraph of node {} with output({}...): {}'.
format(node.name, node.output[0], node.op_type))
# node inputs are not passed directly to the subgraph
# it's up to the node dispatcher to prepare subgraph input
# for example, with Scan/Loop, subgraph input shape would be trimmed from node input shape
# besides, inputs in subgraph could shadow implicit inputs
subgraph_inputs = set([
i.name for i in list(subgraph.initializer) + list(subgraph.input)
])
subgraph_implicit_input = set([
name for name in self.known_vi_.keys()
if not name in subgraph_inputs
])
tmp_graph = helper.make_graph(
list(subgraph.node), 'tmp',
list(subgraph.input) +
[self.known_vi_[i] for i in subgraph_implicit_input], [
helper.make_tensor_value_info(i.name,
onnx.TensorProto.UNDEFINED, None)
for i in subgraph.output
])
tmp_graph.initializer.extend([
i for i in self.out_mp_.graph.initializer
if i.name in subgraph_implicit_input
])
tmp_graph.initializer.extend(subgraph.initializer)
self.tmp_mp_.graph.CopyFrom(tmp_graph)
symbolic_shape_inference = SymbolicShapeInference(
self.int_max_, self.auto_merge_, self.guess_output_rank_,
self.verbose_)
all_shapes_inferred = False
symbolic_shape_inference._preprocess(self.tmp_mp_)
symbolic_shape_inference.suggested_merge_ = self.suggested_merge_.copy()
while symbolic_shape_inference.run_:
all_shapes_inferred = symbolic_shape_inference._infer_impl(
self.tmp_mp_, self.sympy_data_.copy())
symbolic_shape_inference._update_output_from_vi()
if use_node_input:
# if subgraph uses node input, it needs to update to merged dims
subgraph.ClearField('input')
subgraph.input.extend(
symbolic_shape_inference.out_mp_.graph.input[:len(node.input)])
subgraph.ClearField('output')
subgraph.output.extend(symbolic_shape_inference.out_mp_.graph.output)
subgraph.ClearField('value_info')
subgraph.value_info.extend(
symbolic_shape_inference.out_mp_.graph.value_info)
subgraph.ClearField('node')
subgraph.node.extend(symbolic_shape_inference.out_mp_.graph.node)
# for new symbolic dims from subgraph output, add to main graph symbolic dims
subgraph_shapes = [
get_shape_from_type_proto(o.type)
for o in symbolic_shape_inference.out_mp_.graph.output
]
subgraph_new_symbolic_dims = set([
d for s in subgraph_shapes
if s for d in s if type(d) == str and not d in self.symbolic_dims_
])
new_dims = {}
for d in subgraph_new_symbolic_dims:
assert d in symbolic_shape_inference.symbolic_dims_
new_dims[d] = symbolic_shape_inference.symbolic_dims_[d]
self.symbolic_dims_.update(new_dims)
return symbolic_shape_inference
def _get_int_values(self, node, broadcast=False):
values = [self._try_get_value(node, i) for i in range(len(node.input))]
if all([v is not None for v in values]):
# some shape compute is in floating point, cast to int for sympy
for i, v in enumerate(values):
if type(v) != np.ndarray:
continue
if len(v.shape) > 1:
new_v = None # ignore value for rank > 1
elif len(v.shape) == 0:
new_v = int(np.asscalar(v))
else:
assert len(v.shape) == 1
new_v = [int(vv) for vv in v]
values[i] = new_v
values_len = [len(v) if type(v) == list else 0 for v in values]
max_len = max(values_len)
if max_len >= 1 and broadcast:
# broadcast
for i, v in enumerate(values):
if v is None:
continue # don't broadcast if value is unknown
if type(v) == list:
if len(v) < max_len:
values[i] = v * max_len
else:
assert len(v) == max_len
else:
values[i] = [v] * max_len
return values
def _compute_on_sympy_data(self, node, op_func):
assert len(node.output) == 1
values = self._get_int_values(node, broadcast=True)
if all([v is not None for v in values]):
new_shape = []
is_list = [type(v) == list for v in values]
as_list = any(is_list)
if as_list:
data = [op_func(vs) for vs in zip(*values)]
self.sympy_data_[node.output[0]] = data
new_shape = np.array(data).shape
else:
data = op_func(values)
self.sympy_data_[node.output[0]] = data
new_shape = np.array(data).shape
vi = self.known_vi_[node.output[0]]
def _pass_on_sympy_data(self, node):
assert len(node.input) == 1 or node.op_type == 'Reshape'
self._compute_on_sympy_data(node, lambda x: x[0])
def _pass_on_shape_and_type(self, node):
vi = self.known_vi_[node.output[0]]
vi.CopyFrom(
helper.make_tensor_value_info(node.output[0], self.known_vi_[
node.input[0]].type.tensor_type.elem_type,
self._get_shape(node, 0)))
def _new_symbolic_dim(self, prefix, dim):
new_dim = '{}_d{}'.format(prefix, dim)
if new_dim in self.suggested_merge_:
v = self.suggested_merge_[new_dim]
new_dim = sympy.Integer(int(v)) if is_literal(v) else v
else:
self.symbolic_dims_[new_dim] = sympy.Symbol(new_dim, integer=True)
return new_dim
def _new_symbolic_dim_from_output(self, node, out_idx=0, dim=0):
return self._new_symbolic_dim('{}{}_o{}_'.format(
node.op_type, list(self.out_mp_.graph.node).index(node), out_idx),
dim)
def _new_symbolic_shape(self, rank, node, out_idx=0):
return [
self._new_symbolic_dim_from_output(node, out_idx, i)
for i in range(rank)
]
def _compute_conv_pool_shape(self, node):
sympy_shape = self._get_sympy_shape(node, 0)
if len(node.input) > 1:
W_shape = self._get_sympy_shape(node, 1)
rank = len(W_shape) - 2 # number of spatial axes
kernel_shape = W_shape[-rank:]
sympy_shape[1] = W_shape[0]
else:
W_shape = None
kernel_shape = get_attribute(node, 'kernel_shape')
rank = len(kernel_shape)
assert len(sympy_shape) == rank + 2
# only need to symbolic shape inference if input has symbolic dims in spatial axes
is_symbolic_dims = [not is_literal(i) for i in sympy_shape[-rank:]]
if not any(is_symbolic_dims):
shape = get_shape_from_type_proto(self.known_vi_[node.output[0]]
.type)
if len(shape) > 0:
assert len(sympy_shape) == len(shape)
sympy_shape[-rank:] = [sympy.Integer(d) for d in shape[-rank:]]
return sympy_shape
dilations = get_attribute(node, 'dilations', [1] * rank)
strides = get_attribute(node, 'strides', [1] * rank)
effective_kernel_shape = [(k - 1) * d + 1
for k, d in zip(kernel_shape, dilations)]
pads = get_attribute(node, 'pads')
if pads is None:
pads = [0] * (2 * rank)
auto_pad = get_attribute(node, 'auto_pad',
b'NOTSET').decode('utf-8')
if auto_pad != 'VALID' and auto_pad != 'NOTSET':
try:
residual = [
sympy.Mod(d, s)
for d, s in zip(sympy_shape[-rank:], strides)
]
total_pads = [
max(0, (k - s) if r == 0 else (k - r))
for k, s, r in zip(effective_kernel_shape, strides,
residual)
]
except TypeError: # sympy may throw TypeError: cannot determine truth value of Relational
total_pads = [
max(0, (k - s))
for k, s in zip(effective_kernel_shape, strides)
] # assuming no residual if sympy throws error
elif auto_pad == 'VALID':
total_pads = []
else:
total_pads = [0] * rank
else:
assert len(pads) == 2 * rank
total_pads = [p1 + p2 for p1, p2 in zip(pads[:rank], pads[rank:])]
ceil_mode = get_attribute(node, 'ceil_mode', 0)
for i in range(rank):
effective_input_size = sympy_shape[-rank + i]
if len(total_pads) > 0:
effective_input_size = effective_input_size + total_pads[i]
if ceil_mode:
strided_kernel_positions = sympy.ceiling(
(effective_input_size - effective_kernel_shape[i]) /
strides[i])
else:
strided_kernel_positions = (
effective_input_size - effective_kernel_shape[i]
) // strides[i]
sympy_shape[-rank + i] = strided_kernel_positions + 1
return sympy_shape
def _check_merged_dims(self, dims, allow_broadcast=True):
if allow_broadcast:
dims = [d for d in dims if not (is_literal(d) and int(d) <= 1)]
if not all([d == dims[0] for d in dims]):
self._add_suggested_merge(dims, apply=True)
def _compute_matmul_shape(self, node, output_dtype=None):
lhs_shape = self._get_shape(node, 0)
rhs_shape = self._get_shape(node, 1)
lhs_rank = len(lhs_shape)
rhs_rank = len(rhs_shape)
lhs_reduce_dim = 0
rhs_reduce_dim = 0
assert lhs_rank > 0 and rhs_rank > 0
if lhs_rank == 1 and rhs_rank == 1:
new_shape = []
elif lhs_rank == 1:
rhs_reduce_dim = -2
new_shape = rhs_shape[:rhs_reduce_dim] + [rhs_shape[-1]]
elif rhs_rank == 1:
lhs_reduce_dim = -1
new_shape = lhs_shape[:lhs_reduce_dim]
else:
lhs_reduce_dim = -1
rhs_reduce_dim = -2
new_shape = self._broadcast_shapes(
lhs_shape[:-2], rhs_shape[:-2]) + [lhs_shape[-2]
] + [rhs_shape[-1]]
# merge reduce dim
self._check_merged_dims(
[lhs_shape[lhs_reduce_dim], rhs_shape[rhs_reduce_dim]],
allow_broadcast=False)
if output_dtype is None:
# infer output_dtype from input type when not specified
output_dtype = self.known_vi_[node.input[
0]].type.tensor_type.elem_type
vi = self.known_vi_[node.output[0]]
vi.CopyFrom(
helper.make_tensor_value_info(node.output[0], output_dtype,
new_shape))
def _infer_ArrayFeatureExtractor(self, node):
data_shape = self._get_shape(node, 0)
indices_shape = self._get_shape(node, 1)
vi = self.known_vi_[node.output[0]]
vi.CopyFrom(
helper.make_tensor_value_info(node.output[0], self.known_vi_[
node.input[0]].type.tensor_type.elem_type, data_shape[:-1] +
indices_shape))
def _infer_symbolic_compute_ops(self, node):
funcs = {
'Add': lambda l: l[0] + l[1],
'Div': lambda l: l[0] // l[1], # integer div in sympy
'Equal': lambda l: l[0] == l[1],
'Max':
lambda l: l[1] if is_literal(l[0]) and int(l[0]) < -self.int_max_ else (l[0] if is_literal(l[1]) and int(l[1]) < -self.int_max_ else sympy.Max(l[0], l[1])),
'Min':
lambda l: l[1] if is_literal(l[0]) and int(l[0]) > self.int_max_ else (l[0] if is_literal(l[1]) and int(l[1]) > self.int_max_ else sympy.Min(l[0], l[1])),
'Mul': lambda l: l[0] * l[1],
'Sub': lambda l: l[0] - l[1],
'Where': lambda l: l[1] if l[0] else l[2]
}
assert node.op_type in funcs
self._compute_on_sympy_data(node, funcs[node.op_type])
def _infer_Cast(self, node):
self._pass_on_sympy_data(node)
def _infer_CategoryMapper(self, node):
input_type = self.known_vi_[node.input[0]].type.tensor_type.elem_type
if input_type == onnx.TensorProto.STRING:
output_type = onnx.TensorProto.INT64
else:
output_type = onnx.TensorProto.STRING
vi = self.known_vi_[node.output[0]]
vi.CopyFrom(
helper.make_tensor_value_info(node.output[0], output_type,
self._get_shape(node, 0)))
def _infer_Transpose(self, node):
input_shape = self._get_shape(node, 0)
perm = get_attribute(node, 'perm')
output_shape = np.array(input_shape)[perm].tolist()
vi = self.known_vi_[node.output[0]]
vi.CopyFrom(
helper.make_tensor_value_info(node.output[0], self.known_vi_[
node.input[0]].type.tensor_type.elem_type, output_shape))
def _infer_Compress(self, node):
input_shape = self._get_shape(node, 0)
# create a new symbolic dimension for Compress output
compress_len = self._new_symbolic_dim_from_output(node)
axis = get_attribute(node, 'axis')
if axis == None:
# when axis is not specified, input is flattened before compress so output is 1D
output_shape = [compress_len]
else:
output_shape = input_shape
output_shape[handle_negative_axis(axis, len(
input_shape))] = compress_len
vi = self.known_vi_[node.output[0]]
vi.CopyFrom(
helper.make_tensor_value_info(node.output[0], self.known_vi_[
node.input[0]].type.tensor_type.elem_type, output_shape))
def _infer_Concat(self, node):
if any([i in self.sympy_data_ for i in node.input]):
values = self._get_int_values(node)
if all([v is not None for v in values]):
assert 0 == get_attribute(node, 'axis')
self.sympy_data_[node.output[0]] = []
for i in range(len(node.input)):
value = values[i]
if type(value) == list:
self.sympy_data_[node.output[0]].extend(value)
else:
self.sympy_data_[node.output[0]].append(value)
sympy_shape = self._get_sympy_shape(node, 0)
axis = handle_negative_axis(
get_attribute(node, 'axis'), len(sympy_shape))
for i_idx in range(1, len(node.input)):
input_shape = self._get_sympy_shape(node, i_idx)
if input_shape:
sympy_shape[axis] = sympy_shape[axis] + input_shape[axis]
self._update_computed_dims(sympy_shape)
# merge symbolic dims for non-concat axes
for d in range(len(sympy_shape)):
if d == axis:
continue
dims = [
self._get_shape(node, i_idx)[d]
for i_idx in range(len(node.input))
if self._get_shape(node, i_idx)
]
if all([d == dims[0] for d in dims]):
continue
merged = self._merge_symbols(dims)
if type(merged) == str:
sympy_shape[d] = self.symbolic_dims_[merged] if merged else None
else:
sympy_shape[d] = merged
vi = self.known_vi_[node.output[0]]
vi.CopyFrom(
helper.make_tensor_value_info(
node.output[0], self.known_vi_[node.input[0]].type.tensor_type.
elem_type, get_shape_from_sympy_shape(sympy_shape)))
def _infer_Conv(self, node):
sympy_shape = self._compute_conv_pool_shape(node)
self._update_computed_dims(sympy_shape)
vi = self.known_vi_[node.output[0]]
vi.CopyFrom(
helper.make_tensor_value_info(
node.output[0], vi.type.tensor_type.elem_type,
get_shape_from_sympy_shape(sympy_shape)))
def _infer_ConstantOfShape(self, node):
sympy_shape = self._get_int_values(node)[0]
vi = self.known_vi_[node.output[0]]
if sympy_shape is not None:
if type(sympy_shape) != list:
sympy_shape = [sympy_shape]
self._update_computed_dims(sympy_shape)
# update sympy data if output type is int, and shape is known
if vi.type.tensor_type.elem_type == onnx.TensorProto.INT64 and all(
[is_literal(x) for x in sympy_shape]):
self.sympy_data_[node.output[0]] = np.ones(
[int(x) for x in sympy_shape],
dtype=np.int64) * numpy_helper.to_array(
get_attribute(node, 'value', 0))
else:
# create new dynamic shape
sympy_shape = self._new_symbolic_shape(
self._get_shape_rank(node, 0), node)
vi.CopyFrom(
helper.make_tensor_value_info(
node.output[0], vi.type.tensor_type.elem_type,
get_shape_from_sympy_shape(sympy_shape)))
def _infer_Expand(self, node):
expand_to_shape = self._try_get_value(node, 1)
if expand_to_shape is not None:
# new_shape's dim can come from shape value
self._update_computed_dims(expand_to_shape)
shape = self._get_shape(node, 0)
new_shape = self._broadcast_shapes(
shape, get_shape_from_sympy_shape(expand_to_shape))
vi = self.known_vi_[node.output[0]]
vi.CopyFrom(
helper.make_tensor_value_info(node.output[0], self.known_vi_[
node.input[0]].type.tensor_type.elem_type, new_shape))
def _infer_Gather(self, node):
data_shape = self._get_shape(node, 0)
axis = handle_negative_axis(
get_attribute(node, 'axis', 0), len(data_shape))
indices_shape = self._get_shape(node, 1)
new_shape = data_shape[:axis] + indices_shape + data_shape[axis + 1:]
vi = self.known_vi_[node.output[0]]
vi.CopyFrom(
helper.make_tensor_value_info(node.output[
0], vi.type.tensor_type.elem_type, new_shape))
if node.input[0] in self.sympy_data_:
assert 0 == get_attribute(node, 'axis',
0) # only handle 1D sympy compute
idx = self._get_value(node, 1)
data = self.sympy_data_[node.input[0]]
if type(data) == list:
if type(idx) == np.ndarray and len(idx.shape) == 1:
self.sympy_data_[node.output[0]] = [
data[int(i)] for i in idx
]
else:
self.sympy_data_[node.output[0]] = data[int(idx)]
else:
assert idx == 0
self.sympy_data_[node.output[0]] = data
def _infer_GatherElements(self, node):
indices_shape = self._get_shape(node, 1)
vi = self.known_vi_[node.output[0]]
vi.CopyFrom(
helper.make_tensor_value_info(node.output[0], self.known_vi_[
node.input[0]].type.tensor_type.elem_type, indices_shape))
def _infer_GatherND(self, node):
data_shape = self._get_shape(node, 0)
data_rank = len(data_shape)
indices_shape = self._get_shape(node, 1)
indices_rank = len(indices_shape)
last_index_dimension = indices_shape[-1]
assert is_literal(
last_index_dimension) and last_index_dimension <= data_rank
new_shape = indices_shape[:-1] + data_shape[last_index_dimension:]
vi = self.known_vi_[node.output[0]]
vi.CopyFrom(
helper.make_tensor_value_info(node.output[0], self.known_vi_[
node.input[0]].type.tensor_type.elem_type, new_shape))
def _infer_If(self, node):
# special case for constant condition, in case there are mismatching shape from the non-executed branch
subgraphs = [
get_attribute(node, 'then_branch'),
get_attribute(node, 'else_branch')
]
cond = self._try_get_value(node, 0)
if cond is not None:
if cond > 0:
subgraphs[1].CopyFrom(subgraphs[0])
else:
subgraphs[0].CopyFrom(subgraphs[1])
for i_sub, subgraph in enumerate(subgraphs):
subgraph_infer = self._onnx_infer_subgraph(
node, subgraph, use_node_input=False)
for i_out in range(len(node.output)):
vi = self.known_vi_[node.output[i_out]]
if i_sub == 0:
vi.CopyFrom(subgraph.output[i_out])
vi.name = node.output[i_out]
else:
assert all([
d1 == d2
for d1, d2 in zip(vi.type.tensor_type.shape.dim,
subgraph.output[
i_out].type.tensor_type.shape.dim)
])
# pass on sympy data from subgraph, if cond is constant
if cond is not None and i_sub == (0 if cond > 0 else 1):
if subgraph.output[
i_out].name in subgraph_infer.sympy_data_:
self.sympy_data_[vi.name] = subgraph_infer.sympy_data_[
subgraph.output[i_out].name]
def _infer_Loop(self, node):
subgraph = get_attribute(node, 'body')
assert len(subgraph.input) == len(node.input)
for i, si in enumerate(subgraph.input):
subgraph_name = si.name
si.CopyFrom(self.known_vi_[node.input[i]])
si.name = subgraph_name
self._onnx_infer_subgraph(node, subgraph)
# create a new symbolic dimension for iteration dependent dimension
loop_iter_dim = self._new_symbolic_dim_from_output(node)
num_loop_carried = len(node.input) - 2
for i in range(len(node.output)):
vi = self.known_vi_[node.output[i]]
vi.CopyFrom(
subgraph.output[i + 1]
) # first subgraph output is condition, not in node output
if i >= num_loop_carried:
subgraph_vi_dim = subgraph.output[i +
1].type.tensor_type.shape.dim
vi.type.tensor_type.shape.ClearField('dim')
vi_dim = vi.type.tensor_type.shape.dim
vi_dim.add().dim_param = loop_iter_dim
vi_dim.extend(list(subgraph_vi_dim))
vi.name = node.output[i]
def _infer_MatMul(self, node):
self._compute_matmul_shape(node)
def _infer_MatMulInteger(self, node):
self._compute_matmul_shape(node, onnx.TensorProto.INT32)
def _infer_NonMaxSuppression(self, node):
selected = self._new_symbolic_dim_from_output(node)
vi = self.known_vi_[node.output[0]]
vi.CopyFrom(
helper.make_tensor_value_info(node.output[
0], onnx.TensorProto.INT64, [selected, 3]))
def _infer_NonZero(self, node):
input_rank = self._get_shape_rank(node, 0)
# create a new symbolic dimension for NonZero output
nz_len = self._new_symbolic_dim_from_output(node, 0, 1)
vi = self.known_vi_[node.output[0]]
vi.CopyFrom(
helper.make_tensor_value_info(node.output[
0], vi.type.tensor_type.elem_type, [input_rank, nz_len]))
def _infer_OneHot(self, node):
shape = self._get_shape(node, 0)
axis = get_attribute(node, 'axis', -1)
axis = handle_negative_axis(axis, len(shape) + 1)
new_shape = shape[:axis] + [self._new_symbolic_dim_from_output(node)
] + shape[axis:]
vi = self.known_vi_[node.output[0]]
vi.CopyFrom(
helper.make_tensor_value_info(node.output[0], self.known_vi_[
node.input[2]].type.tensor_type.elem_type, new_shape))
def _infer_Pad(self, node):
if get_opset(self.out_mp_) <= 10:
pads = get_attribute(node, 'pads')
else:
pads = self._try_get_value(node, 1)
vi = self.known_vi_[node.output[0]]
output_shape = get_shape_from_type_proto(vi.type)
if len(output_shape) == 0 or None in output_shape:
sympy_shape = self._get_sympy_shape(node, 0)
rank = len(sympy_shape)
if pads is not None:
assert len(pads) == 2 * rank
new_sympy_shape = [
d + pad_up + pad_down
for d, pad_up, pad_down in zip(sympy_shape, pads[:rank],
pads[rank:])
]
self._update_computed_dims(new_sympy_shape)
else:
# dynamic pads, create new symbolic dimensions
new_sympy_shape = self._new_symbolic_shape(rank, node)
output_tp = self.known_vi_[node.input[0]].type.tensor_type.elem_type
vi.CopyFrom(
helper.make_tensor_value_info(node.output[
0], output_tp, get_shape_from_sympy_shape(new_sympy_shape)))
def _infer_Pool(self, node):
sympy_shape = self._compute_conv_pool_shape(node)
self._update_computed_dims(sympy_shape)
for o in node.output:
if not o:
continue
vi = self.known_vi_[o]
vi.CopyFrom(
helper.make_tensor_value_info(o, vi.type.tensor_type.elem_type,
get_shape_from_sympy_shape(
sympy_shape)))
def _infer_Range(self, node):
vi = self.known_vi_[node.output[0]]
input_data = self._get_int_values(node)
if all([i is not None for i in input_data]):
start = as_scalar(input_data[0])
limit = as_scalar(input_data[1])
delta = as_scalar(input_data[2])
new_sympy_shape = [
sympy.Max(sympy.ceiling((limit - start) / delta), 0)
]
else:
new_dim = self._new_symbolic_dim_from_output(node)
new_sympy_shape = [self.symbolic_dims_[new_dim]]
self._update_computed_dims(new_sympy_shape)
vi.CopyFrom(
helper.make_tensor_value_info(
node.output[0], self.known_vi_[node.input[0]].type.tensor_type.
elem_type, get_shape_from_sympy_shape(new_sympy_shape)))
def _infer_ReduceProd(self, node):
axes = get_attribute(node, 'axes')
keep_dims = get_attribute(node, 'keepdims')
if keep_dims == 0 and axes == [0]:
data = self._get_int_values(node)[0]
if data is not None:
self.sympy_data_[node.output[0]] = sympy_reduce_product(data)
def _infer_Reshape(self, node):
shape_value = self._try_get_value(node, 1)
vi = self.known_vi_[node.output[0]]
if shape_value is None:
shape_shape = self._get_shape(node, 1)
assert len(shape_shape) == 1
shape_rank = shape_shape[0]
assert is_literal(shape_rank)
vi.CopyFrom(
helper.make_tensor_value_info(
node.output[0], vi.type.tensor_type.elem_type,
get_shape_from_sympy_shape(
self._new_symbolic_shape(shape_rank, node))))
else:
input_shape = self._get_shape(node, 0)
input_sympy_shape = self._get_sympy_shape(node, 0)
total = int(1)
for d in input_sympy_shape:
total = total * d
new_sympy_shape = []
deferred_dim_idx = -1
non_deferred_size = int(1)
for i, d in enumerate(shape_value):
if type(d) == sympy.Symbol:
new_sympy_shape.append(d)
elif d == 0:
new_sympy_shape.append(input_sympy_shape[i])
non_deferred_size = non_deferred_size * input_sympy_shape[i]
else:
new_sympy_shape.append(d)
if d == -1:
deferred_dim_idx = i
elif d != 0:
non_deferred_size = non_deferred_size * d
assert new_sympy_shape.count(-1) < 2
if -1 in new_sympy_shape:
new_dim = total // non_deferred_size
new_sympy_shape[deferred_dim_idx] = new_dim
self._update_computed_dims(new_sympy_shape)
vi.CopyFrom(
helper.make_tensor_value_info(
node.output[0], vi.type.tensor_type.elem_type,
get_shape_from_sympy_shape(new_sympy_shape)))
self._pass_on_sympy_data(node)
def _infer_Resize(self, node):
vi = self.known_vi_[node.output[0]]
input_sympy_shape = self._get_sympy_shape(node, 0)
if get_opset(self.out_mp_) <= 10:
scales = self._try_get_value(node, 1)
if scales is not None:
new_sympy_shape = [
sympy.simplify(sympy.floor(d * s))
for d, s in zip(input_sympy_shape, scales)
]
self._update_computed_dims(new_sympy_shape)
vi.CopyFrom(
helper.make_tensor_value_info(
node.output[0], self.known_vi_[node.input[
0]].type.tensor_type.elem_type,
get_shape_from_sympy_shape(new_sympy_shape)))
else:
roi = self._try_get_value(node, 1)
scales = self._try_get_value(node, 2)
sizes = self._try_get_value(node, 3)
if sizes is not None:
new_sympy_shape = [
sympy.simplify(sympy.floor(s)) for s in sizes
]
self._update_computed_dims(new_sympy_shape)
elif roi is not None and scales is not None:
rank = len(scales)
assert len(roi) == 2 * rank
roi_start = list(roi)[:rank]
roi_end = list(roi)[rank:]
scales = list(scales)
new_sympy_shape = [
sympy.simplify(sympy.floor(d * (end - start) * scale))
for d, start, end, scale in zip(input_sympy_shape,
roi_start, roi_end, scales)
]
self._update_computed_dims(new_sympy_shape)
else:
new_sympy_shape = self._new_symbolic_shape(
self._get_shape_rank(node, 0), node)
vi.CopyFrom(
helper.make_tensor_value_info(node.output[0], self.known_vi_[
node.input[0]].type.tensor_type.elem_type,
get_shape_from_sympy_shape(
new_sympy_shape)))
def _infer_Scan(self, node):
subgraph = get_attribute(node, 'body')
num_scan_inputs = get_attribute(node, 'num_scan_inputs')
scan_input_axes = get_attribute(node, 'scan_input_axes',
[0] * num_scan_inputs)
num_scan_states = len(node.input) - num_scan_inputs
scan_input_axes = [
handle_negative_axis(
ax, self._get_shape_rank(node, i + num_scan_states))
for i, ax in enumerate(scan_input_axes)
]
# We may have cases where the subgraph has optionial inputs that appear in both subgraph's input and initializer,
# but not in the node's input. In such cases, the input model might be invalid, but let's skip those optional inputs.
assert len(subgraph.input) >= len(node.input)
subgraph_inputs = subgraph.input[:len(node.input)]
for i, si in enumerate(subgraph_inputs):
subgraph_name = si.name
si.CopyFrom(self.known_vi_[node.input[i]])
if i >= num_scan_states:
scan_input_dim = si.type.tensor_type.shape.dim
scan_input_dim.remove(scan_input_dim[scan_input_axes[
i - num_scan_states]])
si.name = subgraph_name
self._onnx_infer_subgraph(node, subgraph)
num_scan_outputs = len(node.output) - num_scan_states
scan_output_axes = get_attribute(node, 'scan_output_axes',
[0] * num_scan_outputs)
scan_input_dim = get_shape_from_type_proto(self.known_vi_[node.input[
-1]].type)[scan_input_axes[-1]]
for i, o in enumerate(node.output):
vi = self.known_vi_[o]
if i >= num_scan_states:
shape = get_shape_from_type_proto(subgraph.output[i].type)
new_dim = handle_negative_axis(
scan_output_axes[i - num_scan_states], len(shape) + 1)
shape = shape[:new_dim] + [scan_input_dim] + shape[new_dim:]
vi.CopyFrom(
helper.make_tensor_value_info(o, subgraph.output[
i].type.tensor_type.elem_type, shape))
else:
vi.CopyFrom(subgraph.output[i])
vi.name = o
def _infer_ScatterElements(self, node):
data_shape = self._get_shape(node, 0)
vi = self.known_vi_[node.output[0]]
vi.CopyFrom(
helper.make_tensor_value_info(node.output[0], self.known_vi_[
node.input[0]].type.tensor_type.elem_type, data_shape))
def _infer_Shape(self, node):
self.sympy_data_[node.output[0]] = self._get_sympy_shape(node, 0)
def _infer_Size(self, node):
sympy_shape = self._get_sympy_shape(node, 0)
self.sympy_data_[node.output[0]] = sympy_reduce_product(sympy_shape)
self.known_vi_[node.output[0]].CopyFrom(
helper.make_tensor_value_info(node.output[0],
onnx.TensorProto.INT64, []))
def _infer_Slice(self, node):
if get_opset(self.out_mp_) <= 9:
axes = get_attribute(node, 'axes')
starts = get_attribute(node, 'starts')
ends = get_attribute(node, 'ends')
steps = [1] * len(axes)
else:
starts = as_list(self._try_get_value(node, 1), keep_none=True)
ends = as_list(self._try_get_value(node, 2), keep_none=True)
axes = self._try_get_value(node, 3)
steps = self._try_get_value(node, 4)
if axes is None and not (starts is None and ends is None):
axes = list(
range(0, len(starts if starts is not None else ends)))
if steps is None and not (starts is None and ends is None):
steps = [1] * len(starts if starts is not None else ends)
axes = as_list(axes, keep_none=True)
steps = as_list(steps, keep_none=True)
new_sympy_shape = self._get_sympy_shape(node, 0)
if starts is None or ends is None:
if axes is None:
for i in range(len(new_sympy_shape)):
new_sympy_shape[i] = self._new_symbolic_dim_from_output(
node, 0, i)
else:
new_sympy_shape = get_shape_from_sympy_shape(new_sympy_shape)
for i in axes:
new_sympy_shape[i] = self._new_symbolic_dim_from_output(
node, 0, i)
else:
for i, s, e, t in zip(axes, starts, ends, steps):
idx = handle_negative_axis(i, len(new_sympy_shape))
if is_literal(e):
if e >= self.int_max_:
e = new_sympy_shape[i]
elif e <= -self.int_max_:
e = 0 if s > 0 else -1
elif is_literal(new_sympy_shape[i]):
if e < 0:
e = e + new_sympy_shape[i]
e = min(e, new_sympy_shape[i])
else:
if e > 0:
e = sympy.Min(
e, new_sympy_shape[i]
) if e > 1 else e #special case for slicing first to make computation easier
else:
e = new_sympy_shape[i] + e
else:
if is_literal(new_sympy_shape[i]):
e = sympy.Min(e, new_sympy_shape[i])
else:
try:
if e >= new_sympy_shape[i]:
e = new_sympy_shape[i]
except Exception:
print(
'Unable to determine if {} <= {}, treat as equal'
.format(e, new_sympy_shape[i]))
e = new_sympy_shape[i]
if is_literal(s) and int(s) < 0:
s = new_sympy_shape[i] + s
new_sympy_shape[idx] = (e - s + t + (-1 if t > 0 else 1)) // t
self._update_computed_dims(new_sympy_shape)
vi = self.known_vi_[node.output[0]]
vi.CopyFrom(
helper.make_tensor_value_info(
node.output[0], vi.type.tensor_type.elem_type,
get_shape_from_sympy_shape(new_sympy_shape)))
# handle sympy_data if needed, for slice in shape computation
if node.input[0] in self.sympy_data_:
assert [0] == axes
assert len(starts) == 1
assert len(ends) == 1
self.sympy_data_[node.output[0]] = self.sympy_data_[node.input[0]][
starts[0]:ends[0]]
def _infer_Split(self, node):
input_sympy_shape = self._get_sympy_shape(node, 0)
axis = handle_negative_axis(
get_attribute(node, 'axis', 0), len(input_sympy_shape))
split = get_attribute(node, 'split')
if not split:
num_outputs = len(node.output)
split = [input_sympy_shape[axis] /
sympy.Integer(num_outputs)] * num_outputs
self._update_computed_dims(split)
else:
split = [sympy.Integer(s) for s in split]
for i_o in range(len(split)):
vi = self.known_vi_[node.output[i_o]]
vi.CopyFrom(
helper.make_tensor_value_info(
node.output[i_o], self.known_vi_[node.input[
0]].type.tensor_type.elem_type,
get_shape_from_sympy_shape(input_sympy_shape[:axis] + [
split[i_o]
] + input_sympy_shape[axis + 1:])))
self.known_vi_[vi.name] = vi
def _infer_Squeeze(self, node):
self._pass_on_sympy_data(node)
def _infer_Tile(self, node):
repeats_value = self._get_value(node, 1)
input_sympy_shape = self._get_sympy_shape(node, 0)
new_sympy_shape = []
for i, d in enumerate(input_sympy_shape):
new_dim = d * repeats_value[i]
new_sympy_shape.append(new_dim)
self._update_computed_dims(new_sympy_shape)
vi = self.known_vi_[node.output[0]]
vi.CopyFrom(
helper.make_tensor_value_info(
node.output[0], vi.type.tensor_type.elem_type,
get_shape_from_sympy_shape(new_sympy_shape)))
def _infer_TopK(self, node):
rank = self._get_shape_rank(node, 0)
axis = handle_negative_axis(get_attribute(node, 'axis', -1), rank)
new_shape = self._get_shape(node, 0)
if get_opset(self.out_mp_) <= 9:
k = get_attribute(node, 'k')
else:
k = self._get_int_values(node)[1]
if k == None:
k = self._new_symbolic_dim_from_output(node)
else:
k = as_scalar(k)
if type(k) in [int, str]:
new_shape[axis] = k
else:
new_sympy_shape = self._get_sympy_shape(node, 0)
new_sympy_shape[axis] = k
self._update_computed_dims(
new_sympy_shape
) # note that TopK dim could be computed in sympy_data, so need to update computed_dims when it enters shape
new_shape = get_shape_from_sympy_shape(new_sympy_shape)
for i_o in range(len(node.output)):
vi = self.known_vi_[node.output[i_o]]
vi.CopyFrom(
helper.make_tensor_value_info(node.output[
i_o], vi.type.tensor_type.elem_type, new_shape))
def _infer_Unsqueeze(self, node):
self._pass_on_sympy_data(node)
def _infer_ZipMap(self, node):
map_key_type = None
if get_attribute(node, 'classlabels_int64s') is not None:
map_key_type = onnx.TensorProto.INT64
elif get_attribute(node, 'classlabels_strings') is not None:
map_key_type = onnx.TensorProto.STRING
assert map_key_type is not None
new_vi = onnx.ValueInfoProto()
new_vi.name = node.output[0]
new_vi.type.sequence_type.elem_type.map_type.value_type.tensor_type.elem_type = onnx.TensorProto.FLOAT
new_vi.type.sequence_type.elem_type.map_type.key_type = map_key_type
vi = self.known_vi_[node.output[0]]
vi.CopyFrom(new_vi)
def _infer_impl(self, in_mp, start_sympy_data={}):
self.sympy_data_ = start_sympy_data
self.out_mp_.graph.ClearField('value_info')
self._apply_suggested_merge(graph_input_only=True)
self.input_symbols_ = set()
for i in self.out_mp_.graph.input:
input_dims = i.type.tensor_type.shape.dim
for i_dim in range(len(input_dims)):
if get_dim_from_type_proto(input_dims[i_dim]) is None:
# some models use None for symbolic dim in input, replace it with a string
input_dims[i_dim].dim_param = self._new_symbolic_dim(i.name,
i_dim)
self.input_symbols_.update([
d for d in get_shape_from_type_proto(i.type) if type(d) == str
])
for s in self.input_symbols_:
if s in self.suggested_merge_:
s_merge = self.suggested_merge_[s]
assert s_merge in self.symbolic_dims_
self.symbolic_dims_[s] = self.symbolic_dims_[s_merge]
else:
self.symbolic_dims_[s] = sympy.Symbol(s, integer=True)
# create a temporary ModelProto for single node inference
# note that we remove initializer to have faster inference
# for tensor ops like Reshape/Tile/Expand that read initializer, we need to do sympy computation based inference anyways
self.tmp_mp_ = onnx.ModelProto()
self.tmp_mp_.CopyFrom(self.out_mp_)
self.tmp_mp_.graph.ClearField('initializer')
for node in self.out_mp_.graph.node:
assert all([i in self.known_vi_ for i in node.input if i])
self._onnx_infer_single_node(node)
if node.op_type in self.dispatcher_:
self.dispatcher_[node.op_type](node)
if self.verbose_ > 2:
print(node.op_type + ': ' + node.name)
for i, name in enumerate(node.input):
print(' Input {}: {} {}€5€5€5€5€5'.format(
i, name, 'initializer'
if name in self.initializers_ else ''))
# onnx automatically merge dims with value, i.e. Mul(['aaa', 'bbb'], [1000, 1]) -> [1000, 'bbb']
# symbolic shape inference needs to apply merge of 'aaa' -> 1000 in this case
if node.op_type in [
'Add', 'Sub', 'Mul', 'Div', 'MatMul', 'MatMulInteger',
'MatMulInteger16', 'Where', 'Sum'
]:
vi = self.known_vi_[node.output[0]]
out_rank = len(get_shape_from_type_proto(vi.type))
in_shapes = [
self._get_shape(node, i) for i in range(len(node.input))
]
for d in range(out_rank - (2 if node.op_type in [
'MatMul', 'MatMulInteger', 'MatMulInteger16'
] else 0)):
in_dims = [
s[len(s) - out_rank + d] for s in in_shapes
if len(s) + d >= out_rank
]
if len(in_dims) > 1:
self._check_merged_dims(in_dims, allow_broadcast=True)
for i_o in range(len(node.output)):
vi = self.known_vi_[node.output[i_o]]
out_type = vi.type
out_type_kind = out_type.WhichOneof('value')
# only TensorProto and SparseTensorProto have shape
if out_type_kind != 'tensor_type' and out_type_kind != 'sparse_tensor_type':
continue
out_shape = get_shape_from_type_proto(vi.type)
out_type_undefined = out_type.tensor_type.elem_type == onnx.TensorProto.UNDEFINED
if self.verbose_ > 2:
print(' {}: {} {}'.format(node.output[
i_o], str(out_shape), vi.type.tensor_type.elem_type))
if node.output[i_o] in self.sympy_data_:
print(' Sympy Data: ' + str(self.sympy_data_[
node.output[i_o]]))
if None in out_shape or out_type_undefined:
if self.auto_merge_:
if node.op_type in [
'Add', 'Sub', 'Mul', 'Div', 'MatMul',
'MatMulInteger', 'MatMulInteger16', 'Concat',
'Where', 'Sum'
]:
shapes = [
self._get_shape(node, i)
for i in range(len(node.input))
]
if node.op_type in [
'MatMul', 'MatMulInteger', 'MatMulInteger16'
]:
# only support auto merge for MatMul for dim < rank-2 when rank > 2
assert len(shapes[0]) > 2 and dim_idx[0] < len(
shapes[0]) - 2
assert len(shapes[1]) > 2 and dim_idx[1] < len(
shapes[1]) - 2
elif node.op_type == 'Expand':
# auto merge for cases like Expand([min(batch, 1), min(seq, 512)], [batch, seq])
shapes = [
self._get_shape(node, 0),
self._get_value(node, 1)
]
else:
shapes = []
if shapes:
for idx in range(len(out_shape)):
if out_shape[idx] is not None:
continue
dim_idx = [
len(s) - len(out_shape) + idx
for s in shapes
]
assert all([d >= 0 for d in dim_idx])
self._add_suggested_merge([
s[i] if is_literal(s[i]) else str(s[i])
for s, i in zip(shapes, dim_idx)
])
self.run_ = True
else:
self.run_ = False
else:
self.run_ = False
# create new dynamic dims for ops not handled by symbolic shape inference
if self.run_ == False and not node.op_type in self.dispatcher_:
is_unknown_op = (out_type_undefined and
len(out_shape) == 0)
if is_unknown_op:
# unknown op to ONNX, maybe from higher opset or other domain
# only guess the output rank from input 0 when using guess_output_rank option
out_rank = self._get_shape_rank(
node, 0) if self.guess_output_rank_ else -1
else:
# valid ONNX op, but not handled by symbolic shape inference, just assign dynamic shape
out_rank = len(out_shape)
if out_rank >= 0:
new_shape = self._new_symbolic_shape(out_rank, node,
i_o)
vi.CopyFrom(
helper.make_tensor_value_info(
vi.name, self.known_vi_[node.input[
0]].type.tensor_type.elem_type,
get_shape_from_sympy_shape(new_shape)))
if self.verbose_ > 0:
if is_unknown_op:
print(
"Possible unknown op: {} node: {}, guessing {} shape"
.format(node.op_type, node.name,
vi.name))
if self.verbose_ > 2:
print(' {}: {} {}'.format(
node.output[i_o],
str(new_shape),
vi.type.tensor_type.elem_type))
self.run_ = True
continue # continue the inference after guess, no need to stop as no merge is needed
if self.verbose_ > 0 or not self.auto_merge_ or out_type_undefined:
print('Stopping at incomplete shape inference at ' +
node.op_type + ': ' + node.name)
print('node inputs:')
for i in node.input:
print(self.known_vi_[i])
print('node outputs:')
for o in node.output:
print(self.known_vi_[o])
if self.auto_merge_ and not out_type_undefined:
print('Merging: ' + str(self.suggested_merge_))
return False
self.run_ = False
return True
def _update_output_from_vi(self):
for output in self.out_mp_.graph.output:
if output.name in self.known_vi_:
tmp_output = self.known_vi_[output.name]
output.CopyFrom(tmp_output)
@staticmethod
def infer_shapes(in_mp,
int_max=2**31 - 1,
fixed_input_shape=None,
auto_merge=True,
guess_output_rank=False,
verbose=0):
if get_opset(in_mp) < 7:
print('Only support shape inferencing models of opset 7 and above.')
return
symbolic_shape_inference = SymbolicShapeInference(
int_max, auto_merge, guess_output_rank, verbose)
all_shapes_inferred = False
symbolic_shape_inference._preprocess(
in_mp, input_shapes=fixed_input_shape)
try:
while symbolic_shape_inference.run_:
all_shapes_inferred = symbolic_shape_inference._infer_impl(
in_mp)
symbolic_shape_inference._update_output_from_vi()
if not all_shapes_inferred:
print('!' * 10)
symbolic_shape_inference.out_mp_ = shape_inference.infer_shapes(
symbolic_shape_inference.out_mp_)
#onnx.save(symbolic_shape_inference.out_mp_, 'tmp.onnx')
except:
print('Stopping at incomplete shape inference')
symbolic_shape_inference.out_mp_ = shape_inference.infer_shapes(
symbolic_shape_inference.out_mp_)
return symbolic_shape_inference.out_mp_.graph
x2paddle/x2paddle/decoder/paddle_decoder.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import paddle.fluid as fluid
class PaddleDecoder(object):
def __init__(self,
model_dir,
model_filename='__model__',
params_filename=None):
exe = fluid.Executor(fluid.CPUPlace())
[self.program, feed, fetchs] = fluid.io.load_inference_model(
model_dir,
exe,
model_filename=model_filename,
params_filename=params_filename)
x2paddle/x2paddle/decoder/tf_decoder.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from x2paddle.core.graph import GraphNode, Graph
from x2paddle.core.fluid_code import FluidCode
from tensorflow.python.framework import tensor_util
from tensorflow.core.framework import attr_value_pb2
import tensorflow as tf
import copy as cp
import numpy
import sys
class TFGraphNode(GraphNode):
def __init__(self, layer, layer_name=None, data_format="NHWC"):
if layer_name is None:
super(TFGraphNode, self).__init__(
layer,
layer.name.replace('/', '_').replace('-', '_').replace('^', ''))
else:
super(TFGraphNode, self).__init__(
layer,
layer_name.replace('/', '_').replace('-', '_').replace('^', ''))
self.layer_type = layer.op
self.tf_data_format = data_format
self.pd_data_format = "NCHW"
self.fluid_code = FluidCode()
self.dtype_map = {
1: "float32",
3: "int32",
4: "uint8",
9: "int64",
10: "bool"
}
@property
def out_shapes(self):
if self.layer_type == "OneShotIterator" or self.layer_type == "IteratorV2":
values = self.layer.attr["output_shapes"].list.shape
else:
values = self.layer.attr["_output_shapes"].list.shape
out_shapes = list()
for value in values:
shape = [dim.size for dim in value.dim]
out_shapes.append(shape)
return out_shapes
@property
def dtype(self):
keys = ['dtype', 'Tidx', 'T', 'DstT']
for k in keys:
dtype = self.layer.attr[k].type
if dtype > 0:
break
if dtype == 0:
dtype = self.layer.attr['output_types'].list.type[0]
if dtype not in self.dtype_map:
raise Exception("Dtype[{}] of node({}) not in dtype_map".format(
dtype, self.layer.name))
return self.dtype_map[dtype]
@property
def raw_dtype(self):
keys = ['dtype', 'Tidx', 'T', 'DstT']
for k in keys:
dtype = self.layer.attr[k].type
if dtype > 0:
break
return dtype
@property
def value(self):
assert self.layer_type == "Const", "Only Const node has value."
attr = self.layer.attr['value']
field = getattr(attr, attr.WhichOneof('value'))
return tensor_util.MakeNdarray(field)
def get_attr(self, name):
if name not in self.layer.attr:
return None
attr = self.layer.attr[name]
field = attr.WhichOneof('value')
value = getattr(attr, field) if field else None
if isinstance(value, attr_value_pb2.AttrValue.ListValue):
result = list(value.ListFields()[0][1])
for i in range(len(result)):
if isinstance(result[i], int):
result[i] = int(result[i])
try:
if isinstance(result[i], long):
result[i] = int(result[i])
except:
pass
return result
else:
return value
class TFGraph(Graph):
def __init__(self, model, data_format="NHWC"):
super(TFGraph, self).__init__(model)
self.identity_map = dict()
self.multi_out_ops = ['Split', 'SplitV', 'IteratorV2']
self.tf_data_format = data_format
def build(self):
for layer in self.model.node:
if layer.op == 'Assert':
continue
self.node_map[layer.name.replace('/', '_').replace(
'-', '_')] = TFGraphNode(
layer, data_format=self.tf_data_format)
for layer_name, node in self.node_map.items():
if node.layer_type == 'Const':
continue
for in_node in node.layer.input:
in_node = in_node.replace('/', '_').replace('-', '_').replace(
'^', '')
if in_node not in self.node_map:
if in_node.strip().split(':')[0] in self.node_map:
self.connect(in_node.strip().split(':')[0], layer_name)
else:
raise Exception(
'input[{}] of node[{}] does not exist in node_map'.
format(in_node, layer_name))
else:
self.connect(in_node, layer_name)
super(TFGraph, self).build()
for layer in self.model.node:
if layer.op == 'Assert':
for ipt in layer.input:
ipt_name = ipt.replace('-', '_').replace('/', '_')
if ipt_name in self.output_nodes:
idx = self.output_nodes.index(ipt_name)
del self.output_nodes[idx]
# tensorflow graph optimize
self._remove_isolated_node()
self._optimize_dialiation_conv()
self._remove_identity_node()
self._remove_cast_node()
def get_node(self, node_name, copy=False):
items = node_name.strip().split(':')
items[0] = items[0].replace('/', '_').replace('-', '_')
if items[0] in self.identity_map:
items[0] = self.identity_map[items[0]]
new_node_name = ":".join(items)
node = super(TFGraph, self).get_node(new_node_name, copy)
if node is None:
return None
if node.layer_type == "Switch":
if hasattr(node, 'index'):
del node.index
if len(items) == 1 and node.layer_type in self.multi_out_ops:
node.index = 0
return node
def remove_node(self, node_name):
if node_name not in self.node_map:
raise Exception("Node[{}] not in graph".format(node_name))
inputs = self.node_map[node_name].inputs
outputs = self.node_map[node_name].outputs
# assert len(inputs) == 1
input_node = self.node_map[inputs[0]]
idx = input_node.outputs.index(node_name)
del input_node.outputs[idx]
for output in outputs:
node = self.node_map[output]
idx = node.inputs.index(node_name)
node.inputs[idx] = inputs[0]
input_node.outputs.append(output)
del self.node_map[node_name]
idx = self.topo_sort.index(node_name)
del self.topo_sort[idx]
def _optimize_dialiation_conv(self):
for name in list(self.node_map.keys()):
node = self.node_map[name]
if node.layer_type == "SpaceToBatchND":
is_dilation = True
out_node0 = self.node_map[node.outputs[0]]
if out_node0.layer_type != 'ExpandDims':
is_dilation = False
continue
out_node1 = self.node_map[out_node0.outputs[0]]
if out_node1.layer_type != 'Conv2D':
is_dilation = False
continue
out_node2 = self.node_map[out_node1.outputs[0]]
if out_node2.layer_type != 'Squeeze':
is_dilation = False
continue
out_node3 = self.node_map[out_node2.outputs[0]]
if out_node3.layer_type != 'BatchToSpaceND':
is_dilation = False
continue
if is_dilation:
node.skip = True
out_node3.skip = True
block_shape = self.node_map[node.inputs[1]]
out_node1.dilation = block_shape.value.tolist()
def _remove_isolated_node(self):
# delete isolated nodes
isolated_nodes = list()
for node_name in self.node_map.keys():
if len(self.get_node(node_name).inputs) == 0 and len(
self.get_node(node_name).outputs) == 0:
isolated_nodes.append(node_name)
for node_name in isolated_nodes:
del self.node_map[node_name]
if node_name in self.input_nodes:
idx = self.input_nodes.index(node_name)
del self.input_nodes[idx]
if node_name in self.output_nodes:
idx = self.output_nodes.index(node_name)
del self.output_nodes[idx]
idx = self.topo_sort.index(node_name)
del self.topo_sort[idx]
def _remove_identity_node(self):
identity_ops = [
'Identity', 'StopGradient', 'Switch', 'Merge',
'PlaceholderWithDefault', 'IteratorGetNext'
]
identity_node = list()
for node_name, node in self.node_map.items():
if node.layer_type in identity_ops:
identity_node.append(node_name)
for node_name in identity_node:
node = self.get_node(node_name)
input_node = self.get_node(node.inputs[0])
self.remove_node(node_name)
self.identity_map[node_name] = input_node.layer_name
if node_name in self.output_nodes:
idx = self.output_nodes.index(node_name)
self.output_nodes[idx] = input_node.layer_name
def _remove_cast_node(self):
cast_node = list()
for node_name, node in self.node_map.items():
if node.layer_type == "Cast":
input = self.get_node(node.inputs[0])
if input.layer_type != "Placeholder" or len(input.outputs) != 1:
continue
cast_node.append(node_name)
for node_name in cast_node:
node = self.get_node(node_name)
input_node = self.get_node(node.inputs[0])
input_node.layer.attr["dtype"].type = node.raw_dtype
self.remove_node(node_name)
self.identity_map[node_name] = input_node.layer_name
if node_name in self.output_nodes:
idx = self.output_nodes.index(node_name)
self.output_nodes[idx] = input_node.layer_name
def data_format_propagation(self, node):
current_node = self.node_map[node.layer_name]
outputs = current_node.outputs
if len(outputs) == 0:
return
for out in outputs:
next_node = self.node_map[out]
next_node.tf_data_format = node.tf_data_format
self.data_format_propagation(next_node)
class TFDecoder(object):
def __init__(self, pb_model, data_format="NHWC", define_input_shape=False):
try:
self.sess = tf.compat.v1.Session()
except:
self.sess = tf.Session()
self.input_info = dict()
self.define_input_shape = define_input_shape
with open(pb_model, 'rb') as f:
try:
graph_def = tf.compat.v1.GraphDef()
except:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
input_map = self._check_input_shape(graph_def)
self._fix_output_shape(graph_def)
self.sess.graph.as_default()
tf.import_graph_def(graph_def, name='', input_map=input_map)
try:
initializer = tf.compat.v1.global_variables_initializer()
except:
initializer = tf.global_variables_initializer()
self.sess.run(initializer)
self.tf_graph = TFGraph(
self.sess.graph._as_graph_def(add_shapes=True)[0], data_format)
self.tf_graph.build()
def _fix_output_shape(self, graph):
for i in range(len(graph.node)):
node = graph.node[i]
if node.op == "swish_f32":
graph.node[i].attr['_disable_call_shape_inference'].b = False
def _check_input_shape(self, graph_def):
numpy.random.seed(13)
graph_def = cp.deepcopy(graph_def)
input_map = dict()
for layer in graph_def.node:
if layer.op != "Placeholder" and layer.op != "OneShotIterator" and layer.op != "IteratorV2":
continue
graph_node = TFGraphNode(layer)
dtype = graph_node.layer.attr['dtype'].type
need_define_shape = 0
if self.define_input_shape:
need_define_shape = 3
elif graph_node.layer.attr[
'shape'].shape.unknown_rank or not graph_node.get_attr(
"shape"):
need_define_shape = 1
else:
value = graph_node.layer.attr["shape"].shape
shape = [dim.size for dim in value.dim]
if shape.count(-1) > 1:
need_define_shape = 2
if need_define_shape == 1:
try:
shape = graph_node.out_shapes[0]
if len(shape) > 0 and shape.count(-1) < 2:
need_define_shape = 0
except:
pass
if need_define_shape > 0:
shape = None
if graph_node.get_attr("shape"):
value = value = graph_node.layer.attr["shape"].shape
shape = [dim.size for dim in value.dim]
if need_define_shape == 1:
print("Unknown shape for input tensor[tensor name: \"{}\"]".
format(layer.name))
elif need_define_shape == 2:
print(
"\nShape[now is {}] for input tensor[tensor name: \"{}\"] not support yet"
.format(shape, layer.name))
else:
print(
"Define shape[now is {}] for input tensor[tensor name: \"{}\']"
.format(shape, layer.name))
print(
"Use your keyboard type the shape of input tensor below :)")
right_shape_been_input = False
while not right_shape_been_input:
try:
shape = raw_input(
"Shape of Input(e.g. None,224,224,3): ")
except:
shape = input("Shape of Input(e.g. None,224,224,3): ")
if shape.count("None") > 1:
print("Only 1 dimension can be None, type again:)")
else:
right_shape_been_input = True
shape = [
None if dim == "None" else int(dim)
for dim in shape.strip().split(',')
]
assert shape.count(None) <= 1, "Only one dimension can be None"
try:
x2paddle_input = tf.compat.v1.placeholder(
dtype=dtype,
shape=shape,
name="x2paddle_{}".format(layer.name))
except:
x2paddle_input = tf.placeholder(
dtype=dtype,
shape=shape,
name="x2paddle_{}".format(layer.name))
input_map["{}:0".format(layer.name)] = x2paddle_input
if shape.count(None) > 0:
shape[shape.index(None)] = -1
self.input_info["x2paddle_{}".format(layer.name)] = (shape,
dtype)
else:
value = graph_node.layer.attr["shape"].shape
shape = [dim.size for dim in value.dim]
self.input_info[layer.name] = (shape, dtype)
return input_map
# trick method
# should be removed after PaddlePaddle V1.6 been released
def infer_tensor(self, graph_node):
if hasattr(graph_node, "index"):
tensor_name = graph_node.layer.name + ":{}".format(graph_node.index)
else:
tensor_name = graph_node.layer.name + ":0"
feed = dict()
for input_name, info in self.input_info.items():
(shape, dtype) = cp.deepcopy(info)
input_tensor = self.sess.graph.get_tensor_by_name(input_name + ":0")
if shape.count(-1) > 0:
shape[shape.index(-1)] = 2
feed[input_tensor] = numpy.random.random_sample(shape)
output_tensor = self.sess.graph.get_tensor_by_name(tensor_name)
return self.sess.run([output_tensor], feed)[0]
def infer_shape_tensor(self, graph_node, out_shape=None):
if hasattr(graph_node, "index"):
tensor_name = graph_node.layer.name + ":{}".format(graph_node.index)
else:
tensor_name = graph_node.layer.name + ":0"
feed = dict()
batch_size = [2, 3, 5]
results = list()
for b in batch_size:
for input_name, info in self.input_info.items():
(shape, dtype) = cp.deepcopy(info)
input_tensor = self.sess.graph.get_tensor_by_name(input_name +
":0")
if shape.count(-1) > 0:
shape[shape.index(-1)] = b
feed[input_tensor] = numpy.random.random_sample(shape)
output_tensor = self.sess.graph.get_tensor_by_name(tensor_name)
results.append(self.sess.run([output_tensor], feed)[0].flatten())
compare01 = (results[0] == results[1])
compare12 = (results[1] == results[2])
if compare01.all() and compare12.all():
return results[0].tolist()
if (compare01 == compare12).all():
index = numpy.argwhere(compare01 == False).flatten()
if index.shape[0] != 1:
raise Exception("There's not only one unstable dimension")
results[0][index[0]] = -1
index = numpy.argwhere(results[0] < 0).flatten()
if index.shape[0] > 2:
print("Warning: More than two dimension less than zero")
if index.shape[0] == 2 and out_shape is not None:
if out_shape[index[1]] > 0:
results[0][index[1]] = out_shape[index[1]]
else:
results[0][index[0]] = out_shape[index[0]]
return results[0].tolist()
else:
raise Exception("Couldn't infer a stable shape shape tensor value")
def infer_tensor_shape(self, graph_node):
if hasattr(graph_node, "index"):
tensor_name = graph_node.layer.name + ":{}".format(graph_node.index)
else:
tensor_name = graph_node.layer.name + ":0"
feed = dict()
batch_size = [2, 3, 5]
shapes = list()
for b in batch_size:
for input_name, info in self.input_info.items():
(shape, dtype) = cp.deepcopy(info)
input_tensor = self.sess.graph.get_tensor_by_name(input_name +
":0")
if shape.count(-1) > 0:
shape[shape.index(-1)] = b
feed[input_tensor] = numpy.random.random_sample(shape)
output_tensor = self.sess.graph.get_tensor_by_name(tensor_name)
shape = self.sess.run([output_tensor], feed)[0].shape
shapes.append(numpy.array(shape))
compare01 = (shapes[0] == shapes[1])
compare12 = (shapes[1] == shapes[2])
if compare01.all() and compare12.all():
return shape[0].tolist()
if (compare01 == compare12).all():
index = numpy.argwhere(compare01 == False).flatten()
if index.shape[0] != 1:
raise Exception("There's not only one unstable dimension")
if index[0] != 0:
raise Exception("Batch size not in the first dimension")
shapes[0][0] = -1
return shapes[0].tolist()
x2paddle/x2paddle/op_mapper/caffe_custom_layer/__init__.py 0 → 100644
浏览文件 @ cf7f9b88
from .register import get_registered_layers
#custom layer import begins
from . import roipooling
from . import priorbox
from . import permute
from . import detectionoutput
from . import normalize
from . import select
from . import shufflechannel
from . import convolutiondepthwise
from . import axpy
from . import upsample
from . import relu6
#custom layer import ends
custom_layers = get_registered_layers()
def set_args(f, params):
""" set args for function 'f' using the parameters in node.layer.param
Args:
f (function): a python function object
params (object): a object contains attributes needed by f's arguments
Returns:
arg_names (list): a list of argument names
kwargs (dict): a dict contains needed arguments
"""
argc = f.__code__.co_argcount
arg_list = f.__code__.co_varnames[0:argc]
kwargs = {}
for arg_name in arg_list:
if hasattr(params, arg_name) and params is not None:
kwargs[arg_name] = getattr(params, arg_name)
return arg_list, kwargs
def has_layer(layer_type):
""" test whether this layer exists in custom layer
"""
return layer_type in custom_layers
def get_params(layer, layer_type):
import re
if layer_type.lower() == "deconvolution" or layer_type.lower(
) == "convolutiondepthwise":
param_name = '_'.join(('convolution', 'param'))
elif layer_type.lower() == "normalize":
param_name = '_'.join(('norm', 'param'))
elif len(layer_type) - len(re.sub("[A-Z]", "", layer_type)) >= 2:
s = ''
tmp_name = ''
for i, ch in enumerate(layer_type):
if i == 0:
s += ch.lower()
continue
elif ch.isupper() and layer_type[i - 1].islower():
tmp_name += (s + '_')
s = ''
s += ch.lower()
tmp_name += s
param_name = '_'.join((tmp_name, 'param'))
else:
param_name = '_'.join((layer_type.lower(), 'param'))
return getattr(layer, param_name, None)
def compute_output_shape(node):
""" compute the output shape of custom layer
"""
layer_type = node.layer_type
assert layer_type in custom_layers, "layer[%s] not exist in custom layers" % (
layer_type)
shape_func = custom_layers[layer_type]['shape']
layer = node.layer
params = get_params(layer, layer_type)
arg_names, kwargs = set_args(shape_func, params)
input_shape = node.input_shape
return shape_func(input_shape, **kwargs)
def make_custom_layer(node):
""" get the code which implement the custom layer function
"""
layer_type = node.layer_type
assert layer_type in custom_layers, "layer[%s] not exist in custom layers" % (
layer_type)
layer_func = custom_layers[layer_type]['layer']
import inspect
return inspect.getsource(layer_func), layer_func
def deal_weights(node, data=None):
""" deal the weights of the custom layer
"""
layer_type = node.layer_type
weights_func = custom_layers[layer_type]['weights']
name = node.layer_name
return weights_func(name, data)
x2paddle/x2paddle/op_mapper/caffe_custom_layer/axpy.py 0 → 100644
浏览文件 @ cf7f9b88
from .register import register
from x2paddle.core.util import *
def axpy_shape(input_shapes):
assert len(input_shapes) == 3, "not valid input shape for axpy layer"
assert len(input_shapes[0]) == len(input_shapes[1]), 'should have same dims'
output_shape = input_shapes[1]
assert (input_shapes[2] == output_shape),\
"shape not consistent for axpy[%s <--> %s]" \
% (str(output_shape), str(input_shapes[2]))
return [output_shape]
def axpy_layer(inputs, input_shape=None, name=None):
alpha = inputs[0]
x = inputs[1]
y = inputs[2]
out = fluid.layers.elementwise_mul(x, alpha, axis=0)
out = fluid.layers.elementwise_add(out, y, name=name)
return out
def axpy_weights(name, data=None):
weights_name = []
return weights_name
register(kind='Axpy', shape=axpy_shape, layer=axpy_layer, weights=axpy_weights)
x2paddle/x2paddle/op_mapper/caffe_custom_layer/convolutiondepthwise.py 0 → 100644
浏览文件 @ cf7f9b88
from .register import register
from x2paddle.core.util import *
import numbers
def convolutiondepthwise_shape(input_shape,
num_output=None,
pad=None,
kernel_size=None,
stride=None,
dilation=None,
pad_h=None,
pad_w=None,
kernel_h=None,
kernel_w=None,
stride_h=None,
stride_w=None):
[k_h, k_w] = [1, 1]
if isinstance(kernel_size, numbers.Number):
[k_h, k_w] = [kernel_size] * 2
elif len(kernel_size) > 0:
k_h = kernel_h if kernel_h > 0 else kernel_size[0]
k_w = kernel_w if kernel_w > 0 else kernel_size[len(kernel_size) - 1]
elif kernel_h > 0 or kernel_w > 0:
k_h = kernel_h
k_w = kernel_w
[s_h, s_w] = [1, 1]
if isinstance(stride, numbers.Number):
[s_h, s_w] = [stride] * 2
elif len(stride) > 0:
s_h = stride_h if stride_h > 0 else stride[0]
s_w = stride_w if stride_w > 0 else stride[len(stride) - 1]
elif stride_h > 0 or stride_w > 0:
s_h = stride_h
s_w = stride_w
[p_h, p_w] = [0, 0]
if isinstance(pad, numbers.Number):
[p_h, p_w] = [pad] * 2
elif len(pad) > 0:
p_h = pad_h if pad_h > 0 else pad[0]
p_w = pad_w if pad_w > 0 else pad[len(pad) - 1]
elif pad_h > 0 or pad_w > 0:
p_h = pad_h
p_w = pad_w
dila_len = len(dilation)
dila_h = 1
dila_w = 1
if dila_len == 2:
dila_h = dilation[0]
dila_w = dilation[1]
elif dila_len == 1:
dila_h = dila_w = dilation[0]
else:
assert dila_len == 0, "invalid length[%s] of dilation in convolution" % (
dila_len)
i_w = input_shape[0][2]
i_h = input_shape[0][3]
o_h = (i_h + 2 * p_h - (dila_h * (k_h - 1) + 1)) / float(s_h) + 1
o_w = (i_w + 2 * p_w - (dila_w * (k_w - 1) + 1)) / float(s_w) + 1
import math
o_h = int(math.floor(o_h))
o_w = int(math.floor(o_w))
c = num_output if num_output is not None else input_shape[0][1]
return [[input_shape[0][0], c, o_h, o_w]]
def convolutiondepthwise_layer(inputs,
num_output=None,
pad=None,
kernel_size=None,
stride=None,
dilation=None,
pad_h=None,
pad_w=None,
kernel_h=None,
kernel_w=None,
stride_h=None,
stride_w=None,
input_shape=None,
name=None):
import numbers
[k_h, k_w] = [1, 1]
if isinstance(kernel_size, numbers.Number):
[k_h, k_w] = [kernel_size] * 2
elif len(kernel_size) > 0:
k_h = kernel_h if kernel_h > 0 else kernel_size[0]
k_w = kernel_w if kernel_w > 0 else kernel_size[len(kernel_size) - 1]
elif kernel_h > 0 or kernel_w > 0:
k_h = kernel_h
k_w = kernel_w
[s_h, s_w] = [1, 1]
if isinstance(stride, numbers.Number):
[s_h, s_w] = [stride] * 2
elif len(stride) > 0:
s_h = stride_h if stride_h > 0 else stride[0]
s_w = stride_w if stride_w > 0 else stride[len(stride) - 1]
elif stride_h > 0 or stride_w > 0:
s_h = stride_h
s_w = stride_w
[p_h, p_w] = [0, 0]
if isinstance(pad, numbers.Number):
[p_h, p_w] = [pad] * 2
elif len(pad) > 0:
p_h = pad_h if pad_h > 0 else pad[0]
p_w = pad_w if pad_w > 0 else pad[len(pad) - 1]
elif pad_h > 0 or pad_w > 0:
p_h = pad_h
p_w = pad_w
input = inputs[0]
dila_len = len(dilation)
dila_h = 1
dila_w = 1
if dila_len == 2:
dila_h = dilation[0]
dila_w = dilation[1]
elif dila_len == 1:
dila_h = dila_w = dilation[0]
else:
assert dila_len == 0, "invalid length[%s] of dilation in convolution" % (
dila_len)
c_in = input_shape[0][1]
c_out = num_output if num_output is not None else input_shape[0][1]
group = int(c_in / (c_in / c_out)) if c_in > c_out else int(c_in /
(c_out / c_in))
out = fluid.layers.conv2d(
input,
dilation=[dila_h, dila_w],
filter_size=[k_h, k_w],
stride=[s_h, s_w],
padding=[p_h, p_w],
groups=group,
num_filters=c_out,
param_attr=name + '_weights',
bias_attr=name + '_bias',
name=name)
return out
def convolutiondepthwise_weights(name, data=None):
weights_name = []
weights_name.append(name + '_weights')
weights_name.append(name + '_bias')
return weights_name
register(
kind='ConvolutionDepthwise',
shape=convolutiondepthwise_shape,
layer=convolutiondepthwise_layer,
weights=convolutiondepthwise_weights)
x2paddle/x2paddle/op_mapper/caffe_custom_layer/detectionoutput.py 0 → 100644
浏览文件 @ cf7f9b88
from .register import register
from x2paddle.core.util import *
def detectionoutput_shape(input_shape):
return [[-1, 6]]
def detectionoutput_layer(inputs,
nms_param=None,
background_label_id=0,
share_location=True,
keep_top_k=100,
confidence_threshold=0.1,
input_shape=None,
name=None):
nms_param_str = nms_param
nms_param = {}
part = nms_param_str.split(',')
for s in part:
if s == '':
break
else:
name, obj = s.split(': ')
if name == 'top_k':
nms_param[name] = int(obj)
else:
nms_param[name] = float(obj)
if nms_param is None:
nms_param = {"nms_threshold": 0.3, "top_k": 10, "eta": 1.0}
mbox_conf_flatten = inputs[1]
mbox_priorbox = inputs[2]
mbox_priorbox_list = fluid.layers.split(mbox_priorbox, 2, dim=1)
pb = mbox_priorbox_list[0]
pbv = mbox_priorbox_list[1]
pb = fluid.layers.reshape(x=pb, shape=[-1, 4])
pbv = fluid.layers.reshape(x=pbv, shape=[-1, 4])
mbox_loc = inputs[0]
mbox_loc = fluid.layers.reshape(x=mbox_loc, shape=[-1, pb.shape[0], 4])
mbox_conf_flatten = fluid.layers.reshape(
x=mbox_conf_flatten, shape=[0, pb.shape[0], -1])
default = {"nms_threshold": 0.3, "top_k": 10, "eta": 1.0}
fields = ['eta', 'top_k', 'nms_threshold']
for f in default.keys():
if f not in nms_param:
nms_param[f] = default[f]
out = fluid.layers.detection_output(
scores=mbox_conf_flatten,
loc=mbox_loc,
prior_box=pb,
prior_box_var=pbv,
background_label=background_label_id,
nms_threshold=nms_param["nms_threshold"],
nms_top_k=nms_param["top_k"],
keep_top_k=keep_top_k,
score_threshold=confidence_threshold,
nms_eta=nms_param["eta"])
return out
def detectionoutput_weights(name, data=None):
weights_name = []
return weights_name
register(
kind='DetectionOutput',
shape=detectionoutput_shape,
layer=detectionoutput_layer,
weights=detectionoutput_weights)
x2paddle/x2paddle/op_mapper/caffe_custom_layer/normalize.py 0 → 100644
浏览文件 @ cf7f9b88
from .register import register
from x2paddle.core.util import *
def normalize_shape(input_shape):
return input_shape
def normalize_layer(inputs,
across_spatial=None,
channel_shared=None,
input_shape=None,
name=None):
assert across_spatial == False, "Only support across_spatial == False for Normalize"
input = inputs[0]
l2_norm = fluid.layers.l2_normalize(input, axis=1, name=name + '_l2')
scale_param = fluid.layers.create_parameter(
shape=[1] if channel_shared else [1, 1, 1, input_shape[0][1]],
dtype=input.dtype,
attr=fluid.ParamAttr(name=name + '_scale'))
scale_param = fluid.layers.reshape(x=scale_param, \
shape=[1] if channel_shared else [input_shape[0][1]])
out = fluid.layers.elementwise_mul(
x=l2_norm, y=scale_param, axis=-1 if channel_shared else 1)
return out
def normalize_weights(name, data=None):
weights_name = [name + '_scale']
return weights_name
register(
kind='Normalize',
shape=normalize_shape,
layer=normalize_layer,
weights=normalize_weights)
x2paddle/x2paddle/op_mapper/caffe_custom_layer/permute.py 0 → 100644
浏览文件 @ cf7f9b88
from .register import register
from x2paddle.core.util import *
def permute_shape(input_shape, order=None):
inshape = input_shape[0]
output_shape = []
for ii in order:
assert ii < len(inshape), "invalid order for permute[%s]" % (name)
output_shape.append(inshape[ii])
return [output_shape]
def permute_layer(inputs, order=None, input_shape=None, name=None):
input = inputs[0]
order = list(order)
out = fluid.layers.transpose(input, perm=order, name=name)
return out
def permute_weights(name, data=None):
weights_name = []
return weights_name
register(
kind='Permute',
shape=permute_shape,
layer=permute_layer,
weights=permute_weights)
x2paddle/x2paddle/op_mapper/caffe_custom_layer/priorbox.py 0 → 100644
浏览文件 @ cf7f9b88
from .register import register
from x2paddle.core.util import *
def priorbox_shape(input_shape, max_size=None, aspect_ratio=None):
fc_shape = input_shape[0]
N = 1
if not max_size == None:
N += 1
if not aspect_ratio == None:
N += 2 * len(aspect_ratio)
N_bbx = fc_shape[2] * fc_shape[3] * N
output_shape = [1, 2, 4 * N_bbx]
return [output_shape]
def priorbox_layer(inputs,
step=0.0,
offset=0.5,
min_size=None,
max_size=[],
aspect_ratio=[1.0],
flip=False,
clip=False,
variance=[0.1, 0.1, 0.2, 0.2],
input_shape=None,
name=None):
input = inputs[0]
image = inputs[1]
steps = tuple(step) if type(step) is list or type(step) is tuple else (step,
step)
box, variance_ = fluid.layers.prior_box(
input,
image,
min_sizes=min_size,
max_sizes=max_size,
aspect_ratios=aspect_ratio,
variance=variance,
flip=flip,
clip=clip,
steps=steps,
offset=offset,
name=name,
min_max_aspect_ratios_order=True)
box = fluid.layers.reshape(box, [1, 1, -1])
variance_ = fluid.layers.reshape(variance_, [1, 1, -1])
out = fluid.layers.concat([box, variance_], axis=1)
return out
def priorbox_weights(name, data=None):
weights_name = []
return weights_name
register(
kind='PriorBox',
shape=priorbox_shape,
layer=priorbox_layer,
weights=priorbox_weights)
x2paddle/x2paddle/op_mapper/caffe_custom_layer/register.py 0 → 100644
浏览文件 @ cf7f9b88
""" this module provides 'register' for registering customized layers
"""
g_custom_layers = {}
def register(kind, shape, layer, weights):
""" register a custom layer or a list of custom layers
Args:
@kind (str or list): type name of the layer
@shape (function): a function to generate the shape of layer's output
@layer (function): a function to generate the paddle code of layer
@weights (function): a function to deal with weights data
Returns:
None
"""
assert type(shape).__name__ == 'function', 'shape should be a function'
assert type(layer).__name__ == 'function', 'layer should be a function'
if type(kind) is str:
kind = [kind]
else:
assert type(
kind) is list, 'invalid param "kind" for register, not a list or str'
for k in kind:
assert type(
k) is str, 'invalid param "kind" for register, not a list of str'
assert k not in g_custom_layers, 'this type[%s] has already been registered' % (
k)
g_custom_layers[k] = {
'shape': shape,
'layer': layer,
'weights': weights
}
def get_registered_layers():
return g_custom_layers
x2paddle/x2paddle/op_mapper/caffe_custom_layer/relu6.py 0 → 100644
浏览文件 @ cf7f9b88
from .register import register
from x2paddle.core.util import *
def relu6_shape(input_shape):
return input_shape
def relu6_layer(inputs, input_shape=None, name=None):
input = inputs[0]
out = fluid.layers.relu6(x=input)
return out
def relu6_weights(name, data=None):
weights_name = []
return weights_name
register(
kind='ReLU6', shape=relu6_shape, layer=relu6_layer, weights=relu6_weights)
x2paddle/x2paddle/op_mapper/caffe_custom_layer/roipooling.py 0 → 100644
浏览文件 @ cf7f9b88
from .register import register
from x2paddle.core.util import *
def roipooling_shape(input_shape, pooled_w=None, pooled_h=None):
base_fea_shape = input_shapes[0]
rois_shape = input_shapes[1]
output_shape = base_fea_shape
output_shape[0] = rois_shape[0]
output_shape[2] = pooled_h
output_shape[3] = pooled_w
return [output_shape]
def roipooling_layer(inputs,
pooled_w=None,
pooled_h=None,
spatial_scale=None,
input_shape=None,
name=None):
input = inputs[0]
roi = inputs[1]
roi = fluid.layers.slice(roi, axes=[1], starts=[1], ends=[5])
out = fluid.layers.roi_pool(
input,
roi,
pooled_height=pooled_h,
pooled_width=pooled_w,
spatial_scale=spatial_scale)
return out
def roipooling_weights(name, data=None):
weights_name = []
return weights_name
register(
kind='ROIPooling',
shape=roipooling_shape,
layer=roipooling_layer,
weights=roipooling_weights)
x2paddle/x2paddle/op_mapper/caffe_custom_layer/select.py 0 → 100644
浏览文件 @ cf7f9b88
from .register import register
from x2paddle.core.util import *
def select_shape(input_shape, axis=None, slice_point=None):
inshape = input_shape[0]
slice_point = slice_point
start = slice_point[0]
if len(slice_point) == 2:
end = slice_point[1]
else:
end = input_shape[axis]
assert end > start, "invalid slice_point with [start:%d, end:%d]" % (start,
end)
output_shape = input_shape
output_shape[axis] = end - start
return [output_shape]
def select_layer(inputs,
axis=None,
slice_point=None,
input_shape=None,
name=None):
input = inputs[0]
maxint32 = 2147483647
slice_point = [0] + slice_point
slice_point.append(maxint32)
i = 0
out = []
for i in range(len(slice_point)):
out.append(
fluid.layers.slice(
input,
axes=[axis],
starts=[slice_point[i]],
ends=[slice_point[i + 1]],
name=name + '_' + str(i)))
if i == len(slice_point) - 2:
break
return out
def select_weights(name, data=None):
weights_name = []
return weights_name
register(
kind='Select',
shape=select_shape,
layer=select_layer,
weights=select_weights)
x2paddle/x2paddle/op_mapper/caffe_custom_layer/shufflechannel.py 0 → 100644
浏览文件 @ cf7f9b88
from .register import register
from x2paddle.core.util import *
def shufflechannel_shape(input_shape):
return input_shape
def shufflechannel_layer(inputs, group=None, input_shape=None, name=None):
input = inputs[0]
out = fluid.layers.shuffle_channel(x=input, group=group)
return out
def shufflechannel_weights(name, data=None):
weights_name = []
return weights_name
register(
kind='ShuffleChannel',
shape=shufflechannel_shape,
layer=shufflechannel_layer,
weights=shufflechannel_weights)
x2paddle/x2paddle/op_mapper/caffe_custom_layer/upsample.py 0 → 100644
浏览文件 @ cf7f9b88
# -*- coding: utf-8 -*-
################################################################################
#
# Copyright (c) 2020 Baidu.com, Inc. All Rights Reserved
#
################################################################################
"""
Author: Drift
Email: wutuobang@baidu.com
Date: 2020/04/22 18:45
"""
from .register import register
from x2paddle.core.util import *
def upsample_shape(input_shapes, scale):
"""
:param input_shapes:
:param scale:
:return:
"""
assert len(input_shapes) == 1, "not valid input shape for upsample layer"
assert type(scale) is int
input_shape = input_shapes[0]
new_h = scale * input_shape[2]
new_w = scale * input_shape[3]
output_shape = [input_shape[0], input_shape[1], new_h, new_w]
return [output_shape]
def upsample_layer(inputs, scale, input_shape=None, name=None):
"""
:param inputs:
:param scale:
:param input_shape:
:param name:
:return:
"""
x = inputs[0]
out = fluid.layers.resize_nearest(
x, align_corners=False, scale=scale, name=name)
return out
def upsample_weights(name, data=None):
"""
:param name:
:param data:
:return:
"""
weights_name = []
return weights_name
register(
kind='Upsample',
shape=upsample_shape,
layer=upsample_layer,
weights=upsample_weights)
x2paddle/x2paddle/op_mapper/caffe_op_mapper.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numbers
import numpy as np
from x2paddle.decoder.caffe_decoder import CaffeGraph
from x2paddle.core.op_mapper import OpMapper
from x2paddle.core.util import *
from x2paddle.op_mapper import caffe_shape
from x2paddle.op_mapper.caffe_custom_layer import *
class CaffeOpMapper(OpMapper):
directly_map_ops = {
'AbsVal': 'abs',
'Sigmoid': 'sigmoid',
'TanH': 'tanh',
}
def __init__(self, decoder):
super(CaffeOpMapper, self).__init__()
self.graph = decoder.caffe_graph
self.weights = dict()
resolver = decoder.resolver
self.used_custom_layers = {}
print("Total nodes: {}".format(len(self.graph.topo_sort)))
for node_name in self.graph.topo_sort:
node = self.graph.get_node(node_name)
if node.layer_type == 'DepthwiseConvolution':
node.layer_type = 'ConvolutionDepthwise'
op = node.layer_type
if hasattr(self, op):
self.set_node_shape(node)
func = getattr(self, op)
func(node)
elif op in custom_layers:
self.set_node_shape(node, is_fluid_op=False)
self.deal_custom_layer(node)
elif op in self.directly_map_ops:
self.set_node_shape(node)
self.directly_map(node)
else:
raise Exception(
"The op {} in model is not supported yet.".format(op))
def op_checker(self):
unsupported_ops = set()
for node_name in self.graph.topo_sort:
node = self.graph.get_node(node_name)
op = node.layer_type
if not hasattr(self, op) and op not in custom_layers:
unsupported_ops.add(op)
if len(unsupported_ops) == 0:
return True
else:
print("There are {} ops not supported yet, list as below".format(
len(unsupported_ops)))
for op in unsupported_ops:
print(op)
return False
def set_node_shape(self, node, is_fluid_op=True):
inputs = node.inputs
input_shape = []
for i, nm in enumerate(inputs):
last_node = self.graph.get_node(nm)
tmp = node.layer.bottom[i]
idx = list(last_node.layer.top).index(tmp)
input_shape.append(last_node.output_shape[idx])
node.input_shape = input_shape
func_name = 'shape_' + node.layer_type.lower()
if is_fluid_op:
node.output_shape = getattr(caffe_shape, func_name)(node.layer,
input_shape)
else:
node.output_shape = compute_output_shape(node)
def adjust_parameters(self, node):
data = node.data
# When using the protobuf-backend, each parameter initially has four dimensions.
# In certain cases (like FC layers), we want to eliminate the singleton dimensions.
# This implementation takes care of the common cases. However, it does leave the
# potential for future issues.
# The Caffe-backend does not suffer from this problem.
data = list(data)
squeeze_indices = [1] # Squeeze biases.
if node.layer_type == 'InnerProduct':
squeeze_indices.append(0) # Squeeze FC.
for idx in squeeze_indices:
if idx >= len(data):
continue
d = data[idx]
assert len(
d.shape
) == 4, 'invalid shape[%s] from caffe when adjust_parameters' % (
str(d.shape))
shape_old = d.shape
sq_axis = None
if idx == 0:
sq_axis = (0, 1)
elif idx == 1:
sq_axis = (0, 1, 2)
else:
continue
data[idx] = np.squeeze(d, axis=sq_axis)
shape_new = data[idx].shape
return data
def get_kernel_parameters(self, kind, params):
assert kind in ['Convolution', 'Pooling', 'Deconvolution']
[k_h, k_w] = [1, 1]
if isinstance(params.kernel_size, numbers.Number):
[k_h, k_w] = [params.kernel_size] * 2
elif len(params.kernel_size) > 0:
k_h = params.kernel_h if params.kernel_h > 0 else params.kernel_size[
0]
k_w = params.kernel_w if params.kernel_w > 0 else params.kernel_size[
len(params.kernel_size) - 1]
elif params.kernel_h > 0 or params.kernel_w > 0:
k_h = params.kernel_h
k_w = params.kernel_w
[s_h, s_w] = [1, 1]
if isinstance(params.stride, numbers.Number):
[s_h, s_w] = [params.stride] * 2
elif len(params.stride) > 0:
s_h = params.stride_h if params.stride_h > 0 else params.stride[0]
s_w = params.stride_w if params.stride_w > 0 else params.stride[len(
params.stride) - 1]
elif params.stride_h > 0 or params.stride_w > 0:
s_h = params.stride_h
s_w = params.stride_w
[p_h, p_w] = [0, 0]
if isinstance(params.pad, numbers.Number):
[p_h, p_w] = [params.pad] * 2
elif len(params.pad) > 0:
p_h = params.pad_h if params.pad_h > 0 else params.pad[0]
p_w = params.pad_w if params.pad_w > 0 else params.pad[len(
params.pad) - 1]
elif params.pad_h > 0 or params.pad_w > 0:
p_h = params.pad_h
p_w = params.pad_w
dila_h = dila_w = 1
group = 1
c_o = 1
if kind in ['Convolution', 'Deconvolution']:
c_o = params.num_output
dila_len = len(params.dilation)
if dila_len == 2:
dila_h = params.dilation[0]
dila_w = params.dilation[1]
elif dila_len == 1:
dila_h = dila_w = params.dilation[0]
else:
assert dila_len == 0, "invalid length[%s] of dilation in convolution" % (
dila_len)
if kind in ['Convolution', 'Deconvolution']:
group = params.group
kernel = [k_h, k_w]
stride = [s_h, s_w]
pad = [p_h, p_w]
dilation = [dila_h, dila_w]
return c_o, kernel, stride, pad, dilation, group
def get_input_name(self, node):
if hasattr(node, "index"):
return node.layer_name + "[{}]".format(node.index)
else:
return node.layer_name
def Input(self, node):
shape = list(node.layer.input_param.shape[0].dim)[1:]
dtype = 'float32'
attr = {
'dtype': string(dtype),
'shape': shape,
'name': string(node.layer_name)
}
node.fluid_code.add_layer(
"data", inputs=None, output=node, param_attr=attr)
def MemoryData(self, node):
# TODO(syf): Paddlepaddle can't fully support
shape = node.output_shape[0][1:]
dtype = 'float32'
attr = {
'dtype': string(dtype),
'shape': shape,
'name': string(node.layer_name)
}
node.fluid_code.add_layer(
"data", inputs=None, output=node.layer_name + '0', param_attr=attr)
node.fluid_code.add_note('{} = [{}]'.format(node.layer_name,
node.layer_name + '0'))
def Convolution(self, node):
data = node.data
params = node.layer.convolution_param
channel, kernel, stride, pad, dilation, group = self.get_kernel_parameters(
node.layer_type, params)
if data is None:
data = []
print(
'The parameter of {} (type is {}) is not set. So we set the parameters as 0'
.format(node.layer_name, node.layer_type))
input_c = node.input_shape[0][1]
output_c = channel
data.append(
np.zeros([output_c, input_c, kernel[0], kernel[1]]).astype(
'float32'))
data.append(np.zeros([output_c, ]).astype('float32'))
else:
data = self.adjust_parameters(node)
self.weights[node.layer_name + '_weights'] = data[0]
if len(data) == 2:
self.weights[node.layer_name + '_bias'] = data[1]
assert len(node.inputs
) == 1, 'The count of Convolution node\'s input is not 1.'
input = self.graph.get_bottom_node(node, idx=0, copy=True)
attr = {
'filter_size': kernel,
'num_filters': channel,
'stride': stride,
'padding': pad,
'dilation': dilation,
'groups': group,
'name': string(node.layer_name),
'param_attr': string(node.layer_name + '_weights'),
'bias_attr': False
if len(data) == 1 else string(node.layer_name + '_bias'),
}
node.fluid_code.add_layer(
"conv2d", inputs=input, output=node, param_attr=attr)
def Deconvolution(self, node):
data = node.data
params = node.layer.convolution_param
channel, kernel, stride, pad, dilation, group = self.get_kernel_parameters(
node.layer_type, params)
if data is None:
data = []
print(
'The parameter of {} (type is {}) is not set. So we set the parameters as 0'
.format(node.layer_name, node.layer_type))
input_c = node.input_shape[0][1]
output_c = channel
data.append(
np.zeros([output_c, input_c, kernel[0], kernel[1]]).astype(
'float32'))
data.append(np.zeros([output_c, ]).astype('float32'))
else:
data = self.adjust_parameters(node)
self.weights[node.layer_name + '_weights'] = data[0]
if len(data) == 2:
self.weights[node.layer_name + '_bias'] = data[1]
assert len(node.inputs
) == 1, 'The count of Deconvolution node\'s input is not 1.'
input = self.graph.get_bottom_node(node, idx=0, copy=True)
attr = {
'output_size': None,
'filter_size': kernel,
'num_filters': channel,
'stride': stride,
'padding': pad,
'dilation': dilation,
'groups': group,
'name': string(node.layer_name),
'param_attr': string(node.layer_name + '_weights'),
'bias_attr': False
if len(data) == 1 else string(node.layer_name + '_bias')
}
node.fluid_code.add_layer(
"conv2d_transpose", inputs=input, output=node, param_attr=attr)
def Pooling(self, node):
params = node.layer.pooling_param
ceil_mode = getattr(params, 'ceil_mode', True)
global_pool = getattr(params, 'global_pooling', False)
kernel_default = [1, 1]
channel, kernel, stride, pad, dilation, group = self.get_kernel_parameters(
node.layer_type, params)
if params.pool == 0:
pool_type = 'max'
else:
pool_type = 'avg'
assert len(
node.inputs) == 1, 'The count of Pooling node\'s input is not 1.'
input = self.graph.get_bottom_node(node, idx=0, copy=True)
attr = {
'pool_size': kernel,
'pool_stride': stride,
'pool_padding': pad,
'ceil_mode': ceil_mode,
'pool_type': string(pool_type),
'exclusive': False,
'global_pooling': global_pool,
'name': string(node.layer_name)
}
node.fluid_code.add_layer(
"pool2d", inputs=input, output=node, param_attr=attr)
def LRN(self, node):
assert len(node.inputs) == 1, 'The count of LRN node\'s input is not 1.'
params = node.layer.lrn_param
# The window size must be an odd value. For a window
# size of (2*n+1), Paddle defines depth_radius = n.
assert params.local_size % 2 == 1
# Caffe scales by (alpha/(2*n+1)), whereas Paddle
# just scales by alpha (as does Krizhevsky's paper).
# We'll account for that here.
alpha = params.alpha / float(params.local_size)
input = self.graph.get_bottom_node(node, idx=0, copy=True)
attr = {
'n': params.local_size,
'k': params.k,
'alpha': alpha,
'beta': params.beta,
'name': string(node.layer_name)
}
node.fluid_code.add_layer(
"lrn", inputs=input, output=node, param_attr=attr)
def InnerProduct(self, node):
data = node.data
params = node.layer.inner_product_param
if data is None:
print(
'The parameter of {} (type is {}) is not set. So we set the parameters as 0.'
.format(node.layer_name, node.layer_type))
input_c = node.input_shape[0][1]
output_c = params.num_output
data = []
data.append(
np.zeros([input_c, output_c]).astype('float32').astype(
'float32'))
data.append(
np.zeros([output_c]).astype('float32').astype('float32'))
else:
data = self.adjust_parameters(node)
# Reshape the parameters to Paddle's ordering
transpose_order = (1, 0)
w = data[0]
fc_shape = w.shape
output_channels = fc_shape[0]
w = w.reshape((output_channels, -1))
w = w.transpose(transpose_order)
data[0] = w
self.weights[node.layer_name + '_weights'] = data[0]
if len(data) == 2:
self.weights[node.layer_name + '_bias'] = data[1]
assert len(node.inputs
) == 1, 'The count of InnerProduct node\'s input is not 1.'
#params = node.layer.inner_product_param
assert params.axis == 1
assert params.bias_term == True
input = self.graph.get_bottom_node(node, idx=0, copy=True)
attr = {
'size': params.num_output,
'name': string(node.layer_name),
'act': None,
'param_attr': string(node.layer_name + '_weights'),
'bias_attr': False
if len(data) == 1 else string(node.layer_name + '_bias')
}
node.fluid_code.add_layer(
"fc", inputs=input, output=node, param_attr=attr)
def Softmax(self, node):
assert len(
node.inputs) == 1, 'The count of Softmax node\'s input is not 1.'
input = self.graph.get_bottom_node(node, idx=0, copy=True)
params = node.layer.softmax_param
axis = params.axis
shape = node.input_shape[0]
dims = len(shape)
axis = axis + dims if axis < 0 else axis
attr = {'axis': axis, 'name': string(node.layer_name + '_softmax')}
node.fluid_code.add_layer(
"softmax", inputs=input, output=node, param_attr=attr)
def Slice(self, node):
assert len(
node.inputs) == 1, 'The count of Slice node\'s input is not 1.'
input = self.graph.get_bottom_node(node, idx=0, copy=True)
top_len = len(node.layer.top)
params = node.layer.slice_param
axis = params.axis
slice_dim = params.slice_dim
if slice_dim != 1 and axis == 1:
axis = slice_dim
output_shape = node.output_shape
sections_list = []
for s in output_shape:
sections_list.append(s[axis])
attr = {
'num_or_sections': sections_list,
'dim': axis,
'name': string(node.layer_name)
}
node.fluid_code.add_layer(
"split", inputs=input, output=node.layer_name, param_attr=attr)
def Concat(self, node):
assert len(
node.inputs
) >= 1, 'The count of Concat node\'s input is not more than 1.'
inputs = []
for i in range(len(node.inputs)):
input = self.graph.get_bottom_node(node, idx=i, copy=True)
inputs.append(input)
params = node.layer.concat_param
axis = params.axis
attr = {'axis': axis, 'name': string(node.layer_name)}
node.fluid_code.add_layer(
"concat", inputs=inputs, output=node, param_attr=attr)
def ReLU(self, node):
"""
:param node:
:return:
"""
assert len(
node.inputs) == 1, 'The count of ReLU node\'s input is not 1.'
input = self.graph.get_bottom_node(node, idx=0, copy=True)
params = node.layer.relu_param
if params.HasField('negative_slope') and params.negative_slope != 0:
negative_slope = float(params.negative_slope)
attr = {'alpha': negative_slope}
node.fluid_code.add_layer(
'leaky_relu', inputs=input, output=node, param_attr=attr)
else:
node.fluid_code.add_layer('relu', inputs=input, output=node)
def PReLU(self, node):
assert len(
node.inputs) == 1, 'The count of PReLU node\'s input is not 1.'
input = self.graph.get_bottom_node(node, idx=0, copy=True)
params = node.layer.prelu_param
mode_bool = params.channel_shared
if mode_bool:
mode = 'all'
else:
mode = 'channel'
data = node.data
assert data is not None, 'The parameter of {} (type is {}) is not set. You need to use python package of caffe to set the default value.'.format(
node.layer_name, node.layer_type)
self.weights[node.layer_name + '_weights'] = data[0]
attr = {
'mode': string(mode),
'param_attr': string(node.layer_name + '_weights'),
'name': string(node.layer_name)
}
node.fluid_code.add_layer(
"prelu", inputs=input, output=node, param_attr=attr)
def Accuracy(self, node):
assert len(
node.inputs) == 2, 'The count of Accuracy node\'s input is not 2.'
inputs = []
inputs[0] = None
inputs[1] = None
i = 0
for shape in node.input_shape:
if shape[1] == 1:
input = self.graph.get_bottom_node(node, idx=i, copy=True)
inputs[1] = input
else:
input = self.graph.get_bottom_node(node, idx=i, copy=True)
inputs[0] = input
i += 1
params = node.layer.accuracy_param
top_k = params.top_k
axis = params.axis
ignore_label = params.ignore_label
assert axis == 1, 'PaddlePaddle can not support the situation when the axis is not 1.'
assert not ignore_label >= 0, 'PaddlePaddle can not support the situation when the model has ignore label.'
attr = {'k': top_k}
node.fluid_code.add_layer(
"accuracy", inputs=inputs, output=node, param_attr=attr)
def Eltwise(self, node):
assert len(
node.inputs) == 2, 'The count of TanH node\'s input is not 2.'
params = node.layer.eltwise_param
mode = params.operation
inputs = []
input0 = self.graph.get_bottom_node(node, idx=0, copy=True)
inputs.append(input0)
input1 = self.graph.get_bottom_node(node, idx=1, copy=True)
inputs.append(input1)
if mode == 0:
inputs_dict = {}
inputs_dict['x'] = inputs[0]
inputs_dict['y'] = inputs[1]
attr = {'act': None, 'name': string(node.layer_name)}
node.fluid_code.add_layer(
"elementwise_mul",
inputs=inputs_dict,
output=node,
param_attr=attr)
elif mode == 1:
if hasattr(params, 'coeff') and len(params.coeff) == 2:
coeff = params.coeff
input1_name = self.get_input_name(inputs[0])
attr = {
'shape': [1],
'value': coeff[0],
'dtype': '{}.dtype'.format(input1_name)
}
node.fluid_code.add_layer(
"fill_constant",
inputs=None,
output=node.layer_name + '_const1',
param_attr=attr)
attr = {'act': None, 'name': string(node.layer_name + '_mul1')}
node.fluid_code.add_layer(
"elementwise_mul",
inputs=input1_name + ', ' + node.layer_name + '_const1',
output=node.layer_name + '_mul1',
param_attr=attr)
input2_name = self.get_input_name(inputs[1])
attr = {
'shape': [1],
'value': coeff[1],
'dtype': '{}.dtype'.format(input2_name)
}
node.fluid_code.add_layer(
"fill_constant",
inputs=None,
output=node.layer_name + '_const2',
param_attr=attr)
attr = {'act': None, 'name': string(node.layer_name + '_mul2')}
node.fluid_code.add_layer(
"elementwise_mul",
inputs=input2_name + ', ' + node.layer_name + '_const2',
output=node.layer_name + '_mul2',
param_attr=attr)
attr = {'act': None, 'name': string(node.layer_name)}
node.fluid_code.add_layer(
"elementwise_add",
inputs='{}_mul1, {}_mul2'.format(node.layer_name,
node.layer_name),
output=node,
param_attr=attr)
else:
inputs_dict = {}
inputs_dict['x'] = inputs[0]
inputs_dict['y'] = inputs[1]
attr = {'act': None, 'name': string(node.layer_name)}
node.fluid_code.add_layer(
"elementwise_add",
inputs=inputs_dict,
output=node,
param_attr=attr)
else:
inputs_dict = {}
inputs_dict['x'] = inputs[0]
inputs_dict['y'] = inputs[1]
attr = {'act': None, 'name': string(node.layer_name)}
node.fluid_code.add_layer(
"elementwise_max",
inputs=inputs_dict,
output=node,
param_attr=attr)
def BatchNorm(self, node):
assert len(
node.inputs) == 1, 'The count of BatchNorm node\'s input is not 1.'
input = self.graph.get_bottom_node(node, idx=0, copy=True)
params = node.layer.batch_norm_param
if hasattr(params, 'eps'):
eps = params.eps
else:
eps = 1e-5
if node.data is None or len(node.data) != 3:
print(
'The parameter of {} (type is {}) is not set. So we set the parameters as 0'
.format(node.layer_name, node.layer_type))
input_c = node.input_shape[0][1]
mean = np.zeros([input_c, ]).astype('float32')
variance = np.zeros([input_c, ]).astype('float32')
scale = 0
else:
node.data = [np.squeeze(i).astype('float32') for i in node.data]
mean, variance, scale = node.data
# Prescale the stats
scaling_factor = 1.0 / scale if scale != 0 else 0
mean *= scaling_factor
variance *= scaling_factor
self.weights[node.layer_name + '_mean'] = mean
self.weights[node.layer_name + '_variance'] = variance
attr = {
'is_test': True,
'param_attr': None,
'bias_attr': None,
'moving_mean_name': string(node.layer_name + '_mean'),
'moving_variance_name': string(node.layer_name + '_variance'),
'epsilon': eps,
'name': string(node.layer_name)
}
node.fluid_code.add_layer(
"batch_norm", inputs=input, output=node, param_attr=attr)
def Scale(self, node):
if node.data is None:
print(
'The parameter of {} (type is {}) is not set. So we set the parameters as 0'
.format(node.layer_name, node.layer_type))
input_c = node.input_shape[0][1]
self.weights[node.layer_name + '_scale'] = np.zeros([
input_c,
]).astype('float32')
self.weights[node.layer_name + '_offset'] = np.zeros([
input_c,
]).astype('float32')
else:
self.weights[node.layer_name + '_scale'] = np.squeeze(node.data[
0]).astype('float32')
self.weights[node.layer_name + '_offset'] = np.squeeze(node.data[
1]).astype('float32')
params = node.layer.scale_param
axis = params.axis
num_axes = params.num_axes
inputs = []
if len(node.inputs) == 2:
# for two tensor, here resets axis to 1. Maybe there is a bug for unkown case.
axis = 1
bias_shape = node.input_shape[0][axis:axis + num_axes]
input0 = self.graph.get_bottom_node(node, idx=0, copy=True)
input1 = self.graph.get_bottom_node(node, idx=1, copy=True)
inputs_dict = {}
inputs_dict['x'] = input0
inputs_dict['y'] = input1
attr = {'axis': axis, 'name': string(node.layer_name + '_mul')}
node.fluid_code.add_layer(
"elementwise_mul",
inputs=inputs_dict,
output=node.layer_name + '_mul',
param_attr=attr)
else:
bias_shape = node.input_shape[0][axis:axis + num_axes]
input0 = self.graph.get_bottom_node(node, idx=0, copy=True)
input0_name = self.get_input_name(input0)
attr = {
'dtype': '{}.dtype'.format(input0_name),
'shape': bias_shape,
'name': string(node.layer_name + '_cparam1'),
'attr': string(node.layer_name + '_scale'),
'is_bias': True,
'default_initializer': 'Constant(value=1.0)'
}
node.fluid_code.add_layer(
"create_parameter", inputs=None, output=node, param_attr=attr)
inputs_dict = {}
inputs_dict['x'] = input0
inputs_dict['y'] = node
attr = {'axis': axis, 'name': string(node.layer_name + '_mul')}
node.fluid_code.add_layer(
"elementwise_mul",
inputs=inputs_dict,
output=node.layer_name + '_mul',
param_attr=attr)
scale_shape = bias_shape
input0_name = self.get_input_name(input0)
attr = {
'dtype': '{}.dtype'.format(input0_name),
'shape': scale_shape,
'name': string(node.layer_name + '_cparam2'),
'attr': string(node.layer_name + '_offset'),
'is_bias': True,
'default_initializer': 'Constant(value=1.0)'
}
node.fluid_code.add_layer(
"create_parameter",
inputs=None,
output=node.layer_name + '_offset_param',
param_attr=attr)
attr = {'axis': axis, 'name': string(node.layer_name + '_add')}
node.fluid_code.add_layer(
"elementwise_add",
inputs='{}_mul, {}_offset_param'.format(node.layer_name,
node.layer_name),
output=node,
param_attr=attr)
def Reshape(self, node):
input = self.graph.get_bottom_node(node, idx=0, copy=True)
top_count = len(input.layer.top)
is_inplace = False if top_count == 1 else True
output_shape = node.output_shape[0]
attr = {
'shape': output_shape,
'inplace': is_inplace,
'act': None,
'name': string(node.layer_name)
}
node.fluid_code.add_layer(
"reshape", inputs=input, output=node, param_attr=attr)
def ArgMax(self, node):
assert len(node.inputs) == 1 and len(
node.outputs
) == 1, 'The count of ArgMax node\'s input and output is not 1.'
input = self.graph.get_bottom_node(node, idx=0, copy=True)
input_shape = node.input_shape[0]
params = node.layer.argmax_param
out_max_val = params.out_max_val if hasattr(params,
out_max_val) else False
top_k = params.top_k if hasattr(params, top_k) else 1
axis = parmas.axis if hasattr(params, axis) else -1
if axis < 0:
axis += len(input_shape)
if out_max_val is True:
attr = {'k': top_k, 'name': string(node.layer_name + '_topk')}
node.fluid_code.add_layer(
"topk",
inputs=input,
output='{}_topk_var, {}_index_var'.format(node.layer_name,
node.layer_name),
param_attr=attr)
attr = {'dtype': '{}_topk_var.dtype'.format(node.layer_name)}
node.fluid_code.add_layer(
"cast",
inputs='{}_index_var'.format(node.layer_name),
output='{}_index_var'.format(node.layer_name),
param_attr=attr)
attr = {'axis': axis, 'name': string(node.layer_name)}
node.fluid_code.add_layer(
"concat",
inputs='{}_topk_var, {}_index_var'.format(node.layer_name,
node.layer_name),
output=node,
param_attr=attr)
else:
attr = {'k': top_k, 'name': string(node.layer_name)}
node.fluid_code.add_layer(
"topk",
inputs=input,
output='_, {}'.format(node.layer_name),
param_attr=attr)
def Crop(self, node):
assert len(
node.inputs) == 2, 'The count of Crop node\'s input is not 2.'
input = self.graph.get_bottom_node(node, idx=0, copy=True)
example = self.graph.get_bottom_node(node, idx=1, copy=True)
params = node.layer.crop_param
axis = params.axis
input_shape = node.input_shape[0]
if axis < 0:
axis += len(input_shape)
offset_real = [0] * len(input_shape)
if hasattr(params, "offset") and len(params.offset) > 0:
offset = list(params.offset)
assert (len(input_shape) - axis
) == len(offset), "invalid offset[%s] in crop layer" % (
str(offset))
offset_real = [0] * axis + offset
attr = {'offsets': list(offset_real), 'name': string(node.layer_name)}
node.fluid_code.add_layer(
"crop",
inputs={'x': input,
'shape': node.input_shape[1]},
output=node,
param_attr=attr)
def Flatten(self, node):
assert len(
node.
inputs) == 1, 'The count of DetectionOutput node\'s input is not 1.'
input = self.graph.get_bottom_node(node, idx=0, copy=True)
shape = node.output_shape[0]
attr = {'shape': shape, 'name': string(node.layer_name)}
node.fluid_code.add_layer(
"reshape", inputs=input, output=node, param_attr=attr)
def Power(self, node):
assert len(
node.inputs) == 1, 'The count of Permute node\'s input is not 1.'
input = self.graph.get_bottom_node(node, idx=0, copy=True)
params = node.layer.power_param
power = params.power
scale = params.scale
shift = params.shift
attr = {
'scale': scale,
'bias': shift,
'bias_after_scale': True,
'name': string(node.layer_name + '_scale')
}
node.fluid_code.add_layer(
"scale", inputs=input, output=node, param_attr=attr)
attr = {'factor': power, 'name': string(node.layer_name)}
node.fluid_code.add_layer(
"pow", inputs=node, output=node, param_attr=attr)
def Reduction(self, node):
assert len(
node.inputs) == 1, 'The count of Reduction node\'s input is not 1.'
input = self.graph.get_bottom_node(node, idx=0, copy=True)
params = node.layer.reduction_param
operation = params.operation
axis = params.axis
coeff = params.coeff
assert operation >= 1 and operation <= 4, "reduction reduction [%s] error" % (
operation)
input_len = len(node.input_shape[0])
if axis < 0:
axis += input_len + 1
dim = list(range(input_len))
if operation == 1: ## operation = SUM
attr = {
'dim': dim[axis:],
'keep_dim': False,
'name': string(node.layer_name)
}
node.fluid_code.add_layer(
"reduce_sum", inputs=input, output=node, param_attr=attr)
elif operation == 2: ## operation = ASUM
attr = {'name': string(node.layer_name + '_abs')}
node.fluid_code.add_layer(
"abs", inputs=input, output=node, param_attr=attr)
attr = {
'dim': dim[axis:],
'keep_dim': False,
'name': string(node.layer_name)
}
node.fluid_code.add_layer(
"reduce_sum", inputs=node, output=node, param_attr=attr)
elif operation == 3: ## operation = SUMSQ
attr = {'factor': 2.0, 'name': string(node.layer_name + '_pow')}
node.fluid_code.add_layer(
"pow", inputs=input, output=node, param_attr=attr)
attr = {
'dim': dim[axis:],
'keep_dim': False,
'name': string(node.layer_name)
}
node.fluid_code.add_layer(
"reduce_sum", inputs=node, output=node, param_attr=attr)
else: ## operation = MEAN
attr = {
'dim': dim[axis:],
'keep_dim': False,
'name': string(node.layer_name)
}
node.fluid_code.add_layer(
"reduce_mean", inputs=node, output=node, param_attr=attr)
attr = {'scale': coeff}
node.fluid_code.add_layer(
"scale", inputs=node, output=node, param_attr=attr)
def deal_custom_layer(self, node):
op = node.layer_type
custom_code, func = make_custom_layer(node)
params = get_params(node.layer, node.layer_type)
arg_names, kwargs = set_args(func, params)
kwargs['name'] = string(node.layer_name)
kwargs['input_shape'] = node.input_shape
data = node.data
if data is not None:
data = self.adjust_parameters(node)
weights_name = deal_weights(node)
for i in range(len(data)):
self.weights[weights_name[i]] = data[i]
inputs_node = []
for i in range(len(node.inputs)):
input = self.graph.get_bottom_node(node, idx=i, copy=True)
if i == 1 and op == 'DetectionOutput':
input = self.graph.get_bottom_node(node, idx=i, copy=True)
while input is not None \
and input.layer_type != 'Softmax' \
and input.layer_type != 'Sigmoid':
input = self.graph.get_bottom_node(input, idx=0, copy=True)
assert input is not None, 'This kind of DetectionOutput is not supported!'
input = self.graph.get_bottom_node(input, idx=0, copy=True)
inputs_node.append(input)
node.fluid_code.add_layer(
func.__code__.co_name,
inputs=inputs_node,
output=node,
param_attr=kwargs,
is_custom_layer=True)
if op not in self.used_custom_layers:
self.used_custom_layers[op] = custom_code
def directly_map(self, node):
assert node.layer_type in self.directly_map_ops
op_info = self.directly_map_ops[node.layer_type]
input = self.graph.get_bottom_node(node, idx=0, copy=True)
attr = {'name': string(node.layer_name)}
node.fluid_code.add_layer(
op_info, inputs=input, output=node, param_attr=attr)
x2paddle/x2paddle/op_mapper/caffe_shape.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
import numbers
from functools import reduce
def get_kernel_parameters(params):
[k_h, k_w] = [1, 1]
if isinstance(params.kernel_size, numbers.Number):
[k_h, k_w] = [params.kernel_size] * 2
elif len(params.kernel_size) > 0:
k_h = params.kernel_h if params.kernel_h > 0 else params.kernel_size[0]
k_w = params.kernel_w if params.kernel_w > 0 else params.kernel_size[
len(params.kernel_size) - 1]
elif params.kernel_h > 0 or params.kernel_w > 0:
k_h = params.kernel_h
k_w = params.kernel_w
[s_h, s_w] = [1, 1]
if isinstance(params.stride, numbers.Number):
[s_h, s_w] = [params.stride] * 2
elif len(params.stride) > 0:
s_h = params.stride_h if params.stride_h > 0 else params.stride[0]
s_w = params.stride_w if params.stride_w > 0 else params.stride[len(
params.stride) - 1]
elif params.stride_h > 0 or params.stride_w > 0:
s_h = params.stride_h
s_w = params.stride_w
[p_h, p_w] = [0, 0]
if isinstance(params.pad, numbers.Number):
[p_h, p_w] = [params.pad] * 2
elif len(params.pad) > 0:
p_h = params.pad_h if params.pad_h > 0 else params.pad[0]
p_w = params.pad_w if params.pad_w > 0 else params.pad[len(params.pad) -
1]
elif params.pad_h > 0 or params.pad_w > 0:
p_h = params.pad_h
p_w = params.pad_w
dila_h = dila_w = 1
if hasattr(params, 'dilation'):
dila_len = len(params.dilation)
if dila_len == 2:
dila_h = params.dilation[0]
dila_w = params.dilation[1]
elif dila_len == 1:
dila_h = dila_w = params.dilation[0]
else:
assert dila_len == 0, "invalid length[%s] of dilation in convolution" % (
dila_len)
return dila_h, dila_w, p_h, p_w, k_h, k_w, s_h, s_w
def get_strided_kernel_output_shape(params, input_shape, round_func):
i_h = input_shape[2]
i_w = input_shape[3]
dila_h, dila_w, pad_h, pad_w, kernel_h, kernel_w, stride_h, stride_w = get_kernel_parameters(
params)
o_h = (i_h + 2 * pad_h - (dila_h *
(kernel_h - 1) + 1)) / float(stride_h) + 1
o_w = (i_w + 2 * pad_w - (dila_w *
(kernel_w - 1) + 1)) / float(stride_w) + 1
o_h = int(round_func(o_h))
o_w = int(round_func(o_w))
has_c_o = hasattr(params, 'num_output')
c = params.num_output if has_c_o else input_shape[1]
return [[input_shape[0], c, o_h, o_w]]
def shape_convolution(layer, input_shape):
params = layer.convolution_param
return get_strided_kernel_output_shape(params, input_shape[0], math.floor)
def shape_deconvolution(layer, input_shape):
h_i = input_shape[0][2]
w_i = input_shape[0][3]
params = layer.convolution_param
dila_h, dila_w, pad_h, pad_w, kernel_h, kernel_w, stride_h, stride_w = get_kernel_parameters(
params)
h_o = (h_i - 1) * stride_h - 2 * pad_h + dila_h * (kernel_h - 1) + 1
w_o = (w_i - 1) * stride_w - 2 * pad_w + dila_w * (kernel_w - 1) + 1
has_c_o = hasattr(params, 'num_output')
c = params.num_output if has_c_o else input_shape.channels
return [[input_shape[0][0], c, h_o, w_o]]
def shape_pooling(layer, input_shape):
params = layer.pooling_param
global_pool = getattr(params, 'global_pooling', False)
if global_pool:
return [[input_shape[0][0], input_shape[0][1], 1, 1]]
ceil_mode = getattr(params, 'ceil_mode', True)
if ceil_mode is True:
method = math.ceil
else:
method = math.floor
return get_strided_kernel_output_shape(params, input_shape[0], method)
def shape_innerproduct(layer, input_shape):
params = layer.inner_product_param
return [[input_shape[0][0], params.num_output]]
def shape_lrn(layer, input_shape):
return input_shape
def shape_relu(layer, input_shape):
return input_shape
def shape_softmax(layer, input_shape):
return input_shape
def shape_input(layer, input_shape):
return [list(layer.input_param.shape[0].dim)]
def shape_memorydata(layer, input_shape):
params = layer.memory_data_param
shape = []
shape.append(int(params.batch_size))
shape.append(int(params.channels))
shape.append(int(params.height))
shape.append(int(params.width))
return [shape]
def shape_concat(layer, input_shape):
params = layer.concat_param
axis = params.axis
output_shape = None
for shape in input_shape:
if output_shape is None:
output_shape = []
for i in range(len(shape)):
output_shape.append(shape[i])
else:
output_shape[axis] += shape[axis]
return [output_shape]
def shape_slice(layer, input_shape):
inshape = input_shape[0]
top_len = len(layer.top)
params = layer.slice_param
axis = params.axis
slice_dim = params.slice_dim
if slice_dim != 1 and axis == 1:
axis = slice_dim
points = list(params.slice_point)
count = inshape[axis]
if len(points) == 0:
assert count % top_len == 0, "the parameter of Slice is wrong"
part = count / top_len
t = part
while t < count:
points.append(int(t))
t += part
points = [0] + points + [count]
output_shape = []
for i in range(len(points)):
shape = []
for ii in range(len(inshape)):
shape.append(inshape[ii])
size = points[i + 1] - points[i]
shape[axis] = size
output_shape.append(shape)
if i == len(points) - 2:
break
return output_shape
def shape_prelu(layer, input_shape):
return input_shape
def shape_sigmoid(layer, input_shape):
return input_shape
def shape_absval(layer, input_shape):
return input_shape
def shape_accuracy(layer, input_shape):
return [[1]]
def shape_tanh(layer, input_shape):
return input_shape
def shape_eltwise(layer, input_shape):
return [input_shape[0]]
def shape_batchnorm(layer, input_shape):
return input_shape
def shape_scale(layer, input_shape):
return input_shape
def shape_reshape(layer, input_shape):
def count(num_list):
return reduce(lambda a, b: a * b, num_list)
inshape = input_shape[0]
params = layer.reshape_param
axis = params.axis if hasattr(params, 'axis') else 0
num_axes = params.num_axes if hasattr(params, 'num_axes') else -1
if inshape[0] == -1:
inshape[0] = 1
input_count = count(inshape)
input_num_axes = len(inshape)
input_start_axis = axis
start_axis = input_start_axis if input_start_axis >= 0 \
else input_num_axes + input_start_axis + 1
assert start_axis >= 0, "[Reshape]axis %d out of range" % (input_start_axis)
assert start_axis <= input_num_axes, "[Reshape]axis %d out of range for %d-D input data"\
% (input_start_axis, input_num_axes)
assert num_axes >= -1, "[Reshape]num_axes must be >= 0, or -1 for all"
end_axis = input_num_axes if num_axes == -1 else start_axis + num_axes
assert end_axis <= input_num_axes, "end_axis[%d] = axis[%d] + num_axes[%d] is out of range"\
% (end_axis, start_axis, num_axes)
num_axes_replaced = end_axis - start_axis
num_axes_retained = input_num_axes - num_axes_replaced
num_new_axes = len(list(params.shape.dim))
outshape = []
for i in range(start_axis):
outshape.append(inshape[i])
for i in range(num_new_axes):
outshape.append(params.shape.dim[i])
for i in range(end_axis, input_num_axes):
outshape.append(inshape[i])
assert len(outshape) == num_axes_retained + num_new_axes,\
"[Reshape]invalid dims of output shape[%s]" % (str(outshape))
inferred_axis = -1
copy_axes = []
constant_count = 1
for i in range(num_new_axes):
top_dim = params.shape.dim[i]
if top_dim == 0:
copy_axes.append(i)
copy_axis_index = start_axis + i
outshape[copy_axis_index] = inshape[copy_axis_index]
elif top_dim == -1:
assert inferred_axis == -1, "[Reshape]new shape contains multiple -1 dims"
inferred_axis = i
else:
constant_count *= top_dim
if inferred_axis >= 0:
explicit_count = constant_count
l = inshape[0:start_axis]
if len(l) > 0:
explicit_count *= count(l)
l = inshape[end_axis:]
if len(l) > 0:
explicit_count *= count(l)
for i in range(len(copy_axes)):
explicit_count *= outshape[start_axis + copy_axes[i]]
assert input_count % explicit_count == 0, "[Reshape]botom count[%d] "\
"must be divisible by product of the specified dimensions[%d] "\
% (input_count, explicit_count)
outshape[start_axis + inferred_axis] = int(input_count / explicit_count)
output_count = count(outshape)
assert output_count == input_count, "[Reshape]output count[%d] must match input count[%d]" % (
output_count, input_count)
outshape[0] = -1
return [outshape]
def shape_argmax(layer, input_shape):
inshape = input_shape[0]
params = layer.argmax_param
out_max_val = params.out_max_val if hasattr(params, out_max_val) else False
top_k = params.top_k if hasattr(params, top_k) else 1
axis = parmas.axis if hasattr(params, axis) else -1
if axis < 0:
axis += len(inshape)
assert (axis + 1 == len(inshape)
), 'only can be applied on the last dimension[axis:%d, %s] now,'\
'make sure you have set axis param in xxx.prototxt file' \
% (axis, str(inshape))
outshape = inshape
outshape[-1] = top_k
if out_max_val is True:
outshape[-1] *= 2
return [outshape]
def shape_crop(layer, input_shape):
assert len(input_shape) == 2, "the number of crop's inputs must be 2"
return [input_shape[1]]
def shape_flatten(layer, input_shape):
assert len(input_shape) == 1, "the number of flatten's inputs must be 1"
inshape = input_shape[0]
params = layer.flatten_param
start_axis = params.axis
end_axis = params.end_axis
if start_axis < 0:
start_axis += len(inshape)
if end_axis < 0:
end_axis += len(inshape) + 1
assert start_axis <= end_axis, 'invalid axis[%d] or end_axis[%d] params'\
% (start_axis, end_axis)
output_shape = inshape[0:start_axis]
if len(inshape[start_axis:end_axis]) != 0:
flat_sz = reduce(lambda a, b: a * b, inshape[start_axis:end_axis])
output_shape += [flat_sz]
output_shape += inshape[end_axis:len(inshape)]
output_shape[0] = -1
return [output_shape]
def shape_power(layer, input_shape):
return input_shape
def shape_reduction(layer, input_shape):
params = layer.reduction_param
axis = params.axis
if axis < 0:
axis += len(input_shape[0]) + 1
assert axis <= len(input_shape[0]), 'invalid axis[%d] error' % (axis)
return [input_shape[0:axis]]
x2paddle/x2paddle/op_mapper/onnx2paddle/onnx_op_mapper.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from x2paddle.op_mapper.onnx2paddle.opset9 import OpSet9, custom_layers
from x2paddle.core.op_mapper import OpMapper
from x2paddle.decoder.onnx_decoder import ONNXGraph, ONNXGraphNode, ONNXGraphDataNode
class ONNXOpMapper(OpMapper):
def __init__(self, decoder):
super(ONNXOpMapper, self).__init__()
self.support_op_sets = [9, ]
self.default_op_set = 9
self.graph = decoder.graph
self.opset = self.create_opset(decoder)
if not self.op_checker():
raise Exception("Model are not supported yet.")
#mapping op
print("Total nodes: {}".format(
sum([
isinstance(node, ONNXGraphNode)
for name, node in self.graph.node_map.items()
])))
print("Nodes converting ...")
for node_name in self.graph.topo_sort:
node = self.graph.get_node(node_name)
op = node.layer_type
if hasattr(self.opset, op):
func = getattr(self.opset, op)
func(node)
elif op in self.opset.default_op_mapping:
self.opset.directly_map(node)
elif op in custom_layers:
self.opset.deal_custom_layer(node)
elif op in self.opset.elementwise_ops:
self.opset.elementwise_map(node)
print("Nodes converted.")
self.weights = self.opset.weights
self.omit_nodes = self.opset.omit_nodes
self.used_custom_layers = self.opset.used_custom_layers
def op_checker(self):
unsupported_ops = set()
for node_name in self.graph.topo_sort:
node = self.graph.get_node(node_name)
op = node.layer_type
if not hasattr(self.opset, op) and \
op not in self.opset.default_op_mapping and \
op not in custom_layers and \
op not in self.opset.elementwise_ops:
unsupported_ops.add(op)
if len(unsupported_ops) == 0:
return True
else:
print("There are {} ops not supported yet, list as below".format(
len(unsupported_ops)))
for op in unsupported_ops:
print(op)
return False
def create_opset(self, decoder):
run_op_set = self.default_op_set
opset = ''
if decoder.op_set in self.support_op_sets:
opset = 'OpSet' + str(decoder.op_set)
elif decoder.op_set < self.default_op_set:
opset = 'OpSet' + str(self.default_op_set)
else:
for op_set in self.support_op_sets:
if decoder.op_set > op_set:
run_op_set = op_set
else:
break
opset = 'OpSet' + str(run_op_set)
print(
'Now, onnx2paddle support convert onnx model opset_verison {},'
'opset_verison of your onnx model is {}, automatically treated as op_set: {}.'
.format(self.support_op_sets, decoder.op_set, run_op_set))
return eval(opset)(decoder)
x2paddle/x2paddle/op_mapper/onnx2paddle/opset9/__init__.py 0 → 100644
浏览文件 @ cf7f9b88
from .opset import OpSet9
from .custom_layer import custom_layers
x2paddle/x2paddle/op_mapper/onnx2paddle/opset9/custom_layer/__init__.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .register import get_registered_layers
custom_layers = get_registered_layers()
def set_args(f, params):
""" set args for function 'f' using the parameters in node.layer.param
Args:
f (function): a python function object
params (object): a object contains attributes needed by f's arguments
Returns:
arg_names (list): a list of argument names
kwargs (dict): a dict contains needed arguments
"""
argc = f.__code__.co_argcount
arg_list = f.__code__.co_varnames[0:argc]
kwargs = {}
for arg_name in arg_list:
if hasattr(params, arg_name) and params is not None:
kwargs[arg_name] = getattr(params, arg_name)
return arg_list, kwargs
def has_layer(layer_type):
""" test whether this layer exists in custom layer
"""
return layer_type in custom_layers
def get_params(layer, layer_type):
import re
if layer_type.lower() == "deconvolution" or layer_type.lower(
) == "convolutiondepthwise":
param_name = '_'.join(('convolution', 'param'))
elif layer_type.lower() == "normalize":
param_name = '_'.join(('norm', 'param'))
elif len(layer_type) - len(re.sub("[A-Z]", "", layer_type)) >= 2:
s = ''
tmp_name = ''
for i, ch in enumerate(layer_type):
if i == 0:
s += ch.lower()
continue
elif ch.isupper() and layer_type[i - 1].islower():
tmp_name += (s + '_')
s = ''
s += ch.lower()
tmp_name += s
param_name = '_'.join((tmp_name, 'param'))
else:
param_name = '_'.join((layer_type.lower(), 'param'))
return getattr(layer, param_name, None)
def compute_output_shape(node):
""" compute the output shape of custom layer
"""
layer_type = node.layer_type
assert layer_type in custom_layers, "layer[%s] not exist in custom layers" % (
layer_type)
shape_func = custom_layers[layer_type]['shape']
layer = node.layer
params = get_params(layer, layer_type)
arg_names, kwargs = set_args(shape_func, params)
input_shape = node.input_shape
return shape_func(input_shape, **kwargs)
def make_custom_layer(node):
""" get the code which implement the custom layer function
"""
layer_type = node.layer_type
assert layer_type in custom_layers, "layer[%s] not exist in custom layers" % (
layer_type)
layer_func = custom_layers[layer_type]['layer']
import inspect
return inspect.getsource(layer_func), layer_func
def make_custom_child_func(node):
""" get the code which implement the custom layer function
"""
layer_type = node.layer_type
child_func = custom_layers[layer_type]['child_func']
if child_func is None:
return None, child_func
import inspect
return inspect.getsource(child_func), child_func
def deal_weights(node, data=None):
""" deal the weights of the custom layer
"""
layer_type = node.layer_type
weights_func = custom_layers[layer_type]['weights']
name = node.layer_name
return weights_func(name, data)
x2paddle/x2paddle/op_mapper/onnx2paddle/opset9/custom_layer/register.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" this module provides 'register' for registering customized layers
"""
g_custom_layers = {}
def register(kind, shape, layer, child_func, weights):
""" register a custom layer or a list of custom layers
Args:
@kind (str or list): type name of the layer
@shape (function): a function to generate the shape of layer's output
@layer (function): a function to generate the paddle code of layer
@weights (function): a function to deal with weights data
Returns:
None
"""
assert type(shape).__name__ == 'function', 'shape should be a function'
assert type(layer).__name__ == 'function', 'layer should be a function'
if type(kind) is str:
kind = [kind]
else:
assert type(
kind) is list, 'invalid param "kind" for register, not a list or str'
for k in kind:
assert type(
k) is str, 'invalid param "kind" for register, not a list of str'
assert k not in g_custom_layers, 'this type[%s] has already been registered' % (
k)
g_custom_layers[k] = {
'shape': shape,
'layer': layer,
'child_func': child_func,
'weights': weights
}
def get_registered_layers():
return g_custom_layers
x2paddle/x2paddle/op_mapper/onnx2paddle/opset9/opset.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from x2paddle.decoder.onnx_decoder import ONNXGraph, ONNXGraphNode, ONNXGraphDataNode
from x2paddle.core.graph import GraphNode
from x2paddle.core.fluid_code import Layer
from x2paddle.core.fluid_code import FluidCode
from x2paddle.core.util import string
from x2paddle.op_mapper.onnx2paddle.opset9.custom_layer import *
from functools import reduce
import numpy as np
import onnx
import onnx.numpy_helper as numpy_helper
from onnx.mapping import TENSOR_TYPE_TO_NP_TYPE
import logging as _logging
from collections import OrderedDict
import math
import os
import shutil
_logger = _logging.getLogger(__name__)
def _const_weight_or_none(node, necessary=False):
if 'Constant' in node.layer_type:
return node.value
if isinstance(node, ONNXGraphDataNode):
return node.weight
if necessary:
assert '{} should be an initializer or Constant operator.'.format(
node.layer_name)
return None
def _is_static_shape(shape):
negtive_dims = 0
error_dims = 0
for dim in shape:
if dim < 0:
negtive_dims += 1
if dim < -1:
error_dims += 1
if negtive_dims > 1:
return False
if error_dims > 0:
return False
return True
def _get_same_padding(in_size, kernel_size, stride):
new_size = int(math.ceil(in_size * 1.0 / stride))
pad_size = (new_size - 1) * stride + kernel_size - in_size
pad0 = int(pad_size / 2)
pad1 = pad_size - pad0
return [pad0, pad1]
def print_mapping_info(func):
def run_mapping(*args, **kwargs):
node = args[1]
try:
res = func(*args, **kwargs)
except:
print("convert failed node:{}, op_type is {}".format(
node.layer_name[9:], node.layer_type))
raise
else:
#print("convert successfully node:{}, op_type is {}".format(
# node.layer_name[9:], node.layer_type))
return res
return run_mapping
class OpSet9():
elementwise_ops = {
'Add': 'elementwise_add',
'Div': 'elementwise_div',
'Sub': 'elementwise_sub',
'Mul': 'elementwise_mul',
'Pow': 'elementwise_pow',
}
default_op_mapping_field_values = OrderedDict()
default_op_mapping_field_values['FLUID_OP'] = ''
default_op_mapping_field_values['FLUID_INPUT_ARGS'] = None
default_op_mapping_field_values['FLUID_OUTPUT_ARGS'] = None
default_op_mapping_field_values['ATTR_MAPPING'] = dict()
default_op_mapping_field_values['DEFAULTS'] = dict()
default_op_mapping_field_values['INPUT_PERM'] = None
default_op_mapping_field_values['OUTPUT_PERM'] = None
default_op_mapping_field_values['FILL_NAME_FIELD'] = True
default_op_mapping = {
'Shape': ['shape', ['X'], ['Out']],
'Clip': [
'clip', ['X'], ['Out'], dict(), dict(
min=(np.asarray(
[255, 255, 127, 255], dtype=np.uint8).view(np.float32)[0]),
max=(np.asarray(
[255, 255, 127, 127], dtype=np.uint8).view(np.float32)[0]),
)
],
'Erf': ['erf', ['X'], ['Out']],
'Ceil': ['ceil', ['X'], ['Out']],
'ReduceMean': [
'reduce_mean', ['X'], ['Out'], dict(
axes='dim', keepdims='keep_dim'), dict(keep_dim=1)
],
'ReduceSum': [
'reduce_sum', ['X'], ['Out'], dict(
axes='dim', keepdims='keep_dim'), dict(keep_dim=1)
],
'ReduceMin': [
'reduce_min', ['X'], ['Out'], dict(
axes='dim', keepdims='keep_dim'), dict(keep_dim=1)
],
'ReduceMax': [
'reduce_max', ['X'], ['Out'], dict(
axes='dim', keepdims='keep_dim'), dict(keep_dim=1)
],
#active function
'Relu': ['relu', ['X'], ['Out']],
'LeakyRelu': ['leaky_relu', ['X'], ['Out'], dict(), dict(alpha=.01)],
'Elu': ['elu', ['X'], ['Out'], dict(), dict(alpha=1.)],
'ThresholdedRelu': [
'thresholded_relu', ['X'], ['Out'], dict(alpha='threshold'),
dict(alpha=1.)
],
'Tanh': ['tanh', ['X'], ['Out']],
'Sigmoid': ['sigmoid', ['X'], ['Out']],
'HardSigmoid': [
'hard_sigmoid', ['X'], ['Out'], dict(
alpha='slope', beta='offset'), dict(
slope=.2, offset=.5)
],
'Softsign': ['softsign', ['X'], ['Out']],
'Softplus': ['softplus', ['X'], ['Out']],
'Exp': ['exp', ['X'], ['Out']],
'Softmax': ['softmax', ['X'], ['Out'], dict(), dict(axis=1)],
'Sqrt': ['sqrt', ['X'], ['Out']],
'Floor': ['floor', ['X'], ['Out']],
'Abs': ['abs', ['X'], ['Out']],
}
default_ioa_constraint = {}
def __init__(self, decoder):
super(OpSet9, self).__init__()
self.graph = decoder.graph
self.input_shapes = []
self.weights = dict()
self.omit_nodes = list()
self.used_custom_layers = dict()
@print_mapping_info
def directly_map(self, node, name='', *args, **kwargs):
inputs = node.layer.input
outputs = node.layer.output
op_type = node.layer_type
attrs = node.attr_map
info = self.default_op_mapping[op_type]
info.extend(
list(self.default_op_mapping_field_values.values())[len(info):])
(
fluid_op,
fluid_input_args,
fluid_output_args,
attr_mapping,
default_attrs,
input_perm,
output_perm,
fill_name_field, ) = info
if fluid_op in self.default_ioa_constraint:
for predicate, message in self.default_ioa_constraint[fluid_op]:
assert predicate(inputs, outputs, attrs), message
mapped_attrs = {
attr_mapping.get(key, key): value
for key, value in attrs.items()
}
if '' in mapped_attrs:
mapped_attrs.pop('')
if '_' in mapped_attrs:
mapped_attrs.pop('_')
fluid_attrs = default_attrs.copy()
fluid_attrs.update(mapped_attrs)
inputs = inputs if input_perm is None else list(
map(lambda i: inputs[i], input_perm))
val_inps = []
for idx, ipt in enumerate(inputs):
val_inps.append(self.graph.get_input_node(node, idx=idx, copy=True))
val_outs = outputs if output_perm is None else list(
map(lambda i: outputs[i], output_perm))
attr = fluid_attrs
assert len(val_inps) == 1, 'directly_map error with multi inputs'
if fluid_op not in ['shape', 'erf']:
attr['name'] = string(node.layer_name)
node.fluid_code.add_layer(
fluid_op, inputs=val_inps[0], output=val_outs[0], param_attr=attr)
if fluid_op in ['shape']:
node.fluid_code.add_layer(
'cast',
inputs=val_outs[0],
output=val_outs[0],
param_attr={'dtype': string('int64')})
@print_mapping_info
def deal_custom_layer(self, node):
op = node.layer_type
custom_code, func = make_custom_layer(node)
child_func_code, child_func = make_custom_child_func(node)
params = get_params(node.layer, node.layer_type)
arg_names, kwargs = set_args(func, params)
kwargs['name'] = string(node.layer_name)
node.fluid_code.add_layer(
func.__code__.co_name,
inputs=node.inputs,
output=node,
param_attr=kwargs,
is_custom_layer=True)
if op not in self.used_custom_layers:
self.used_custom_layers[op] = custom_code
if op + '_child_func' not in self.used_custom_layers:
if child_func_code is not None:
self.used_custom_layers[op +
'_child_func'] = child_func_code
@print_mapping_info
def elementwise_map(self, node):
assert node.layer_type in self.elementwise_ops
op_type = self.elementwise_ops[node.layer_type]
val_x = self.graph.get_input_node(node, idx=0, copy=True)
val_y = self.graph.get_input_node(node, idx=1, copy=True)
inputs = {'x': val_x, 'y': val_y}
node.fluid_code.add_layer(
op_type, inputs=inputs, output=node, param_attr=None)
@print_mapping_info
def place_holder(self, node):
self.input_shapes.append(node.out_shapes[0])
shape = node.out_shapes[0]
for i, dim_shape in enumerate(shape):
if dim_shape == 0 and i == 0:
shape[i] = 1
if dim_shape == 0 and i != 0:
assert 'shape of input is not assigned'
attr = {
"dtype": string(node.dtype),
"shape": shape,
"name": string(node.layer_name),
"append_batch_size": 'False'
}
node.fluid_code.add_layer(
"data", inputs=None, output=node, param_attr=attr)
@print_mapping_info
def create_parameter(self, node, parameter=None):
if parameter is not None:
node = parameter
dtype = node.dtype
shape = node.out_shapes[0]
if len(node.weight.shape) == 0:
shape = [1]
self.weights[node.layer_name] = node.weight
attr = {
'dtype': string(dtype),
'shape': shape,
'name': string(node.layer_name),
'default_initializer': 'Constant(0.0)'
}
if dtype == 'bool':
attr['dtype'] = string('int64')
node.fluid_code.add_layer(
"create_parameter", inputs=None, output=node, param_attr=attr)
node.fluid_code.add_layer(
"cast",
inputs=node,
output=node,
param_attr={'dtype': string('bool')})
elif dtype == 'uint8':
attr['dtype'] = string('float32')
node.fluid_code.add_layer(
"create_parameter", inputs=None, output=node, param_attr=attr)
else:
node.fluid_code.add_layer(
"create_parameter", inputs=None, output=node, param_attr=attr)
def _pad_if_asymmetric(self, node, pads, val_name): # pads: SSEE
assert len(pads) & 1 == 0
symmetric = True
ndims = len(pads) // 2
for idx_dim in range(ndims):
if pads[idx_dim] != pads[ndims + idx_dim]:
symmetric = False
break
if symmetric:
return pads[:ndims], val_name
val_padded = self.Pad(node, op_independent=False)
return [0] * ndims, val_padded
def _interpolate(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
inputs = {'input': val_x}
if node.layer_type == 'Resize':
if len(node.layer.input) == 2:
# opset 10
val_scales = self.graph.get_input_node(node, idx=1, copy=True)
inputs['scale'] = val_scales
elif len(node.layer.input) == 3:
# opset 11
val_scales = self.graph.get_input_node(node, idx=2, copy=True)
inputs['scale'] = val_scales
elif len(node.layer.input) == 4:
# opset 11
val_sizes = self.graph.get_input_node(node, idx=3, copy=True)
var_nc, var_hw = val_sizes.layer_name + '_nc', val_sizes.layer_name + '_hw'
node.fluid_code.add_layer(
'split',
inputs=val_sizes,
output=var_nc + ',' + var_hw,
param_attr={
'dim': 0,
'num_or_sections': [2, 2],
})
node.fluid_code.add_layer(
"cast",
inputs=var_hw,
output=var_hw,
param_attr={'dtype': string('int32')})
inputs['out_shape'] = var_hw
elif node.layer_type == 'Upsample':
val_scales = self.graph.get_input_node(node, idx=1, copy=True)
inputs['scale'] = val_scales
attr = {'name': string(node.layer_name)}
mode = node.get_attr('mode', 'nearest')
fluid_op = 'resize_{}'.format(mode)
if 'linear' in mode:
print(
'Warnning: paddle not support op:resize wiht mode: linear, we use bilinear replace linear'
)
fluid_op = 'resize_bilinear'
node.fluid_code.add_layer(
fluid_op, inputs=inputs, output=node, param_attr=attr)
@print_mapping_info
def RoiAlign(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
val_rois = self.graph.get_input_node(node, idx=1, copy=True)
pooled_height = node.get_attr('output_height')
pooled_width = node.get_attr('output_width')
spatial_scale = node.get_attr('spatial_scale')
sampling_ratio = node.get_attr('sampling_ratio')
attr = {
'pooled_height': pooled_height,
'pooled_width': pooled_width,
'spatial_scale': spatial_scale,
'sampling_ratio': sampling_ratio,
}
node.fluid_code.add_layer(
'roi_align',
inputs={'input': val_x,
'rois': val_rois},
output=node,
param_attr=attr)
@print_mapping_info
def MaxRoiPool(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
val_rois = self.graph.get_input_node(node, idx=1, copy=True)
spatial_scale = node.get_attr('spatial_scale')
pooled_height, pooled_width = node.get_attr('pooled_shape')
attr = {
'pooled_height': pooled_height,
'pooled_width': pooled_width,
'spatial_scale': spatial_scale,
}
node.fluid_code.add_layer(
'roi_pool',
inputs={'input': val_x,
'rois': val_rois},
output=node,
param_attr=attr)
@print_mapping_info
def Pad(self, node, op_independent=True):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
pads = node.get_attr('pads')
mode = node.get_attr('mode', 'constant')
value = node.get_attr('value', 0.)
data_shape = val_x.out_shapes[0]
output_shape = node.out_shapes[0]
assume_pad2d = False
attr = {}
paddings = []
if len(pads) == 4:
assume_pad2d |= mode != 'constant'
if data_shape:
assume_pad2d |= data_shape and len(data_shape) == 4 # NCHW
if output_shape:
assume_pad2d |= output_shape and len(output_shape) == 4 # NCHW
if assume_pad2d:
fluid_op = 'pad2d'
attr['data_format'] = string('NCHW')
attr['mode'] = string(mode)
else:
attr = {'pad_value': value}
fluid_op = 'pad'
if len(pads) == 4:
paddings = np.array(pads).reshape(
(-1, 2)).transpose().flatten().tolist() # SSEE -> SESE
elif len(pads) == 8:
paddings = np.array(pads).reshape(
(-1, 4)).transpose().flatten().tolist() # SSEE -> SESE
if sum(paddings[:4]) == 0:
fluid_op = 'pad2d'
paddings = paddings[4:]
attr['mode'] = string(mode)
attr['paddings'] = paddings
if op_independent:
attr['name'] = string(node.layer_name)
node.fluid_code.add_layer(
fluid_op, inputs=val_x, output=node, param_attr=attr)
else:
attr['name'] = string(node.layer_name + '_paded')
node.fluid_code.add_layer(
fluid_op,
inputs=val_x,
output=node.layer_name + '_paded',
param_attr=attr)
return node.layer_name + '_paded'
@print_mapping_info
def Unsqueeze(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
axes = node.get_attr('axes')
attr = {'axes': axes, 'name': string(node.layer_name)}
if len(val_x.out_shapes[0]) == 0:
if node.layer_name:
node.fluid_code.add_layer(
'reshape',
inputs=val_x,
output=node,
param_attr={'shape': [1]})
else:
if str(val_x.dtype) == 'bool':
val_x_cast = val_x.layer_name + '_cast'
node.fluid_code.add_layer(
'cast',
inputs=val_x,
output=val_x_cast,
param_attr={'dtype': string('int64')})
node.fluid_code.add_layer(
'unsqueeze',
inputs=val_x_cast,
output=node,
param_attr=attr)
else:
node.fluid_code.add_layer(
'unsqueeze', inputs=val_x, output=node, param_attr=attr)
@print_mapping_info
def Shrink(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
bias = node.get_attr('bias')
lambd = node.get_attr('lambd')
assert bias == 0.0, 'not support bias!=0'
attr = {'threshold': lambd, 'name': node.layer_name}
node.fluid_code.add_layer(
'hard_shrink', inputs=val_x, output=node, param_attr=attr)
@print_mapping_info
def Constant(self, node):
val_output = self.graph.get_node(node.layer.output[0], copy=True)
value = node.get_attr('value')
dtype = np.dtype(value.dtype)
output_dtype = val_output.dtype
if output_dtype:
assert dtype == output_dtype, 'tensor dtype unmatches storage dtype'
shape = node.get_attr('shape', None)
if shape is None:
shape = val_output.out_shapes[0]
if shape is None:
shape = list(value.shape)
_logger.warning('in (Constant -> %s): '
'attribute "shape" of %s not inferred, '
'using value as 1-D tensor may lead to fails',
val_output.layer_name, val_output.layer_name)
if len(value) == 1:
value = value.tolist()
shape = [1]
value = value[0]
if dtype.name == 'int64':
dtype = 'int32'
attr = {'shape': shape, 'dtype': string(dtype), 'value': value}
node.fluid_code.add_layer(
'fill_constant', inputs=None, output=node, param_attr=attr)
else:
if dtype.name == 'uint8':
dtype = 'int64'
value = np.reshape(value, shape)
self.weights[node.layer_name] = value
attr = {
'dtype': string(dtype),
'shape': shape,
'name': string(node.layer_name),
'default_initializer': 'Constant(0.0)'
}
node.fluid_code.add_layer(
"create_parameter", inputs=None, output=node, param_attr=attr)
@print_mapping_info
def Resize(self, node):
self._interpolate(node)
@print_mapping_info
def Upsample(self, node):
self._interpolate(node)
@print_mapping_info
def InstanceNormalization(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
val_scale = self.graph.get_input_node(node, idx=1, copy=True)
val_b = self.graph.get_input_node(node, idx=2, copy=True)
epsilon = node.get_attr('epsilon', 1e-5)
attr = {
'epsilon': epsilon,
'param_attr': string(val_scale.layer_name),
'bias_attr': string(val_b.layer_name)
}
node.fluid_code.add_layer(
"instance_norm", inputs=val_x, output=node, param_attr=attr)
@print_mapping_info
def Expand(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
val_shape = self.graph.get_input_node(node, idx=1, copy=True)
if len(val_shape.outputs) == 1:
self.omit_nodes.append(val_shape.layer_name)
val_x_dtype = val_x.dtype
name_ones = node.layer_name + '_ones'
attr_ones = {
'shape': val_shape.layer_name,
'dtype': string(val_x_dtype),
'value': 1
}
node.fluid_code.add_layer(
'fill_constant',
inputs=None,
output=name_ones,
param_attr=attr_ones)
inputs = {'x': name_ones, 'y': val_x}
node.fluid_code.add_layer(
'elementwise_mul',
inputs=inputs,
output=node.layer_name,
param_attr=None)
@print_mapping_info
def Gather(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
indices = self.graph.get_input_node(node, idx=1, copy=True)
indices_shape = indices.out_shapes[0]
axis = node.get_attr('axis', 0)
#assert len(
# indices_shape) <= 2, "Gather op don't support dim of indice >2 "
if axis == 0 and len(indices_shape) <= 1:
if len(val_x.out_shapes[0]) <= 1:
node.fluid_code.add_layer(
'gather',
inputs={'input': val_x,
'index': indices},
output=node,
param_attr=None)
elif len(val_x.out_shapes[0]) > 1:
if len(indices_shape) == 0:
gather_ = node.layer_name + '_1'
node.fluid_code.add_layer(
'gather',
inputs={'input': val_x,
'index': indices},
output=gather_,
param_attr=None)
node.fluid_code.add_layer(
'squeeze',
inputs={'input': gather_,
'axes': [0]},
output=node,
param_attr=None)
else:
node.fluid_code.add_layer(
'gather',
inputs={'input': val_x,
'index': indices},
output=node,
param_attr=None)
elif axis > 0 and len(indices_shape) <= 1:
perm = list(range(len(val_x.out_shapes[0])))
perm = [axis] + perm[:axis] + perm[axis + 1:]
attr_trans = {'perm': perm}
name_trans = val_x.layer_name + '_trans'
node.fluid_code.add_layer(
'transpose',
inputs=val_x,
output=name_trans,
param_attr=attr_trans)
node.fluid_code.add_layer(
'gather',
inputs={'input': name_trans,
'index': indices},
output=node,
param_attr=None)
node.fluid_code.add_layer(
'transpose', inputs=node, output=node, param_attr=attr_trans)
if len(indices_shape) < 1:
node.fluid_code.add_layer(
'squeeze',
inputs={'input': node,
'axes': [axis]},
output=node,
param_attr=None)
elif axis == 0 and len(indices_shape) > 1:
if val_x.out_shapes[0] is not None and isinstance(
val_x, ONNXGraphDataNode):
indices_cast = indices.layer_name + '_cast'
node.fluid_code.add_layer(
'cast',
inputs=indices,
output=indices_cast,
param_attr={'dtype': string('int64')})
node.fluid_code.add_layer(
'embedding',
inputs=indices_cast,
output=node,
use_fluid=True,
param_attr={
'param_attr': string(val_x.layer_name),
'size': val_x.out_shapes[0]
})
else:
from functools import reduce
reshape_shape = reduce(lambda x, y: x * y, indices_shape)
indices_reshape = indices.layer_name + '_shape'
node.fluid_code.add_layer(
'reshape',
inputs=indices,
output=indices_reshape,
param_attr={'shape': [reshape_shape, ]})
perm = list(range(len(val_x.out_shapes[0])))
node.fluid_code.add_layer(
'gather',
inputs={'input': val_x,
'index': indices_reshape},
output=node,
param_attr=None)
val_x_shape = val_x.out_shapes[0]
reshaped_shape = []
for i in perm:
reshaped_shape.append(indices_shape[i])
for i in val_x_shape[:axis] + val_x_shape[axis + 1:]:
reshaped_shape.append(i)
node.fluid_code.add_layer(
'reshape',
inputs=node,
output=node,
param_attr={'shape': reshaped_shape})
elif axis > 0 and len(indices_shape) > 1:
from functools import reduce
reshape_shape = reduce(lambda x, y: x * y, indices_shape)
indices_reshape = indices.layer_name + '_shape'
node.fluid_code.add_layer(
'reshape',
inputs=indices,
output=indices_reshape,
param_attr={'shape': [reshape_shape, ]})
perm = list(range(len(val_x.out_shapes[0])))
perm = [axis] + perm[:axis] + perm[axis + 1:]
attr_trans = {'perm': perm}
name_trans = val_x.layer_name + '_transpose'
node.fluid_code.add_layer(
'transpose',
inputs=val_x,
output=name_trans,
param_attr=attr_trans)
node.fluid_code.add_layer(
'gather',
inputs={'input': name_trans,
'index': indices_reshape},
output=node,
param_attr=None)
input_transpose = node.layer_name + '_transpose'
node.fluid_code.add_layer(
'transpose',
inputs=node,
output=input_transpose,
param_attr=attr_trans)
val_x_shape = val_x.out_shapes[0]
reshaped_shape = []
for i in perm:
reshaped_shape.append(indices_shape[i])
for i in val_x_shape[:axis] + val_x_shape[axis + 1:]:
reshaped_shape.append(i)
node.fluid_code.add_layer(
'reshape',
inputs=input_transpose,
output=node,
param_attr={'shape': reshaped_shape})
@print_mapping_info
def ScatterND(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
indices = self.graph.get_input_node(node, idx=1, copy=True)
updates = self.graph.get_input_node(node, idx=2, copy=True)
if len(indices.out_shapes[0]) == 1:
node.fluid_code.add_layer(
'scatter',
inputs={'input': val_x,
'index': indices,
'updates': updates},
output=node,
param_attr=None)
else:
input_inner_indices = node.layer_name + '_input_inner_indices'
node.fluid_code.add_layer(
'scatter_nd',
inputs={
'shape': val_x.out_shapes[0],
'index': indices,
'updates': updates
},
output=input_inner_indices,
param_attr=None)
constant_minus_one = node.layer_name + '_constant_minus_one'
node.fluid_code.add_layer(
'fill_constant',
inputs=None,
output=constant_minus_one,
param_attr={
'shape': updates.out_shapes[0],
'dtype': string(updates.dtype),
'value': -1
})
indices_mask = node.layer_name + '_indices_mask'
node.fluid_code.add_layer(
'scatter_nd',
inputs={
'shape': val_x.out_shapes[0],
'index': indices,
'updates': constant_minus_one
},
output=indices_mask,
param_attr=None)
constant_1 = node.layer_name + '_constant_1'
node.fluid_code.add_layer(
'fill_constant',
inputs=None,
output=constant_1,
param_attr={
'shape': val_x.out_shapes[0],
'dtype': string(val_x.dtype),
'value': 1
})
input_out_indices_mask = node.layer_name + '_input_out_indices_mask'
node.fluid_code.add_layer(
"elementwise_add",
inputs={"x": indices_mask,
"y": constant_1},
output=input_out_indices_mask,
param_attr=None)
input_out_indices = node.layer_name + '_input_out_indices'
node.fluid_code.add_layer(
"elementwise_mul",
inputs={"x": val_x,
"y": input_out_indices_mask},
output=input_out_indices,
param_attr=None)
node.fluid_code.add_layer(
"elementwise_add",
inputs={"x": input_inner_indices,
"y": input_out_indices},
output=node,
param_attr=None)
@print_mapping_info
def Range(self, node):
val_start = self.graph.get_input_node(node, idx=0, copy=True)
val_limit = self.graph.get_input_node(node, idx=1, copy=True)
val_delta = self.graph.get_input_node(node, idx=2, copy=True)
dtype = val_start.dtype
inputs = {'start': val_start, 'end': val_limit, 'step': val_delta}
node.fluid_code.add_layer(
'range',
inputs=inputs,
output=node,
param_attr={'dtype': string(dtype)})
@print_mapping_info
def Slice(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
starts, ends, axes, steps = None, None, None, None
attr = {}
if len(node.inputs) > 1:
starts = self.graph.get_input_node(node, idx=1, copy=True)
ends = self.graph.get_input_node(node, idx=2, copy=True)
if len(node.inputs) > 3:
axes = self.graph.get_input_node(node, idx=3, copy=True)
axes = _const_weight_or_none(axes, necessary=True)
if len(node.inputs) > 4:
steps = self.graph.get_input_node(node, idx=4, copy=True)
steps = _const_weight_or_none(steps)
if steps is not None:
assert steps == 1, "Only support convert op:Slice, which attribute:steps == 1"
attr = {
"axes": axes,
"starts": starts.layer_name,
"ends": ends.layer_name
}
starts_value = _const_weight_or_none(starts)
ends_value = _const_weight_or_none(ends)
if starts_value is not None and ends_value is not None:
self.omit_nodes.append(starts.layer_name)
self.omit_nodes.append(ends.layer_name)
ends_value = ends_value.copy()
for idx in range(len(ends_value)):
if ends_value[idx] > 2**31 - 1:
ends_value[idx] = 2**31 - 1
attr = {
"axes": axes,
"starts": starts_value,
"ends": ends_value
}
else:
if starts.dtype != 'int32':
starts_cast = starts.layer_name + '_cast'
node.fluid_code.add_layer(
'cast',
inputs=starts,
output=starts_cast,
param_attr={'dtype': string('int32')})
attr['starts'] = starts_cast
if ends.dtype != 'int32':
ends_cast = ends.layer_name + '_cast'
node.fluid_code.add_layer(
'cast',
inputs=ends,
output=ends_cast,
param_attr={'dtype': string('int32')})
attr['ends'] = ends_cast
else:
starts = node.get_attr('starts')
ends = node.get_attr('ends')
axes = node.get_attr('axes')
for idx in range(len(ends)):
if ends[idx] > 2**31 - 1:
ends[idx] = 2**31 - 1
attr = {"axes": axes, "starts": starts, "ends": ends}
node.fluid_code.add_layer(
'slice', inputs=val_x, output=node, param_attr=attr)
@print_mapping_info
def ConstantOfShape(self, node):
val_shape = self.graph.get_input_node(node, idx=0, copy=True)
val_y = self.graph.get_node(node.layer.output[0], copy=True)
value = node.get_attr('value')
dtype = value.dtype
value = value.tolist()
assert len(value) == 1, ('given value not Scalar, shape of value > 1, '
'this is not supported')
if len(value) == 1:
value = value[0]
attr = {
'shape': val_shape.layer_name,
'dtype': string(dtype),
'value': value
}
node.fluid_code.add_layer(
'fill_constant', inputs=None, output=node, param_attr=attr)
@print_mapping_info
def Split(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
val_y = self.graph.get_node(node.layer.output[0], copy=True)
fluid_op = 'split'
split = node.get_attr('split')
axis = node.get_attr('axis', 0)
attr = {
'num_or_sections': split,
'dim': axis,
'name': string(node.layer_name)
}
node.fluid_code.add_layer(
'split', inputs=val_x, output=val_y, param_attr=attr)
@print_mapping_info
def Reshape(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
val_shape = self.graph.get_input_node(node, idx=1, copy=True)
val_reshaped = self.graph.get_node(node.layer.output[0], copy=True)
attr = {}
shape_value = _const_weight_or_none(val_shape)
shape_dims = len(val_shape.out_shapes[0])
if shape_value is not None:
node.fluid_code.add_layer(
'reshape',
inputs={'x': val_x},
output=node,
param_attr={'shape': shape_value.tolist()})
elif len(node.out_shapes[0]) > 0 and _is_static_shape(node.out_shapes[
0]):
node.fluid_code.add_layer(
'reshape',
inputs={'x': val_x,
'shape': node.out_shapes[0]},
output=node,
param_attr=attr)
elif val_shape.dtype == 'int64':
val_shape_cast = val_shape.layer_name + '_cast'
node.fluid_code.add_layer(
'cast',
inputs=val_shape,
output=val_shape_cast,
param_attr={'dtype': string('int32')})
# shape may be [], come form Gather by scalar indices
if len(val_shape.out_shapes[0]) > 0:
node.fluid_code.add_layer(
'reshape',
inputs=val_shape_cast,
output=val_shape_cast,
param_attr={'shape': val_shape.out_shapes[0]})
node.fluid_code.add_layer(
'reshape',
inputs={'x': val_x,
'shape': val_shape_cast},
output=node,
param_attr=attr)
else:
# shape may be [], come form Gather by scalar indices
if len(val_shape.out_shapes[0]) > 0:
node.fluid_code.add_layer(
'reshape',
inputs=val_shape,
output=val_shape,
param_attr={'shape': val_shape.out_shapes[0]})
node.fluid_code.add_layer(
'reshape',
inputs={'x': val_x,
'shape': val_shape},
output=node,
param_attr=attr)
@print_mapping_info
def Cast(self, node):
val_input = self.graph.get_input_node(node, idx=0, copy=True)
val_output = self.graph.get_node(node.layer.output[0], copy=True)
dtype = node.get_attr('to')
if not isinstance(dtype, np.dtype):
dtype = TENSOR_TYPE_TO_NP_TYPE[dtype]
output_dtype = val_output.dtype
if output_dtype:
assert dtype == output_dtype, 'dtype of to unmatches output'
attr = {'dtype': string(dtype)}
node.fluid_code.add_layer(
'cast', inputs=val_input, output=node, param_attr=attr)
@print_mapping_info
def Not(self, node):
val_input = self.graph.get_input_node(node, idx=0, copy=True)
node.fluid_code.add_layer('logical_not', inputs=val_input, output=node)
@print_mapping_info
def AveragePool(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
auto_pad = node.get_attr('auto_pad', 'NOTSET')
kernel_shape = node.get_attr("kernel_shape")
poolnd = len(kernel_shape)
strides = node.get_attr("strides")
pad_mode = node.get_attr("pads")
ceil_mode = bool(node.get_attr('ceil_mode', 0))
pads = node.get_attr('pads', [0] * (poolnd * 2))
fluid_op = 'pool{}d'.format(poolnd)
assert 2 <= poolnd <= 3, 'only pool2d and pool3d is supported'
paddings, val_x = self._pad_if_asymmetric(node, pads, val_x)
if auto_pad == "SAME_UPPER" or auto_pad == "SAME_LOWER":
input_shape = val_x.out_shapes[0]
pad_h = _get_same_padding(input_shape[2], kernel_shape[0],
strides[0])
pad_w = _get_same_padding(input_shape[3], kernel_shape[1],
strides[1])
paddings = pad_h + pad_w
attr = {
"pool_size": kernel_shape,
"pool_type": string('avg'),
"pool_stride": strides,
"pool_padding": paddings,
"ceil_mode": ceil_mode,
"exclusive": 'True',
"name": string(node.layer_name)
}
node.fluid_code.add_layer(
fluid_op, inputs=val_x, output=node, param_attr=attr)
@print_mapping_info
def Concat(self, node):
inputs = []
dtypes = set()
for i in range(len(node.layer.input)):
ipt = self.graph.get_input_node(node, idx=i, copy=True)
if isinstance(ipt, str):
inputs.append(ipt)
else:
inputs.append(ipt.layer_name)
dtypes.add(ipt.dtype)
if len(dtypes) > 1:
assert 'Unspported situation happened, please create issue on https://github.com/PaddlePaddle/X2Paddle/issues.'
axis = node.get_attr('axis')
attr = {'axis': axis}
node.fluid_code.add_layer(
'concat', inputs=inputs, output=node, param_attr=attr)
@print_mapping_info
def Flatten(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
axis = node.get_attr('axis', 1)
attr = {"axis": str(axis), "name": string(node.layer_name)}
node.fluid_code.add_layer(
'flatten', inputs=val_x, output=node, param_attr=attr)
@print_mapping_info
def Gemm(self, node):
val_a = self.graph.get_input_node(node, idx=0, copy=True)
val_b = self.graph.get_input_node(node, idx=1, copy=True)
val_c = self.graph.get_input_node(node, idx=2, copy=True)
alpha = node.get_attr('alpha', 1.) # optional
beta = node.get_attr('beta', 1.) # optional
trans_a = bool(node.get_attr('transA', 0)) # optional
trans_b = bool(node.get_attr('transB', 0)) # optional
val_mm = node.layer_name + '_mm'
matmul_inputs = {"x": val_a, "y": val_b}
attr_matmul = {
"transpose_x": trans_a,
"transpose_y": trans_b,
"alpha": alpha,
"name": string(val_mm)
}
node.fluid_code.add_layer(
'matmul',
inputs=matmul_inputs,
output=val_mm,
param_attr=attr_matmul)
if beta != 0:
if beta == 1.:
add_inputs = {"x": val_mm, "y": val_c}
attr = {"name": string(node.layer_name)}
node.fluid_code.add_layer(
"elementwise_add",
inputs=add_inputs,
output=node,
param_attr=attr)
else:
var_beta = node.layer_name + '_beta'
matmul_beta_inputs = {"x": val_c, "y": var_beta}
node.fluid_code.add_layer(
"Constant",
inputs=matmul_beta_inputs,
output=var_beta,
param_attr={'value': beta})
add_inputs = {"x": val_mm, "y": var_beta}
attr = {"name": string(node.layer_name)}
node.fluid_code.add_layer(
"elementwise_add",
inputs=add_inputs,
output=node,
param_attr=attr)
@print_mapping_info
def Sum(self, node):
val_inps = node.layer.input
inputs = {
"x": self.graph.get_input_node(
node, idx=0, copy=True),
"y": self.graph.get_input_node(
node, idx=1, copy=True),
}
node.fluid_code.add_layer("elementwise_add", inputs=inputs, output=node)
for idx, ipt in enumerate(val_inps[2:]):
y = self.graph.get_input_node(node, idx=idx, copy=True)
inputs = {
"x": node.layer_name,
"y": y,
}
node.fluid_code.add_layer(
"elementwise_add", inputs=inputs, output=node)
@print_mapping_info
def MatMul(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
val_y = self.graph.get_input_node(node, idx=1, copy=True)
x_shape = val_x.out_shapes[0]
y_shape = val_y.out_shapes[0]
inputs = {"x": val_x, "y": val_y}
if y_shape[0] == 1 and x_shape[-1] != 1 and x_shape[0] != 1:
y_squeeze = val_y.layer_name + '_squeeze'
node.fluid_code.add_layer(
"squeeze",
inputs=val_y,
output=y_squeeze,
param_attr={'axes': [0]})
inputs['y'] = y_squeeze
node.fluid_code.add_layer(
"matmul", inputs=inputs, output=node, param_attr=None)
else:
node.fluid_code.add_layer(
"matmul", inputs=inputs, output=node, param_attr=None)
@print_mapping_info
def BatchNormalization(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
val_scale = self.graph.get_input_node(node, idx=1, copy=True)
val_b = self.graph.get_input_node(node, idx=2, copy=True)
val_mean = self.graph.get_input_node(node, idx=3, copy=True)
val_var = self.graph.get_input_node(node, idx=4, copy=True)
self.omit_nodes.append(val_scale.layer_name)
self.omit_nodes.append(val_b.layer_name)
self.omit_nodes.append(val_mean.layer_name)
self.omit_nodes.append(val_var.layer_name)
momentum = node.get_attr('momentum', .9)
epsilon = node.get_attr('epsilon', 1e-5)
# Attribute: spatial is used in BatchNormalization-1,6,7
spatial = bool(node.get_attr('spatial'))
attr = {
"momentum": momentum,
"epsilon": epsilon,
"data_layout": string('NCHW'),
"is_test": True,
"param_attr": string(val_scale.layer_name),
"bias_attr": string(val_b.layer_name),
"moving_mean_name": string(val_mean.layer_name),
"moving_variance_name": string(val_var.layer_name),
"use_global_stats": spatial,
"name": string(node.layer_name)
}
node.fluid_code.add_layer(
"batch_norm", inputs=val_x, output=node, param_attr=attr)
@print_mapping_info
def Transpose(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
perm = node.get_attr('perm')
attr = {'perm': perm, "name": string(node.layer_name)}
node.fluid_code.add_layer(
"transpose", inputs=val_x, output=node, param_attr=attr)
@print_mapping_info
def Relu(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
attr = {"name": string(node.layer_name)}
node.fluid_code.add_layer(
"relu", inputs=val_x, output=node, param_attr=attr)
@print_mapping_info
def PRelu(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
val_slope = self.graph.get_input_node(node, idx=1, copy=True)
mode = 'channel'
shape_slope = val_slope.out_shapes[0]
if len(shape_slope) == 1:
mode = 'all'
elif len(shape_slope) > 2:
mode = 'element'
attr = {
"param_attr": string(val_slope.layer_name),
'mode': string(mode)
}
node.fluid_code.add_layer(
"prelu", inputs=val_x, output=node, param_attr=attr)
@print_mapping_info
def Squeeze(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
axes = node.get_attr('axes')
attr = {'axes': axes, "name": string(node.layer_name)}
if len(val_x.out_shapes[0]) == 1:
node.fluid_code.add_layer(
"cast",
inputs=val_x,
output=node,
param_attr={'dtype': string(val_x.dtype)})
else:
node.fluid_code.add_layer(
"squeeze", inputs=val_x, output=node, param_attr=attr)
@print_mapping_info
def Equal(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
val_y = self.graph.get_input_node(node, idx=1, copy=True)
node.fluid_code.add_layer(
"equal",
inputs={'x': val_x,
'y': val_y},
output=node,
param_attr=None)
@print_mapping_info
def Greater(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
val_y = self.graph.get_input_node(node, idx=1, copy=True)
node.fluid_code.add_layer(
"greater_than",
inputs={'x': val_x,
'y': val_y},
output=node,
param_attr=None)
@print_mapping_info
def Where(self, node):
condition = self.graph.get_input_node(node, idx=0, copy=True)
val_x = self.graph.get_input_node(node, idx=1, copy=True)
val_y = self.graph.get_input_node(node, idx=2, copy=True)
not_condition = condition.layer_name + '_not'
node.fluid_code.add_layer(
"logical_not",
inputs=condition,
output=not_condition,
param_attr=None)
cast_not_condition = not_condition + '_cast'
node.fluid_code.add_layer(
"cast",
inputs=not_condition,
output=cast_not_condition,
param_attr={'dtype': string(val_x.dtype)})
cast_condition = condition.layer_name + '_cast'
node.fluid_code.add_layer(
"cast",
inputs=condition,
output=cast_condition,
param_attr={'dtype': string(val_x.dtype)})
mul_val_x = val_x.layer_name + '_mul'
node.fluid_code.add_layer(
"elementwise_mul",
inputs={'x': val_x,
'y': cast_condition},
output=mul_val_x,
param_attr=None)
mul_val_y = val_y.layer_name + '_mul'
node.fluid_code.add_layer(
"elementwise_mul",
inputs={'x': val_y,
'y': cast_not_condition},
output=mul_val_y,
param_attr=None)
node.fluid_code.add_layer(
"elementwise_add",
inputs={'x': mul_val_x,
'y': mul_val_y},
output=node,
param_attr=None)
@print_mapping_info
def NonZero(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
val_x_dim = len(val_x.out_shapes[0])
if val_x_dim == 1:
node.fluid_code.add_layer("nonzero", inputs=val_x, output=val_x)
node.fluid_code.add_layer(
"transpose",
inputs=val_x,
output=node,
param_attr={'perm': [1, 0]})
if val_x_dim > 1:
node.fluid_code.add_layer("nonzero", inputs=val_x, output=val_x)
node.fluid_code.add_layer(
"split",
inputs=val_x,
output=val_x,
param_attr={'num_or_sections': 1,
'dim': val_x_dim})
node.fluid_code.add_layer("concat", inputs=val_x, output=node)
@print_mapping_info
def Identity(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
node.fluid_code.add_layer("assign", inputs=val_x, output=node)
@print_mapping_info
def Tile(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
val_repeats = self.graph.get_input_node(node, idx=1, copy=True)
repeats = _const_weight_or_none(val_repeats)
if repeats is None:
repeats = val_repeats.layer_name
if val_repeats.dtype != 'int32':
attr = {"dtype": string("int32")}
node.fluid_code.add_layer(
"cast",
inputs=repeats,
output="{}.tmp".format(repeats),
param_attr=attr)
repeats = "{}.tmp".format(repeats)
elif isinstance(repeats, int):
repeats = [repeats]
attr = {
'expand_times': repeats,
"name": string(node.layer_name),
}
node.fluid_code.add_layer(
"expand", inputs=val_x, output=node, param_attr=attr)
@print_mapping_info
def MaxPool(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
auto_pad = node.get_attr('auto_pad', 'NOTSET')
assert node.get_attr(
"dilations") is None, 'only dilations = 0 is supported' # optional
kernel_shape = node.get_attr("kernel_shape")
poolnd = len(kernel_shape)
strides = node.get_attr("strides")
pad_mode = node.get_attr("pads")
ceil_mode = bool(node.get_attr('ceil_mode', 0)) # optional
pads = node.get_attr('pads', [0] * (poolnd * 2)) # optional
fluid_op = 'pool{}d'.format(poolnd)
assert 2 <= poolnd <= 3, 'only pool2d and pool3d is supported'
paddings, val_x = self._pad_if_asymmetric(node, pads, val_x)
if auto_pad == "SAME_UPPER" or auto_pad == "SAME_LOWER":
input_shape = val_x.out_shapes[0]
pad_h = _get_same_padding(input_shape[2], kernel_shape[0],
strides[0])
pad_w = _get_same_padding(input_shape[3], kernel_shape[1],
strides[1])
paddings = pad_h + pad_w
attr = {
"pool_size": kernel_shape,
"pool_type": string("max"),
"pool_stride": strides,
"pool_padding": paddings,
"ceil_mode": ceil_mode,
"name": string(node.layer_name),
"exclusive": False
}
node.fluid_code.add_layer(
fluid_op, inputs=val_x, output=node, param_attr=attr)
def _global_pool(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
val_y = self.graph.get_node(node.layer.output[0], copy=True)
fluid_op = 'pool2d'
pool_type = None
if node.layer.op_type == 'GlobalMaxPool':
pool_type = 'max'
elif node.layer.op_type == 'GlobalAveragePool':
pool_type = 'avg'
attr = {
"pool_type": string(pool_type),
"global_pooling": True,
"name": string(node.layer_name)
}
node.fluid_code.add_layer(
fluid_op, inputs=val_x, output=node, param_attr=attr)
@print_mapping_info
def GlobalMaxPool(self, node):
self._global_pool(node)
@print_mapping_info
def GlobalAveragePool(self, node):
self._global_pool(node)
@print_mapping_info
def Conv(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
val_w = self.graph.get_input_node(node, idx=1, copy=True)
val_y = self.graph.get_node(node.layer.output[0], copy=True)
self.omit_nodes.append(val_w.layer_name)
has_bias = len(node.layer.input) == 3
if has_bias:
val_b = self.graph.get_input_node(node, idx=2, copy=True)
self.omit_nodes.append(val_b.layer_name)
auto_pad = node.get_attr('auto_pad', 'NOTSET')
kernel_shape = node.get_attr('kernel_shape')
convnd = len(kernel_shape)
assert 2 <= convnd <= 3, 'only conv2d and conv3d is supported'
num_out_channels = val_w.out_shapes[0][0]
fluid_op = 'conv{}d'.format(convnd)
num_groups = node.get_attr('group', 1)
strides = node.get_attr('strides', [1] * convnd)
dilations = node.get_attr('dilations', [1] * convnd)
pads = node.get_attr('pads', [0] * (convnd * 2))
input_shape = val_x.out_shapes[0]
paddings, val_x = self._pad_if_asymmetric(node, pads, val_x)
if auto_pad == "SAME_UPPER" or auto_pad == "SAME_LOWER":
pad_h = _get_same_padding(input_shape[2], kernel_shape[0],
strides[0])
pad_w = _get_same_padding(input_shape[3], kernel_shape[1],
strides[1])
paddings = pad_h + pad_w
attr = {
"num_filters": num_out_channels,
"filter_size": kernel_shape,
"stride": strides,
"padding": paddings,
"dilation": dilations,
"groups": num_groups,
'param_attr': string(val_w.layer_name),
"name": string(node.layer_name)
}
if has_bias:
attr["bias_attr"] = string(val_b.layer_name)
else:
attr["bias_attr"] = False
node.fluid_code.add_layer(
fluid_op, inputs=val_x, output=node, param_attr=attr)
@print_mapping_info
def ConvTranspose(self, node):
val_x = self.graph.get_input_node(node, idx=0, copy=True)
val_w = self.graph.get_input_node(node, idx=1, copy=True)
val_b = None
if len(node.layer.input) > 2:
val_b = self.graph.get_input_node(node, idx=2, copy=True)
self.omit_nodes.append(val_b.layer_name)
self.omit_nodes.append(val_w.layer_name)
val_y = self.graph.get_node(node.layer.output[0], copy=True)
auto_pad = node.get_attr('auto_pad', 'NOTSET')
out_padding = node.get_attr('output_padding', [0, 0])
kernel_shape = node.get_attr('kernel_shape')
assert kernel_shape, 'kernel_shape not inferred'
convnd = len(kernel_shape)
assert 2 <= convnd <= 3, 'only conv2d_transpose and conv3d_transpose supported'
num_out_channels = val_w.out_shapes[0][1]
fluid_op = 'conv{}d_transpose'.format(convnd)
num_groups = node.get_attr('group', 1)
strides = node.get_attr('strides', [1] * convnd)
dilations = node.get_attr('dilations', [1] * convnd)
output_size = node.get_attr('output_shape', [])
pads = node.get_attr('pads', [0] * (convnd * 2))
paddings, var_x = self._pad_if_asymmetric(node, pads, val_x)
output_size = [0, 0]
output_size[0] = (val_x.out_shapes[0][2] - 1
) * strides[0] - 2 * paddings[0] + dilations[0] * (
kernel_shape[0] - 1) + 1 + out_padding[0]
output_size[1] = (val_x.out_shapes[0][3] - 1
) * strides[1] - 2 * paddings[1] + dilations[1] * (
kernel_shape[1] - 1) + 1 + out_padding[1]
attr = {
'num_filters': num_out_channels,
'output_size': output_size or None,
'filter_size': kernel_shape,
'padding': paddings,
'stride': strides,
'dilation': dilations,
'groups': num_groups,
'param_attr': string(val_w.layer_name),
'bias_attr': None if val_b is None else string(val_b.layer_name),
'name': string(node.layer_name),
}
node.fluid_code.add_layer(
fluid_op, inputs=val_x, output=node, param_attr=attr)
x2paddle/x2paddle/op_mapper/paddle2onnx/opset10/opset.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
import sys
import x2paddle
import os
import numpy as np
import paddle.fluid.core as core
import paddle.fluid as fluid
import onnx
from onnx import helper, onnx_pb
from x2paddle.op_mapper.paddle2onnx.opset9.opset import OpSet9
class OpSet10(OpSet9):
def __init__(self):
super(OpSet10, self).__init__()
def slice(self, op, block):
axes = op.attr('axes')
starts = op.attr('starts')
ends = op.attr('ends')
axes_name = self.get_name(op.type, 'axes')
starts_name = self.get_name(op.type, 'starts')
ends_name = self.get_name(op.type, 'ends')
axes_node = self.make_constant_node(axes_name,
onnx_pb.TensorProto.INT64, axes)
starts_node = self.make_constant_node(starts_name,
onnx_pb.TensorProto.INT64, starts)
ends_node = self.make_constant_node(ends_name,
onnx_pb.TensorProto.INT64, ends)
node = helper.make_node(
"Slice",
inputs=[op.input('Input')[0], starts_name, ends_name, axes_name],
outputs=op.output('Out'), )
return [starts_node, ends_node, axes_node, node]
x2paddle/x2paddle/op_mapper/paddle2onnx/opset11/opset.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
import sys
import x2paddle
import os
import numpy as np
import paddle.fluid.core as core
import paddle.fluid as fluid
import onnx
from onnx import helper, onnx_pb
from x2paddle.op_mapper.paddle2onnx.opset10.opset import OpSet10
class OpSet11(OpSet10):
def __init__(self):
super(OpSet11, self).__init__()
def relu6(self, op, block):
min_name = self.get_name(op.type, 'min')
max_name = self.get_name(op.type, 'max')
min_node = self.make_constant_node(min_name, onnx_pb.TensorProto.FLOAT,
0)
max_node = self.make_constant_node(max_name, onnx_pb.TensorProto.FLOAT,
op.attr('threshold'))
node = helper.make_node(
'Clip',
inputs=[op.input('X')[0], min_name, max_name],
outputs=op.output('Out'), )
return [min_node, max_node, node]
def pad2d(self, op, block):
x_shape = block.var(op.input('X')[0]).shape
paddings = op.attr('paddings')
onnx_pads = []
#TODO support pads is Variable
if op.attr('data_format') == 'NCHW':
pads = [
0, 0, paddings[0], paddings[2], 0, 0, paddings[1], paddings[3]
]
else:
pads = [
0, paddings[0], paddings[2], 0, 0, paddings[1], paddings[3], 0
]
pads_name = self.get_name(op.type, 'pads')
pads_node = self.make_constant_node(pads_name,
onnx_pb.TensorProto.INT64, pads)
constant_value_name = self.get_name(op.type, 'constant_value')
constant_value_node = self.make_constant_node(constant_value_name,
onnx_pb.TensorProto.FLOAT,
op.attr('pad_value'))
node = helper.make_node(
'Pad',
inputs=op.input('X') + [pads_name, constant_value_name],
outputs=op.output('Out'),
mode=op.attr('mode'))
return [pads_node, constant_value_node, node]
def clip(self, op, block):
min_name = self.get_name(op.type, 'min')
max_name = self.get_name(op.type, 'max')
min_node = self.make_constant_node(min_name, onnx_pb.TensorProto.FLOAT,
op.attr('min'))
max_node = self.make_constant_node(max_name, onnx_pb.TensorProto.FLOAT,
op.attr('max'))
node = helper.make_node(
'Clip',
inputs=[op.input('X')[0], min_name, max_name],
outputs=op.output('Out'))
return [min_node, max_node, node]
def bilinear_interp(self, op, block):
input_names = op.input_names
coordinate_transformation_mode = ''
align_corners = op.attr('align_corners')
align_mode = op.attr('align_mode')
if align_corners:
coordinate_transformation_mode = 'align_corners'
elif align_mode == 1:
coordinate_transformation_mode = 'asymmetric'
else:
coordinate_transformation_mode = 'half_pixel'
roi_name = self.get_name(op.type, 'roi')
roi_node = self.make_constant_node(roi_name, onnx_pb.TensorProto.FLOAT,
[1, 1, 1, 1, 1, 1, 1, 1])
if ('OutSize' in input_names and len(op.input('OutSize')) > 0) or (
'SizeTensor' in input_names and
len(op.input('SizeTensor')) > 0):
node_list = list()
empty_name = self.get_name(op.type, 'empty')
empty_tensor = helper.make_tensor(
empty_name,
onnx_pb.TensorProto.FLOAT, (0, ),
np.array([]).astype('float32'),
raw=False)
empty_node = helper.make_node(
'Constant', [], outputs=[empty_name], value=empty_tensor)
shape_name0 = self.get_name(op.type, 'shape')
shape_node0 = helper.make_node(
'Shape', inputs=op.input('X'), outputs=[shape_name0])
starts_name = self.get_name(op.type, 'slice.starts')
starts_node = self.make_constant_node(
starts_name, onnx_pb.TensorProto.INT64, [0])
ends_name = self.get_name(op.type, 'slice.ends')
ends_node = self.make_constant_node(ends_name,
onnx_pb.TensorProto.INT64, [2])
shape_name1 = self.get_name(op.type, 'shape')
shape_node1 = helper.make_node(
'Slice',
inputs=[shape_name0, starts_name, ends_name],
outputs=[shape_name1])
node_list.extend([
roi_node, empty_node, shape_node0, starts_node, ends_node,
shape_node1
])
if 'OutSize' in input_names and len(op.input('OutSize')) > 0:
cast_shape_name = self.get_name(op.type, "shape.cast")
cast_shape_node = helper.make_node(
'Cast',
inputs=op.input('OutSize'),
outputs=[cast_shape_name],
to=onnx_pb.TensorProto.INT64)
node_list.append(cast_shape_node)
else:
concat_shape_name = self.get_name(op.type, "shape.concat")
concat_shape_node = helper.make_node(
"Concat",
inputs=op.input('SizeTensor'),
outputs=[concat_shape_name],
axis=0)
cast_shape_name = self.get_name(op.type, "shape.cast")
cast_shape_node = helper.make_node(
'Cast',
inputs=[concat_shape_name],
outputs=[cast_shape_name],
to=onnx_pb.TensorProto.INT64)
node_list.extend([concat_shape_node, cast_shape_node])
shape_name3 = self.get_name(op.type, "shape.concat")
shape_node3 = helper.make_node(
'Concat',
inputs=[shape_name1, cast_shape_name],
outputs=[shape_name3],
axis=0)
result_node = helper.make_node(
'Resize',
inputs=[op.input('X')[0], roi_name, empty_name, shape_name3],
outputs=op.output('Out'),
mode='linear',
coordinate_transformation_mode=coordinate_transformation_mode)
node_list.extend([shape_node3, result_node])
return node_list
elif 'Scale' in input_names and len(op.input('Scale')) > 0:
node = helper.make_node(
'Resize',
inputs=[op.input('X')[0], roi_name, op.input('Scale')[0]],
outputs=op.output('Out'),
mode='linear',
coordinate_transformation_mode=coordinate_transformation_mode)
else:
out_shape = [op.attr('out_h'), op.attr('out_w')]
scale = op.attr('scale')
if out_shape.count(-1) > 0:
scale_name = self.get_name(op.type, 'scale')
scale_node = self.make_constant_node(scale_name,
onnx_pb.TensorProto.FLOAT,
[1, 1, scale, scale])
node = helper.make_node(
'Resize',
inputs=[op.input('X')[0], roi_name, scale_name],
outputs=op.output('Out'),
mode='nearest',
coordinate_transformation_mode=coordinate_transformation_mode
)
return [scale_node, roi_node, node]
else:
raise Exception("Unexpected situation happend")
return [roi_node, node]
def nearest_interp(self, op, block):
input_names = op.input_names
coordinate_transformation_mode = ''
align_corners = op.attr('align_corners')
if align_corners:
coordinate_transformation_mode = 'align_corners'
else:
coordinate_transformation_mode = 'half_pixel'
roi_name = self.get_name(op.type, 'roi')
roi_node = self.make_constant_node(roi_name, onnx_pb.TensorProto.FLOAT,
[1, 1, 1, 1, 1, 1, 1, 1])
if 'OutSize' in input_names and len(op.input('OutSize')) > 0:
node = helper.make_node(
'Resize',
inputs=[op.input('X')[0], roi_name, op.input('OutSize')[0]],
outputs=op.output('Out'),
mode='nearest',
coordinate_transformation_mode=coordinate_transformation_mode)
elif 'Scale' in input_names and len(op.input('Scale')) > 0:
node = helper.make_node(
'Resize',
inputs=[op.input('X')[0], roi_name, op.input('Scale')[0]],
outputs=op.output('Out'),
mode='nearest',
coordinate_transformation_mode=coordinate_transformation_mode)
else:
out_shape = [op.attr('out_h'), op.attr('out_w')]
scale = op.attr('scale')
if out_shape.count(-1) > 0:
scale_name = self.get_name(op.type, 'scale')
scale_node = self.make_constant_node(scale_name,
onnx_pb.TensorProto.FLOAT,
[1, 1, scale, scale])
node = helper.make_node(
'Resize',
inputs=[op.input('X')[0], roi_name, scale_name],
outputs=op.output('Out'),
mode='nearest',
coordinate_transformation_mode=coordinate_transformation_mode
)
return [scale_node, roi_node, node]
else:
raise Exception("Unexpected situation happend")
return [roi_node, node]
def hard_swish(self, op, block):
min_name = self.get_name(op.type, 'min')
max_name = self.get_name(op.type, 'max')
scale_name = self.get_name(op.type, 'scale')
offset_name = self.get_name(op.type, 'offset')
min_node = self.make_constant_node(min_name, onnx_pb.TensorProto.FLOAT,
0)
max_node = self.make_constant_node(max_name, onnx_pb.TensorProto.FLOAT,
op.attr('threshold'))
scale_node = self.make_constant_node(scale_name,
onnx_pb.TensorProto.FLOAT,
op.attr('scale'))
offset_node = self.make_constant_node(offset_name,
onnx_pb.TensorProto.FLOAT,
op.attr('offset'))
name0 = self.get_name(op.type, 'add')
node0 = helper.make_node(
'Add', inputs=[op.input('X')[0], offset_name], outputs=[name0])
name1 = self.get_name(op.type, 'relu')
node1 = helper.make_node(
'Clip',
inputs=[name0, min_name, max_name],
outputs=[name1], )
name2 = self.get_name(op.type, 'mul')
node2 = helper.make_node(
'Mul', inputs=[op.input('X')[0], name1], outputs=[name2])
node3 = helper.make_node(
'Div', inputs=[name2, scale_name], outputs=op.output('Out'))
return [
min_node, max_node, scale_node, offset_node, node0, node1, node2,
node3
]
def yolo_box(self, op, block):
from .paddle_custom_layer.yolo_box import yolo_box
return yolo_box(op, block)
def multiclass_nms(self, op, block):
from .paddle_custom_layer.multiclass_nms import multiclass_nms
return multiclass_nms(op, block)
x2paddle/x2paddle/op_mapper/paddle2onnx/opset11/paddle_custom_layer/multiclass_nms.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
import sys
import os
import numpy as np
import paddle.fluid.core as core
import paddle.fluid as fluid
import onnx
import logging
from onnx import helper, onnx_pb
def multiclass_nms(op, block):
"""
Convert the paddle multiclass_nms to onnx op.
This op is get the select boxes from origin boxes.
"""
inputs = dict()
outputs = dict()
attrs = dict()
for name in op.input_names:
inputs[name] = op.input(name)
for name in op.output_names:
outputs[name] = op.output(name)
for name in op.attr_names:
attrs[name] = op.attr(name)
result_name = outputs['Out'][0]
background = attrs['background_label']
normalized = attrs['normalized']
if normalized == False:
logging.warn(
"The parameter normalized of multiclass_nms OP of Paddle is False, which has diff with ONNX." \
" Please set normalized=True in multiclass_nms of Paddle, see doc Q4 in https://github.com/PaddlePaddle/X2Paddle/blob/develop/FAQ.md")
#convert the paddle attribute to onnx tensor
name_score_threshold = [outputs['Out'][0] + "@score_threshold"]
name_iou_threshold = [outputs['Out'][0] + "@iou_threshold"]
name_keep_top_k = [outputs['Out'][0] + '@keep_top_k']
name_keep_top_k_2D = [outputs['Out'][0] + '@keep_top_k_1D']
node_score_threshold = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_score_threshold,
value=onnx.helper.make_tensor(
name=name_score_threshold[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=(),
vals=[float(attrs['score_threshold'])]))
node_iou_threshold = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_iou_threshold,
value=onnx.helper.make_tensor(
name=name_iou_threshold[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=(),
vals=[float(attrs['nms_threshold'])]))
node_keep_top_k = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_keep_top_k,
value=onnx.helper.make_tensor(
name=name_keep_top_k[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=(),
vals=[np.int64(attrs['keep_top_k'])]))
node_keep_top_k_2D = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_keep_top_k_2D,
value=onnx.helper.make_tensor(
name=name_keep_top_k_2D[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=[1, 1],
vals=[np.int64(attrs['keep_top_k'])]))
# the paddle data format is x1,y1,x2,y2
kwargs = {'center_point_box': 0}
name_select_nms = [outputs['Out'][0] + "@select_index"]
node_select_nms= onnx.helper.make_node(
'NonMaxSuppression',
inputs=inputs['BBoxes'] + inputs['Scores'] + name_keep_top_k +\
name_iou_threshold + name_score_threshold,
outputs=name_select_nms)
# step 1 nodes select the nms class
node_list = [
node_score_threshold, node_iou_threshold, node_keep_top_k,
node_keep_top_k_2D, node_select_nms
]
# create some const value to use
name_const_value = [result_name+"@const_0",
result_name+"@const_1",\
result_name+"@const_2",\
result_name+"@const_-1"]
value_const_value = [0, 1, 2, -1]
for name, value in zip(name_const_value, value_const_value):
node = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=[name],
value=onnx.helper.make_tensor(
name=name + "@const",
data_type=onnx.TensorProto.INT64,
dims=[1],
vals=[value]))
node_list.append(node)
# In this code block, we will deocde the raw score data, reshape N * C * M to 1 * N*C*M
# and the same time, decode the select indices to 1 * D, gather the select_indices
outputs_gather_1 = [result_name + "@gather_1"]
node_gather_1 = onnx.helper.make_node(
'Gather',
inputs=name_select_nms + [result_name + "@const_1"],
outputs=outputs_gather_1,
axis=1)
node_list.append(node_gather_1)
outputs_squeeze_gather_1 = [result_name + "@sequeeze_gather_1"]
node_squeeze_gather_1 = onnx.helper.make_node(
'Squeeze',
inputs=outputs_gather_1,
outputs=outputs_squeeze_gather_1,
axes=[1])
node_list.append(node_squeeze_gather_1)
outputs_gather_2 = [result_name + "@gather_2"]
node_gather_2 = onnx.helper.make_node(
'Gather',
inputs=name_select_nms + [result_name + "@const_2"],
outputs=outputs_gather_2,
axis=1)
node_list.append(node_gather_2)
#slice the class is not 0
if background == 0:
outputs_nonzero = [result_name + "@nonzero"]
node_nonzero = onnx.helper.make_node(
'NonZero', inputs=outputs_squeeze_gather_1, outputs=outputs_nonzero)
node_list.append(node_nonzero)
else:
name_thresh = [result_name + "@thresh"]
node_thresh = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_thresh,
value=onnx.helper.make_tensor(
name=name_thresh[0] + "@const",
data_type=onnx.TensorProto.INT32,
dims=[1],
vals=[-1]))
node_list.append(node_thresh)
outputs_cast = [result_name + "@cast"]
node_cast = onnx.helper.make_node(
'Cast', inputs=outputs_squeeze_gather_1, outputs=outputs_cast, to=6)
node_list.append(node_cast)
outputs_greater = [result_name + "@greater"]
node_greater = onnx.helper.make_node(
'Greater',
inputs=outputs_cast + name_thresh,
outputs=outputs_greater)
node_list.append(node_greater)
outputs_nonzero = [result_name + "@nonzero"]
node_nonzero = onnx.helper.make_node(
'NonZero', inputs=outputs_greater, outputs=outputs_nonzero)
node_list.append(node_nonzero)
outputs_gather_1_nonzero = [result_name + "@gather_1_nonzero"]
node_gather_1_nonzero = onnx.helper.make_node(
'Gather',
inputs=outputs_gather_1 + outputs_nonzero,
outputs=outputs_gather_1_nonzero,
axis=0)
node_list.append(node_gather_1_nonzero)
outputs_gather_2_nonzero = [result_name + "@gather_2_nonzero"]
node_gather_2_nonzero = onnx.helper.make_node(
'Gather',
inputs=outputs_gather_2 + outputs_nonzero,
outputs=outputs_gather_2_nonzero,
axis=0)
node_list.append(node_gather_2_nonzero)
# reshape scores N * C * M to (N*C*M) * 1
outputs_reshape_scores_rank1 = [result_name + "@reshape_scores_rank1"]
node_reshape_scores_rank1 = onnx.helper.make_node(
"Reshape",
inputs=inputs['Scores'] + [result_name + "@const_-1"],
outputs=outputs_reshape_scores_rank1)
node_list.append(node_reshape_scores_rank1)
# get the shape of scores
outputs_shape_scores = [result_name + "@shape_scores"]
node_shape_scores = onnx.helper.make_node(
'Shape', inputs=inputs['Scores'], outputs=outputs_shape_scores)
node_list.append(node_shape_scores)
# gather the index: 2 shape of scores
outputs_gather_scores_dim1 = [result_name + "@gather_scores_dim1"]
node_gather_scores_dim1 = onnx.helper.make_node(
'Gather',
inputs=outputs_shape_scores + [result_name + "@const_2"],
outputs=outputs_gather_scores_dim1,
axis=0)
node_list.append(node_gather_scores_dim1)
# mul class * M
outputs_mul_classnum_boxnum = [result_name + "@mul_classnum_boxnum"]
node_mul_classnum_boxnum = onnx.helper.make_node(
'Mul',
inputs=outputs_gather_1_nonzero + outputs_gather_scores_dim1,
outputs=outputs_mul_classnum_boxnum)
node_list.append(node_mul_classnum_boxnum)
# add class * M * index
outputs_add_class_M_index = [result_name + "@add_class_M_index"]
node_add_class_M_index = onnx.helper.make_node(
'Add',
inputs=outputs_mul_classnum_boxnum + outputs_gather_2_nonzero,
outputs=outputs_add_class_M_index)
node_list.append(node_add_class_M_index)
# Squeeze the indices to 1 dim
outputs_squeeze_select_index = [result_name + "@squeeze_select_index"]
node_squeeze_select_index = onnx.helper.make_node(
'Squeeze',
inputs=outputs_add_class_M_index,
outputs=outputs_squeeze_select_index,
axes=[0, 2])
node_list.append(node_squeeze_select_index)
# gather the data from flatten scores
outputs_gather_select_scores = [result_name + "@gather_select_scores"]
node_gather_select_scores = onnx.helper.make_node('Gather',
inputs=outputs_reshape_scores_rank1 + \
outputs_squeeze_select_index,
outputs=outputs_gather_select_scores,
axis=0)
node_list.append(node_gather_select_scores)
# get nums to input TopK
outputs_shape_select_num = [result_name + "@shape_select_num"]
node_shape_select_num = onnx.helper.make_node(
'Shape',
inputs=outputs_gather_select_scores,
outputs=outputs_shape_select_num)
node_list.append(node_shape_select_num)
outputs_gather_select_num = [result_name + "@gather_select_num"]
node_gather_select_num = onnx.helper.make_node(
'Gather',
inputs=outputs_shape_select_num + [result_name + "@const_0"],
outputs=outputs_gather_select_num,
axis=0)
node_list.append(node_gather_select_num)
outputs_unsqueeze_select_num = [result_name + "@unsqueeze_select_num"]
node_unsqueeze_select_num = onnx.helper.make_node(
'Unsqueeze',
inputs=outputs_gather_select_num,
outputs=outputs_unsqueeze_select_num,
axes=[0])
node_list.append(node_unsqueeze_select_num)
outputs_concat_topK_select_num = [result_name + "@conat_topK_select_num"]
node_conat_topK_select_num = onnx.helper.make_node(
'Concat',
inputs=outputs_unsqueeze_select_num + name_keep_top_k_2D,
outputs=outputs_concat_topK_select_num,
axis=0)
node_list.append(node_conat_topK_select_num)
outputs_cast_concat_topK_select_num = [
result_name + "@concat_topK_select_num"
]
node_outputs_cast_concat_topK_select_num = onnx.helper.make_node(
'Cast',
inputs=outputs_concat_topK_select_num,
outputs=outputs_cast_concat_topK_select_num,
to=6)
node_list.append(node_outputs_cast_concat_topK_select_num)
# get min(topK, num_select)
outputs_compare_topk_num_select = [result_name + "@compare_topk_num_select"]
node_compare_topk_num_select = onnx.helper.make_node(
'ReduceMin',
inputs=outputs_cast_concat_topK_select_num,
outputs=outputs_compare_topk_num_select,
keepdims=0)
node_list.append(node_compare_topk_num_select)
# unsqueeze the indices to 1D tensor
outputs_unsqueeze_topk_select_indices = [
result_name + "@unsqueeze_topk_select_indices"
]
node_unsqueeze_topk_select_indices = onnx.helper.make_node(
'Unsqueeze',
inputs=outputs_compare_topk_num_select,
outputs=outputs_unsqueeze_topk_select_indices,
axes=[0])
node_list.append(node_unsqueeze_topk_select_indices)
# cast the indices to INT64
outputs_cast_topk_indices = [result_name + "@cast_topk_indices"]
node_cast_topk_indices = onnx.helper.make_node(
'Cast',
inputs=outputs_unsqueeze_topk_select_indices,
outputs=outputs_cast_topk_indices,
to=7)
node_list.append(node_cast_topk_indices)
# select topk scores indices
outputs_topk_select_topk_indices = [result_name + "@topk_select_topk_values",\
result_name + "@topk_select_topk_indices"]
node_topk_select_topk_indices = onnx.helper.make_node(
'TopK',
inputs=outputs_gather_select_scores + outputs_cast_topk_indices,
outputs=outputs_topk_select_topk_indices)
node_list.append(node_topk_select_topk_indices)
# gather topk label, scores, boxes
outputs_gather_topk_scores = [result_name + "@gather_topk_scores"]
node_gather_topk_scores = onnx.helper.make_node(
'Gather',
inputs=outputs_gather_select_scores +
[outputs_topk_select_topk_indices[1]],
outputs=outputs_gather_topk_scores,
axis=0)
node_list.append(node_gather_topk_scores)
outputs_gather_topk_class = [result_name + "@gather_topk_class"]
node_gather_topk_class = onnx.helper.make_node(
'Gather',
inputs=outputs_gather_1_nonzero +
[outputs_topk_select_topk_indices[1]],
outputs=outputs_gather_topk_class,
axis=1)
node_list.append(node_gather_topk_class)
# gather the boxes need to gather the boxes id, then get boxes
outputs_gather_topk_boxes_id = [result_name + "@gather_topk_boxes_id"]
node_gather_topk_boxes_id = onnx.helper.make_node(
'Gather',
inputs=outputs_gather_2_nonzero +
[outputs_topk_select_topk_indices[1]],
outputs=outputs_gather_topk_boxes_id,
axis=1)
node_list.append(node_gather_topk_boxes_id)
# squeeze the gather_topk_boxes_id to 1 dim
outputs_squeeze_topk_boxes_id = [result_name + "@squeeze_topk_boxes_id"]
node_squeeze_topk_boxes_id = onnx.helper.make_node(
'Squeeze',
inputs=outputs_gather_topk_boxes_id,
outputs=outputs_squeeze_topk_boxes_id,
axes=[0, 2])
node_list.append(node_squeeze_topk_boxes_id)
outputs_gather_select_boxes = [result_name + "@gather_select_boxes"]
node_gather_select_boxes = onnx.helper.make_node(
'Gather',
inputs=inputs['BBoxes'] + outputs_squeeze_topk_boxes_id,
outputs=outputs_gather_select_boxes,
axis=1)
node_list.append(node_gather_select_boxes)
# concat the final result
# before concat need to cast the class to float
outputs_cast_topk_class = [result_name + "@cast_topk_class"]
node_cast_topk_class = onnx.helper.make_node(
'Cast',
inputs=outputs_gather_topk_class,
outputs=outputs_cast_topk_class,
to=1)
node_list.append(node_cast_topk_class)
outputs_unsqueeze_topk_scores = [result_name + "@unsqueeze_topk_scores"]
node_unsqueeze_topk_scores = onnx.helper.make_node(
'Unsqueeze',
inputs=outputs_gather_topk_scores,
outputs=outputs_unsqueeze_topk_scores,
axes=[0, 2])
node_list.append(node_unsqueeze_topk_scores)
inputs_concat_final_results = outputs_cast_topk_class + outputs_unsqueeze_topk_scores +\
outputs_gather_select_boxes
outputs_sort_by_socre_results = [result_name + "@concat_topk_scores"]
node_sort_by_socre_results = onnx.helper.make_node(
'Concat',
inputs=inputs_concat_final_results,
outputs=outputs_sort_by_socre_results,
axis=2)
node_list.append(node_sort_by_socre_results)
# select topk classes indices
outputs_squeeze_cast_topk_class = [result_name + "@squeeze_cast_topk_class"]
node_squeeze_cast_topk_class = onnx.helper.make_node(
'Squeeze',
inputs=outputs_cast_topk_class,
outputs=outputs_squeeze_cast_topk_class,
axes=[0, 2])
node_list.append(node_squeeze_cast_topk_class)
outputs_topk_select_classes_indices = [result_name + "@topk_select_topk_classes_scores",\
result_name + "@topk_select_topk_classes_indices"]
node_topk_select_topk_indices = onnx.helper.make_node(
'TopK',
inputs=outputs_squeeze_cast_topk_class + outputs_cast_topk_indices,
outputs=outputs_topk_select_classes_indices,
largest=0)
node_list.append(node_topk_select_topk_indices)
outputs_concat_final_results = outputs['Out']
node_concat_final_results = onnx.helper.make_node(
'Gather',
inputs=outputs_sort_by_socre_results +
[outputs_topk_select_classes_indices[1]],
outputs=outputs_concat_final_results,
axis=1)
node_list.append(node_concat_final_results)
return node_list
x2paddle/x2paddle/op_mapper/paddle2onnx/opset11/paddle_custom_layer/yolo_box.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import onnx
import numpy as np
from onnx import onnx_pb, helper
from x2paddle.op_mapper.paddle2onnx.opset9.paddle_custom_layer.yolo_box import is_static_shape
from x2paddle.op_mapper.paddle2onnx.opset9.paddle_custom_layer.yolo_box import get_old_name
from x2paddle.op_mapper.paddle2onnx.opset9.paddle_custom_layer.yolo_box import MAX_FLOAT32
def yolo_box(op, block):
inputs = dict()
outputs = dict()
attrs = dict()
for name in op.input_names:
inputs[name] = op.input(name)
for name in op.output_names:
outputs[name] = op.output(name)
for name in op.attr_names:
attrs[name] = op.attr(name)
model_name = outputs['Boxes'][0]
input_shape = block.vars[get_old_name(inputs['X'][0])].shape
is_static_shape(input_shape)
image_size = inputs['ImgSize']
input_height = input_shape[2]
input_width = input_shape[3]
class_num = attrs['class_num']
anchors = attrs['anchors']
num_anchors = int(len(anchors)) // 2
downsample_ratio = attrs['downsample_ratio']
input_size = input_height * downsample_ratio
conf_thresh = attrs['conf_thresh']
conf_thresh_mat = np.ones([num_anchors * input_height *
input_width]) * conf_thresh
node_list = []
im_outputs = []
x_shape = [1, num_anchors, 5 + class_num, input_height, input_width]
name_x_shape = [model_name + "@x_shape"]
node_x_shape = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_x_shape,
value=onnx.helper.make_tensor(
name=name_x_shape[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=[5],
vals=x_shape))
node_list.append(node_x_shape)
outputs_x_reshape = [model_name + "@reshape"]
node_x_reshape = onnx.helper.make_node(
'Reshape', inputs=inputs['X'] + name_x_shape, outputs=outputs_x_reshape)
node_list.append(node_x_reshape)
outputs_x_transpose = [model_name + "@x_transpose"]
node_x_transpose = onnx.helper.make_node(
'Transpose',
inputs=outputs_x_reshape,
outputs=outputs_x_transpose,
perm=[0, 1, 3, 4, 2])
node_list.append(node_x_transpose)
range_x = []
range_y = []
for i in range(0, input_width):
range_x.append(i)
for j in range(0, input_height):
range_y.append(j)
name_range_x = [model_name + "@range_x"]
node_range_x = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_range_x,
value=onnx.helper.make_tensor(
name=name_range_x[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=[input_width],
vals=range_x))
node_list.append(node_range_x)
name_range_y = [model_name + "@range_y"]
node_range_y = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_range_y,
value=onnx.helper.make_tensor(
name=name_range_y[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=[input_height],
vals=range_y))
node_list.append(node_range_y)
range_x_new_shape = [1, input_width]
range_y_new_shape = [input_height, 1]
name_range_x_new_shape = [model_name + "@range_x_new_shape"]
node_range_x_new_shape = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_range_x_new_shape,
value=onnx.helper.make_tensor(
name=name_range_x_new_shape[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=[len(range_x_new_shape)],
vals=range_x_new_shape))
node_list.append(node_range_x_new_shape)
name_range_y_new_shape = [model_name + "@range_y_new_shape"]
node_range_y_new_shape = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_range_y_new_shape,
value=onnx.helper.make_tensor(
name=name_range_y_new_shape[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=[len(range_y_new_shape)],
vals=range_y_new_shape))
node_list.append(node_range_y_new_shape)
outputs_range_x_reshape = [model_name + "@range_x_reshape"]
node_range_x_reshape = onnx.helper.make_node(
'Reshape',
inputs=name_range_x + name_range_x_new_shape,
outputs=outputs_range_x_reshape)
node_list.append(node_range_x_reshape)
outputs_range_y_reshape = [model_name + "@range_y_reshape"]
node_range_y_reshape = onnx.helper.make_node(
'Reshape',
inputs=name_range_y + name_range_y_new_shape,
outputs=outputs_range_y_reshape)
node_list.append(node_range_y_reshape)
outputs_grid_x = [model_name + "@grid_x"]
node_grid_x = onnx.helper.make_node(
"Tile",
inputs=outputs_range_x_reshape + name_range_y_new_shape,
outputs=outputs_grid_x)
node_list.append(node_grid_x)
outputs_grid_y = [model_name + "@grid_y"]
node_grid_y = onnx.helper.make_node(
"Tile",
inputs=outputs_range_y_reshape + name_range_x_new_shape,
outputs=outputs_grid_y)
node_list.append(node_grid_y)
outputs_box_x = [model_name + "@box_x"]
outputs_box_y = [model_name + "@box_y"]
outputs_box_w = [model_name + "@box_w"]
outputs_box_h = [model_name + "@box_h"]
outputs_conf = [model_name + "@conf"]
outputs_prob = [model_name + "@prob"]
node_split_input = onnx.helper.make_node(
"Split",
inputs=outputs_x_transpose,
outputs=outputs_box_x + outputs_box_y + outputs_box_w\
+ outputs_box_h + outputs_conf + outputs_prob,
axis=-1,
split=[1, 1, 1, 1, 1, class_num])
node_list.append(node_split_input)
outputs_box_x_sigmoid = [model_name + "@box_x_sigmoid"]
outputs_box_y_sigmoid = [model_name + "@box_y_sigmoid"]
node_box_x_sigmoid = onnx.helper.make_node(
"Sigmoid", inputs=outputs_box_x, outputs=outputs_box_x_sigmoid)
node_list.append(node_box_x_sigmoid)
node_box_y_sigmoid = onnx.helper.make_node(
"Sigmoid", inputs=outputs_box_y, outputs=outputs_box_y_sigmoid)
node_list.append(node_box_y_sigmoid)
outputs_box_x_squeeze = [model_name + "@box_x_squeeze"]
outputs_box_y_squeeze = [model_name + "@box_y_squeeze"]
node_box_x_squeeze = onnx.helper.make_node(
'Squeeze',
inputs=outputs_box_x_sigmoid,
outputs=outputs_box_x_squeeze,
axes=[4])
node_list.append(node_box_x_squeeze)
node_box_y_squeeze = onnx.helper.make_node(
'Squeeze',
inputs=outputs_box_y_sigmoid,
outputs=outputs_box_y_squeeze,
axes=[4])
node_list.append(node_box_y_squeeze)
outputs_box_x_add_grid = [model_name + "@box_x_add_grid"]
outputs_box_y_add_grid = [model_name + "@box_y_add_grid"]
node_box_x_add_grid = onnx.helper.make_node(
"Add",
inputs=outputs_grid_x + outputs_box_x_squeeze,
outputs=outputs_box_x_add_grid)
node_list.append(node_box_x_add_grid)
node_box_y_add_grid = onnx.helper.make_node(
"Add",
inputs=outputs_grid_y + outputs_box_y_squeeze,
outputs=outputs_box_y_add_grid)
node_list.append(node_box_y_add_grid)
name_input_h = [model_name + "@input_h"]
name_input_w = [model_name + "@input_w"]
node_input_h = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_input_h,
value=onnx.helper.make_tensor(
name=name_input_w[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=(),
vals=[input_height]))
node_list.append(node_input_h)
node_input_w = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_input_w,
value=onnx.helper.make_tensor(
name=name_input_w[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=(),
vals=[input_width]))
node_list.append(node_input_w)
outputs_box_x_encode = [model_name + "@box_x_encode"]
outputs_box_y_encode = [model_name + "@box_y_encode"]
node_box_x_encode = onnx.helper.make_node(
'Div',
inputs=outputs_box_x_add_grid + name_input_w,
outputs=outputs_box_x_encode)
node_list.append(node_box_x_encode)
node_box_y_encode = onnx.helper.make_node(
'Div',
inputs=outputs_box_y_add_grid + name_input_h,
outputs=outputs_box_y_encode)
node_list.append(node_box_y_encode)
name_anchor_tensor = [model_name + "@anchor_tensor"]
node_anchor_tensor = onnx.helper.make_node(
"Constant",
inputs=[],
outputs=name_anchor_tensor,
value=onnx.helper.make_tensor(
name=name_anchor_tensor[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=[len(anchors)],
vals=anchors))
node_list.append(node_anchor_tensor)
anchor_shape = [int(num_anchors), 2]
name_anchor_shape = [model_name + "@anchor_shape"]
node_anchor_shape = onnx.helper.make_node(
"Constant",
inputs=[],
outputs=name_anchor_shape,
value=onnx.helper.make_tensor(
name=name_anchor_shape[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=[2],
vals=anchor_shape))
node_list.append(node_anchor_shape)
outputs_anchor_tensor_reshape = [model_name + "@anchor_tensor_reshape"]
node_anchor_tensor_reshape = onnx.helper.make_node(
"Reshape",
inputs=name_anchor_tensor + name_anchor_shape,
outputs=outputs_anchor_tensor_reshape)
node_list.append(node_anchor_tensor_reshape)
name_input_size = [model_name + "@input_size"]
node_input_size = onnx.helper.make_node(
"Constant",
inputs=[],
outputs=name_input_size,
value=onnx.helper.make_tensor(
name=name_input_size[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=(),
vals=[input_size]))
node_list.append(node_input_size)
outputs_anchors_div_input_size = [model_name + "@anchors_div_input_size"]
node_anchors_div_input_size = onnx.helper.make_node(
"Div",
inputs=outputs_anchor_tensor_reshape + name_input_size,
outputs=outputs_anchors_div_input_size)
node_list.append(node_anchors_div_input_size)
outputs_anchor_w = [model_name + "@anchor_w"]
outputs_anchor_h = [model_name + "@anchor_h"]
node_anchor_split = onnx.helper.make_node(
'Split',
inputs=outputs_anchors_div_input_size,
outputs=outputs_anchor_w + outputs_anchor_h,
axis=1,
split=[1, 1])
node_list.append(node_anchor_split)
new_anchor_shape = [1, int(num_anchors), 1, 1]
name_new_anchor_shape = [model_name + "@new_anchor_shape"]
node_new_anchor_shape = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_new_anchor_shape,
value=onnx.helper.make_tensor(
name=name_new_anchor_shape[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=[len(new_anchor_shape)],
vals=new_anchor_shape))
node_list.append(node_new_anchor_shape)
outputs_anchor_w_reshape = [model_name + "@anchor_w_reshape"]
outputs_anchor_h_reshape = [model_name + "@anchor_h_reshape"]
node_anchor_w_reshape = onnx.helper.make_node(
'Reshape',
inputs=outputs_anchor_w + name_new_anchor_shape,
outputs=outputs_anchor_w_reshape)
node_list.append(node_anchor_w_reshape)
node_anchor_h_reshape = onnx.helper.make_node(
'Reshape',
inputs=outputs_anchor_h + name_new_anchor_shape,
outputs=outputs_anchor_h_reshape)
node_list.append(node_anchor_h_reshape)
outputs_box_w_squeeze = [model_name + "@box_w_squeeze"]
node_box_w_squeeze = onnx.helper.make_node(
'Squeeze',
inputs=outputs_box_w,
outputs=outputs_box_w_squeeze,
axes=[4])
node_list.append(node_box_w_squeeze)
outputs_box_h_squeeze = [model_name + "@box_h_squeeze"]
node_box_h_squeeze = onnx.helper.make_node(
'Squeeze',
inputs=outputs_box_h,
outputs=outputs_box_h_squeeze,
axes=[4])
node_list.append(node_box_h_squeeze)
outputs_box_w_exp = [model_name + "@box_w_exp"]
node_box_w_exp = onnx.helper.make_node(
"Exp", inputs=outputs_box_w_squeeze, outputs=outputs_box_w_exp)
node_list.append(node_box_w_exp)
outputs_box_h_exp = [model_name + "@box_h_exp"]
node_box_h_exp = onnx.helper.make_node(
"Exp", inputs=outputs_box_h_squeeze, outputs=outputs_box_h_exp)
node_list.append(node_box_h_exp)
outputs_box_w_encode = [model_name + "box_w_encode"]
outputs_box_h_encode = [model_name + "box_h_encode"]
node_box_w_encode = onnx.helper.make_node(
'Mul',
inputs=outputs_box_w_exp + outputs_anchor_w_reshape,
outputs=outputs_box_w_encode)
node_list.append(node_box_w_encode)
node_box_h_encode = onnx.helper.make_node(
'Mul',
inputs=outputs_box_h_exp + outputs_anchor_h_reshape,
outputs=outputs_box_h_encode)
node_list.append(node_box_h_encode)
outputs_conf_sigmoid = [model_name + "@conf_sigmoid"]
node_conf_sigmoid = onnx.helper.make_node(
'Sigmoid', inputs=outputs_conf, outputs=outputs_conf_sigmoid)
node_list.append(node_conf_sigmoid)
name_conf_thresh = [model_name + "@conf_thresh"]
node_conf_thresh = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_conf_thresh,
value=onnx.helper.make_tensor(
name=name_conf_thresh[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=[num_anchors * input_height * input_width],
vals=conf_thresh_mat))
node_list.append(node_conf_thresh)
conf_shape = [1, int(num_anchors), input_height, input_width, 1]
name_conf_shape = [model_name + "@conf_shape"]
node_conf_shape = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_conf_shape,
value=onnx.helper.make_tensor(
name=name_conf_shape[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=[len(conf_shape)],
vals=conf_shape))
node_list.append(node_conf_shape)
outputs_conf_thresh_reshape = [model_name + "@conf_thresh_reshape"]
node_conf_thresh_reshape = onnx.helper.make_node(
'Reshape',
inputs=name_conf_thresh + name_conf_shape,
outputs=outputs_conf_thresh_reshape)
node_list.append(node_conf_thresh_reshape)
outputs_conf_sub = [model_name + "@conf_sub"]
node_conf_sub = onnx.helper.make_node(
'Sub',
inputs=outputs_conf_sigmoid + outputs_conf_thresh_reshape,
outputs=outputs_conf_sub)
node_list.append(node_conf_sub)
outputs_conf_clip = [model_name + "@conf_clip"]
node_conf_clip = onnx.helper.make_node(
'Clip', inputs=outputs_conf_sub, outputs=outputs_conf_clip)
node_list.append(node_conf_clip)
zeros = [0]
name_zeros = [model_name + "@zeros"]
node_zeros = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_zeros,
value=onnx.helper.make_tensor(
name=name_zeros[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=(),
vals=zeros))
node_list.append(node_zeros)
outputs_conf_clip_bool = [model_name + "@conf_clip_bool"]
node_conf_clip_bool = onnx.helper.make_node(
'Greater',
inputs=outputs_conf_clip + name_zeros,
outputs=outputs_conf_clip_bool)
node_list.append(node_conf_clip_bool)
outputs_conf_clip_cast = [model_name + "@conf_clip_cast"]
node_conf_clip_cast = onnx.helper.make_node(
'Cast',
inputs=outputs_conf_clip_bool,
outputs=outputs_conf_clip_cast,
to=1)
node_list.append(node_conf_clip_cast)
outputs_conf_set_zero = [model_name + "@conf_set_zero"]
node_conf_set_zero = onnx.helper.make_node(
'Mul',
inputs=outputs_conf_sigmoid + outputs_conf_clip_cast,
outputs=outputs_conf_set_zero)
node_list.append(node_conf_set_zero)
outputs_prob_sigmoid = [model_name + "@prob_sigmoid"]
node_prob_sigmoid = onnx.helper.make_node(
'Sigmoid', inputs=outputs_prob, outputs=outputs_prob_sigmoid)
node_list.append(node_prob_sigmoid)
new_shape = [1, int(num_anchors), input_height, input_width, 1]
name_new_shape = [model_name + "@new_shape"]
node_new_shape = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_new_shape,
value=onnx.helper.make_tensor(
name=name_new_shape[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=[len(new_shape)],
vals=new_shape))
node_list.append(node_new_shape)
outputs_conf_new_shape = [model_name + "@_conf_new_shape"]
node_conf_new_shape = onnx.helper.make_node(
'Reshape',
inputs=outputs_conf_set_zero + name_new_shape,
outputs=outputs_conf_new_shape)
node_list.append(node_conf_new_shape)
outputs_score = [model_name + "@score"]
node_score = onnx.helper.make_node(
'Mul',
inputs=outputs_prob_sigmoid + outputs_conf_new_shape,
outputs=outputs_score)
node_list.append(node_score)
outputs_conf_bool = [model_name + "@conf_bool"]
node_conf_bool = onnx.helper.make_node(
'Greater',
inputs=outputs_conf_new_shape + name_zeros,
outputs=outputs_conf_bool)
node_list.append(node_conf_bool)
outputs_box_x_new_shape = [model_name + "@box_x_new_shape"]
node_box_x_new_shape = onnx.helper.make_node(
'Reshape',
inputs=outputs_box_x_encode + name_new_shape,
outputs=outputs_box_x_new_shape)
node_list.append(node_box_x_new_shape)
outputs_box_y_new_shape = [model_name + "@box_y_new_shape"]
node_box_y_new_shape = onnx.helper.make_node(
'Reshape',
inputs=outputs_box_y_encode + name_new_shape,
outputs=outputs_box_y_new_shape)
node_list.append(node_box_y_new_shape)
outputs_box_w_new_shape = [model_name + "@box_w_new_shape"]
node_box_w_new_shape = onnx.helper.make_node(
'Reshape',
inputs=outputs_box_w_encode + name_new_shape,
outputs=outputs_box_w_new_shape)
node_list.append(node_box_w_new_shape)
outputs_box_h_new_shape = [model_name + "@box_h_new_shape"]
node_box_h_new_shape = onnx.helper.make_node(
'Reshape',
inputs=outputs_box_h_encode + name_new_shape,
outputs=outputs_box_h_new_shape)
node_list.append(node_box_h_new_shape)
outputs_pred_box = [model_name + "@pred_box"]
node_pred_box = onnx.helper.make_node(
'Concat',
inputs=outputs_box_x_new_shape + outputs_box_y_new_shape + \
outputs_box_w_new_shape + outputs_box_h_new_shape,
outputs=outputs_pred_box,
axis=4)
node_list.append(node_pred_box)
outputs_conf_cast = [model_name + "conf_cast"]
node_conf_cast = onnx.helper.make_node(
'Cast', inputs=outputs_conf_bool, outputs=outputs_conf_cast, to=1)
node_list.append(node_conf_cast)
outputs_pred_box_mul_conf = [model_name + "@pred_box_mul_conf"]
node_pred_box_mul_conf = onnx.helper.make_node(
'Mul',
inputs=outputs_pred_box + outputs_conf_cast,
outputs=outputs_pred_box_mul_conf)
node_list.append(node_pred_box_mul_conf)
box_shape = [1, int(num_anchors) * input_height * input_width, 4]
name_box_shape = [model_name + "@box_shape"]
node_box_shape = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_box_shape,
value=onnx.helper.make_tensor(
name=name_box_shape[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=[len(box_shape)],
vals=box_shape))
node_list.append(node_box_shape)
outputs_pred_box_new_shape = [model_name + "@pred_box_new_shape"]
node_pred_box_new_shape = onnx.helper.make_node(
'Reshape',
inputs=outputs_pred_box_mul_conf + name_box_shape,
outputs=outputs_pred_box_new_shape)
node_list.append(node_pred_box_new_shape)
outputs_pred_box_x = [model_name + "@_pred_box_x"]
outputs_pred_box_y = [model_name + "@_pred_box_y"]
outputs_pred_box_w = [model_name + "@_pred_box_w"]
outputs_pred_box_h = [model_name + "@_pred_box_h"]
node_pred_box_split = onnx.helper.make_node(
'Split',
inputs=outputs_pred_box_new_shape,
outputs=outputs_pred_box_x + outputs_pred_box_y + outputs_pred_box_w +
outputs_pred_box_h,
axis=2)
node_list.append(node_pred_box_split)
name_number_two = [model_name + "@number_two"]
node_number_two = onnx.helper.make_node(
"Constant",
inputs=[],
outputs=name_number_two,
value=onnx.helper.make_tensor(
name=name_number_two[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=(),
vals=[2]))
node_list.append(node_number_two)
outputs_half_w = [model_name + "@half_w"]
node_half_w = onnx.helper.make_node(
"Div",
inputs=outputs_pred_box_w + name_number_two,
outputs=outputs_half_w)
node_list.append(node_half_w)
outputs_half_h = [model_name + "@half_h"]
node_half_h = onnx.helper.make_node(
"Div",
inputs=outputs_pred_box_h + name_number_two,
outputs=outputs_half_h)
node_list.append(node_half_h)
outputs_pred_box_x1 = [model_name + "@pred_box_x1"]
node_pred_box_x1 = onnx.helper.make_node(
'Sub',
inputs=outputs_pred_box_x + outputs_half_w,
outputs=outputs_pred_box_x1)
node_list.append(node_pred_box_x1)
outputs_pred_box_y1 = [model_name + "@pred_box_y1"]
node_pred_box_y1 = onnx.helper.make_node(
'Sub',
inputs=outputs_pred_box_y + outputs_half_h,
outputs=outputs_pred_box_y1)
node_list.append(node_pred_box_y1)
outputs_pred_box_x2 = [model_name + "@pred_box_x2"]
node_pred_box_x2 = onnx.helper.make_node(
'Add',
inputs=outputs_pred_box_x + outputs_half_w,
outputs=outputs_pred_box_x2)
node_list.append(node_pred_box_x2)
outputs_pred_box_y2 = [model_name + "@pred_box_y2"]
node_pred_box_y2 = onnx.helper.make_node(
'Add',
inputs=outputs_pred_box_y + outputs_half_h,
outputs=outputs_pred_box_y2)
node_list.append(node_pred_box_y2)
outputs_sqeeze_image_size = [model_name + "@sqeeze_image_size"]
node_sqeeze_image_size = onnx.helper.make_node(
"Squeeze",
axes=[0],
inputs=image_size,
outputs=outputs_sqeeze_image_size)
node_list.append(node_sqeeze_image_size)
output_img_height = [model_name + "@img_height"]
output_img_width = [model_name + "@img_width"]
node_image_size_split = onnx.helper.make_node(
"Split",
inputs=outputs_sqeeze_image_size,
outputs=output_img_height + output_img_width,
axis=-1,
split=[1, 1])
node_list.append(node_image_size_split)
output_img_width_cast = [model_name + "@img_width_cast"]
node_img_width_cast = onnx.helper.make_node(
'Cast', inputs=output_img_width, outputs=output_img_width_cast, to=1)
node_list.append(node_img_width_cast)
output_img_height_cast = [model_name + "@img_height_cast"]
node_img_height_cast = onnx.helper.make_node(
'Cast', inputs=output_img_height, outputs=output_img_height_cast, to=1)
node_list.append(node_img_height_cast)
outputs_pred_box_x1_decode = [model_name + "@pred_box_x1_decode"]
outputs_pred_box_y1_decode = [model_name + "@pred_box_y1_decode"]
outputs_pred_box_x2_decode = [model_name + "@pred_box_x2_decode"]
outputs_pred_box_y2_decode = [model_name + "@pred_box_y2_decode"]
node_pred_box_x1_decode = onnx.helper.make_node(
'Mul',
inputs=outputs_pred_box_x1 + output_img_width_cast,
outputs=outputs_pred_box_x1_decode)
node_list.append(node_pred_box_x1_decode)
node_pred_box_y1_decode = onnx.helper.make_node(
'Mul',
inputs=outputs_pred_box_y1 + output_img_height_cast,
outputs=outputs_pred_box_y1_decode)
node_list.append(node_pred_box_y1_decode)
node_pred_box_x2_decode = onnx.helper.make_node(
'Mul',
inputs=outputs_pred_box_x2 + output_img_width_cast,
outputs=outputs_pred_box_x2_decode)
node_list.append(node_pred_box_x2_decode)
node_pred_box_y2_decode = onnx.helper.make_node(
'Mul',
inputs=outputs_pred_box_y2 + output_img_height_cast,
outputs=outputs_pred_box_y2_decode)
node_list.append(node_pred_box_y2_decode)
name_number_one = [model_name + "@one"]
node_number_one = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_number_one,
value=onnx.helper.make_tensor(
name=name_number_one[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=(),
vals=[1]))
node_list.append(node_number_one)
output_new_img_height = [model_name + "@new_img_height"]
node_new_img_height = onnx.helper.make_node(
'Sub',
inputs=output_img_height_cast + name_number_one,
outputs=output_new_img_height)
node_list.append(node_new_img_height)
output_new_img_width = [model_name + "@new_img_width"]
node_new_img_width = onnx.helper.make_node(
'Sub',
inputs=output_img_width_cast + name_number_one,
outputs=output_new_img_width)
node_list.append(node_new_img_width)
outputs_pred_box_x2_sub_w = [model_name + "@pred_box_x2_sub_w"]
node_pred_box_x2_sub_w = onnx.helper.make_node(
'Sub',
inputs=outputs_pred_box_x2_decode + output_new_img_width,
outputs=outputs_pred_box_x2_sub_w)
node_list.append(node_pred_box_x2_sub_w)
outputs_pred_box_y2_sub_h = [model_name + "@pred_box_y2_sub_h"]
node_pred_box_y2_sub_h = onnx.helper.make_node(
'Sub',
inputs=outputs_pred_box_y2_decode + output_new_img_height,
outputs=outputs_pred_box_y2_sub_h)
node_list.append(node_pred_box_y2_sub_h)
outputs_pred_box_x1_clip = [model_name + "@pred_box_x1_clip"]
outputs_pred_box_y1_clip = [model_name + "@pred_box_y1_clip"]
outputs_pred_box_x2_clip = [model_name + "@pred_box_x2_clip"]
outputs_pred_box_y2_clip = [model_name + "@pred_box_y2_clip"]
min_const_name = model_name + "@pred_box_min_const"
max_const_name = model_name + "@pred_box_max_const"
min_const = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=[min_const_name],
value=onnx.helper.make_tensor(
name=min_const_name,
data_type=onnx.TensorProto.FLOAT,
dims=(),
vals=[0.0]))
node_list.append(min_const)
max_const = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=[max_const_name],
value=onnx.helper.make_tensor(
name=max_const_name,
data_type=onnx.TensorProto.FLOAT,
dims=(),
vals=[MAX_FLOAT32]))
node_list.append(max_const)
node_pred_box_x1_clip = onnx.helper.make_node(
'Clip',
inputs=outputs_pred_box_x1_decode + [min_const_name, max_const_name],
outputs=outputs_pred_box_x1_clip)
node_list.append(node_pred_box_x1_clip)
node_pred_box_y1_clip = onnx.helper.make_node(
'Clip',
inputs=outputs_pred_box_y1_decode + [min_const_name, max_const_name],
outputs=outputs_pred_box_y1_clip)
node_list.append(node_pred_box_y1_clip)
node_pred_box_x2_clip = onnx.helper.make_node(
'Clip',
inputs=outputs_pred_box_x2_sub_w + [min_const_name, max_const_name],
outputs=outputs_pred_box_x2_clip)
node_list.append(node_pred_box_x2_clip)
node_pred_box_y2_clip = onnx.helper.make_node(
'Clip',
inputs=outputs_pred_box_y2_sub_h + [min_const_name, max_const_name],
outputs=outputs_pred_box_y2_clip)
node_list.append(node_pred_box_y2_clip)
outputs_pred_box_x2_res = [model_name + "@box_x2_res"]
node_pred_box_x2_res = onnx.helper.make_node(
'Sub',
inputs=outputs_pred_box_x2_decode + outputs_pred_box_x2_clip,
outputs=outputs_pred_box_x2_res)
node_list.append(node_pred_box_x2_res)
outputs_pred_box_y2_res = [model_name + "@box_y2_res"]
node_pred_box_y2_res = onnx.helper.make_node(
'Sub',
inputs=outputs_pred_box_y2_decode + outputs_pred_box_y2_clip,
outputs=outputs_pred_box_y2_res)
node_list.append(node_pred_box_y2_res)
node_pred_box_result = onnx.helper.make_node(
'Concat',
inputs=outputs_pred_box_x1_clip + outputs_pred_box_y1_clip +
outputs_pred_box_x2_res + outputs_pred_box_y2_res,
outputs=outputs['Boxes'],
axis=-1)
node_list.append(node_pred_box_result)
score_shape = [1, input_height * input_width * int(num_anchors), class_num]
name_score_shape = [model_name + "@score_shape"]
node_score_shape = onnx.helper.make_node(
"Constant",
inputs=[],
outputs=name_score_shape,
value=onnx.helper.make_tensor(
name=name_score_shape[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=[len(score_shape)],
vals=score_shape))
node_list.append(node_score_shape)
node_score_new_shape = onnx.helper.make_node(
'Reshape',
inputs=outputs_score + name_score_shape,
outputs=outputs['Scores'])
node_list.append(node_score_new_shape)
return node_list
x2paddle/x2paddle/op_mapper/paddle2onnx/opset9/opset.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
import sys
import x2paddle
import os
import numpy as np
import paddle.fluid.core as core
import paddle.fluid as fluid
import onnx
from onnx import helper, onnx_pb
class OpSet9(object):
def __init__(self):
self.paddle_onnx_dtype_map = {
core.VarDesc.VarType.FP32: onnx_pb.TensorProto.FLOAT,
core.VarDesc.VarType.FP64: onnx_pb.TensorProto.DOUBLE,
core.VarDesc.VarType.INT32: onnx_pb.TensorProto.INT32,
core.VarDesc.VarType.INT16: onnx_pb.TensorProto.INT16,
core.VarDesc.VarType.INT16: onnx_pb.TensorProto.UINT16,
core.VarDesc.VarType.INT64: onnx_pb.TensorProto.INT64,
core.VarDesc.VarType.BOOL: onnx_pb.TensorProto.BOOL
}
self.name_counter = dict()
def get_name(self, op_name, var_name):
name = 'p2o.{}.{}'.format(op_name, var_name)
if name not in self.name_counter:
self.name_counter[name] = 0
else:
self.name_counter[name] += 1
return name + '.{}'.format(self.name_counter[name])
def make_constant_node(self, name, dtype, value=None):
if isinstance(value, list):
dims = (len(value), )
elif value is None:
dims = ()
value = []
else:
dims = ()
value = [value]
tensor = helper.make_tensor(
name=name, data_type=dtype, dims=dims, vals=value)
node = helper.make_node(
'Constant', inputs=[], outputs=[name], value=tensor)
return node
def convert_weights(self, program, scope=None):
var_names = program.global_block().vars
nodes = list()
for name in var_names:
var = program.global_block().var(name)
if name.endswith('feed') or name.endswith('fetch'):
continue
if not var.persistable:
continue
weight = np.array(scope.find_var(name).get_tensor())
tensor = helper.make_tensor(
name=name,
dims=var.shape,
data_type=self.paddle_onnx_dtype_map[var.dtype],
vals=weight.flatten().tolist())
node = helper.make_node(
'Constant', inputs=[], outputs=[name], value=tensor)
nodes.append(node)
return nodes
def conv2d(self, op, block):
kernel_shape = block.var(op.input('Filter')[0]).shape
node = helper.make_node(
'Conv',
inputs=op.input('Input') + op.input('Filter'),
outputs=op.output('Output'),
dilations=op.attr('dilations'),
kernel_shape=kernel_shape[-2:],
strides=op.attr('strides'),
group=op.attr('groups'),
pads=op.attr('paddings') + op.attr('paddings'))
return node
def conv2d_transpose(self, op, block):
kernel_shape = block.var(op.input('Filter')[0]).shape
node = helper.make_node(
'ConvTranspose',
inputs=op.input('Input') + op.input('Filter'),
outputs=op.output('Output'),
dilations=op.attr('dilations'),
kernel_shape=kernel_shape[-2:],
strides=op.attr('strides'),
group=1,
pads=op.attr('paddings') + op.attr('paddings'))
return node
def relu(self, op, block):
node = helper.make_node(
'Relu', inputs=op.input('X'), outputs=op.output('Out'))
return node
def tanh(self, op, block):
node = helper.make_node(
'Tanh', inputs=op.input('X'), outputs=op.output('Out'))
return node
def log(self, op, block):
node = helper.make_node(
'Log', inputs=op.input('X'), outputs=op.output('Out'))
return node
def sigmoid(self, op, block):
node = helper.make_node(
'Sigmoid', inputs=op.input('X'), outputs=op.output('Out'))
return node
def clip(self, op, block):
min_value = op.attr('min')
max_value = op.attr('max')
node = helper.make_node(
'Clip',
inputs=[op.input('X')[0]],
outputs=op.output('Out'),
max=max_value,
min=min_value)
return node
def exp(self, op, block):
node = helper.make_node(
'Exp', inputs=op.input('X'), outputs=op.output('Out'))
return node
def abs(self, op, block):
node = helper.make_node(
'Abs', inputs=op.input('X'), outputs=op.output('Out'))
return node
def leaky_relu(self, op, block):
node = helper.make_node(
'LeakyRelu',
inputs=op.input('X'),
outputs=op.output('Out'),
alpha=op.attr('alpha'))
return node
def elementwise_add(self, op, block):
axis = op.attr('axis')
x_shape = block.var(op.input('X')[0]).shape
y_shape = block.var(op.input('Y')[0]).shape
if len(y_shape) == 1 and axis == 1:
shape_name = self.get_name(op.type, 'shape')
shape_value = [1] * len(x_shape)
shape_value[axis] = y_shape[0]
shape_node = self.make_constant_node(
shape_name, onnx_pb.TensorProto.INT64, shape_value)
temp_value = self.get_name(op.type, 'temp')
y_node = helper.make_node(
'Reshape',
inputs=[op.input('Y')[0], shape_name],
outputs=[temp_value])
node = helper.make_node(
'Add',
inputs=[op.input('X')[0], temp_value],
outputs=op.output('Out'))
return [shape_node, y_node, node]
elif axis == -1 or axis == (len(x_shape) - 1
) or len(x_shape) == len(y_shape):
node = helper.make_node(
'Add',
inputs=[op.input('X')[0], op.input('Y')[0]],
outputs=op.output('Out'))
return node
else:
raise Exception("Unexpected situation happend in elementwise_add")
def elementwise_sub(self, op, block):
axis = op.attr('axis')
x_shape = block.var(op.input('X')[0]).shape
y_shape = block.var(op.input('Y')[0]).shape
if len(y_shape) == 1 and axis == 1:
shape_name = self.get_name(op.type, 'shape')
shape_value = [1] * len(x_shape)
shape_value[axis] = y_shape[0]
shape_node = self.make_constant_node(
shape_name, onnx_pb.TensorProto.INT64, shape_value)
temp_value = self.get_name(op.type, 'temp')
y_node = helper.make_node(
'Reshape',
inputs=[op.input('Y')[0], shape_name],
outputs=[temp_value])
node = helper.make_node(
'Sub',
inputs=[op.input('X')[0], temp_value],
outputs=op.output('Out'))
return [shape_node, y_node, node]
elif axis == -1 or axis == (len(x_shape) - 1
) or len(x_shape) == len(y_shape):
node = helper.make_node(
'Sub',
inputs=[op.input('X')[0], op.input('Y')[0]],
outputs=op.output('Out'))
return node
else:
raise Exception("Unexpected situation happend in elementwise_sub")
def pool2d(self, op, block):
pool_type = {
'max': ('MaxPool', 'GlobalMaxPool'),
'avg': ('AveragePool', 'GlobalAveragePool')
}
if op.attr('global_pooling'):
node = helper.make_node(
pool_type[op.attr('pooling_type')][1],
inputs=op.input('X'),
outputs=op.output('Out'), )
elif op.attr('adaptive'):
raise Excpetion("ONNX cannot support adaptive pool")
else:
input_shape = block.var(op.input('X')[0]).shape
k_size = op.attr('ksize')
paddings = op.attr('paddings')
if input_shape[2] > 0 and input_shape[2] + paddings[0] < k_size[0]:
k_size[0] = input_shape[2] + paddings[0]
if input_shape[3] > 0 and input_shape[3] + paddings[1] < k_size[1]:
k_size[1] = input_shape[3] + paddings[1]
node = helper.make_node(
pool_type[op.attr('pooling_type')][0],
inputs=op.input('X'),
outputs=op.output('Out'),
kernel_shape=k_size,
strides=op.attr('strides'),
pads=op.attr('paddings') + op.attr('paddings'))
return node
def pad2d(self, op, block):
x_shape = block.var(op.input('X')[0]).shape
paddings = op.attr('paddings')
onnx_pads = []
if op.attr('data_format') == 'NCHW':
pads = [
0, 0, paddings[0], paddings[2], 0, 0, paddings[1], paddings[3]
]
else:
pads = [
0, paddings[0], paddings[2], 0, 0, paddings[1], paddings[3], 0
]
#TODO support pads is Variable
node = helper.make_node(
'Pad',
inputs=op.input('X'),
outputs=op.output('Out'),
mode=op.attr('mode'),
value=op.attr('pad_value'),
pads=pads)
return node
def softmax(self, op, block):
axis = op.attr('axis')
shape = block.var(op.output('Out')[0]).shape
if axis < 0:
axis += len(shape)
if axis == len(shape) - 1:
node = helper.make_node(
'Softmax',
inputs=op.input('X'),
outputs=op.output('Out'),
axis=op.attr('axis'))
return node
else:
perm = [i for i in range(len(shape))]
perm[-1] = axis
perm[axis] = len(shape) - 1
transpose_name0 = self.get_name(op.type, 'transpose')
transpose_node0 = helper.make_node(
'Transpose',
inputs=op.input('X'),
outputs=[transpose_name0],
perm=perm)
softmax_name = self.get_name(op.type, 'softmax')
softmax_node = helper.make_node(
'Softmax',
inputs=[transpose_name0],
outputs=[softmax_name],
axis=-1)
transpose_name1 = self.get_name(op.type, 'transpose')
transpose_node1 = helper.make_node(
'Transpose',
inputs=[softmax_name],
outputs=op.output('Out'),
perm=perm)
return [transpose_node0, softmax_node, transpose_node1]
def scale(self, op, block):
scale = op.attr('scale')
bias = op.attr('bias')
if math.fabs(scale - 1.0) < 1e-06 and math.fabs(bias - 0.0) < 1e-06:
node = helper.make_node(
'Identity', inputs=op.input('X'), outputs=op.output('Out'))
return node
else:
scale_name = self.get_name(op.type, 'scale')
bias_name = self.get_name(op.type, 'bias')
scale_node = self.make_constant_node(
scale_name, onnx_pb.TensorProto.FLOAT, scale)
bias_node = self.make_constant_node(bias_name,
onnx_pb.TensorProto.FLOAT, bias)
temp_tensor_name = self.get_name(op.type, 'temporary')
if op.attr('bias_after_scale'):
node1 = helper.make_node(
'Mul',
inputs=[scale_name, op.input('X')[0]],
outputs=[temp_tensor_name])
node2 = helper.make_node(
'Add',
inputs=[bias_name, temp_tensor_name],
outputs=op.output('Out'))
else:
node1 = helper.make_node(
'Add',
inputs=[bias_name, op.input('X')[0]],
outputs=temp_tensor_name)
node2 = helper.make_node(
'Mul',
inputs=[scale_name, temp_tensor_name],
outputs=[op.output('Out')])
return [scale_node, bias_node, node1, node2]
def mul(self, op, block):
x_shape = block.var(op.input('X')[0]).shape
y_shape = block.var(op.input('Y')[0]).shape
out_shape = list(block.var(op.output('Out')[0]).shape)
x_num_col_dims = op.attr('x_num_col_dims')
y_num_col_dims = op.attr('y_num_col_dims')
flatten_x_name = 'flatten_{}'.format(op.input('X')[0])
flatten_y_name = 'flatten_{}'.format(op.input('Y')[0])
shape_name = 'temp_shape_{}'.format(op.output('Out')[0])
temp_out_name = 'temp_{}'.format(op.output('Out')[0])
flatten_x = helper.make_node(
'Flatten',
inputs=op.input('X'),
outputs=[flatten_x_name],
axis=x_num_col_dims)
flatten_y = helper.make_node(
'Flatten',
inputs=op.input('Y'),
outputs=[flatten_y_name],
axis=y_num_col_dims)
shape_node = self.make_constant_node(
shape_name, onnx_pb.TensorProto.INT64, out_shape)
node = helper.make_node(
'MatMul',
inputs=[flatten_x_name, flatten_y_name],
outputs=[temp_out_name])
reshape_out = helper.make_node(
'Reshape',
inputs=[temp_out_name, shape_name],
outputs=op.output('Out'))
return [flatten_x, flatten_y, shape_node, node, reshape_out]
def batch_norm(self, op, block):
kwargs = {
'epsilon': op.attr('epsilon'),
'momentum': op.attr('momentum')
}
inputs = op.input('X') + op.input('Scale') + op.input(
'Bias') + op.input('Mean') + op.input('Variance')
node = helper.make_node(
'BatchNormalization',
inputs=inputs,
outputs=op.output('Y'),
**kwargs)
return node
def instance_norm(self, op, block):
kwargs = {'epsilon': op.attr('epsilon'), }
inputs = op.input('X') + op.input('Scale') + op.input('Bias')
node = helper.make_node(
'InstanceNormalization',
inputs=inputs,
outputs=op.output('Y'),
**kwargs)
return node
def concat(self, op, block):
node = helper.make_node(
'Concat',
inputs=op.input('X'),
outputs=op.output('Out'),
axis=op.attr('axis'))
return node
def sum(self, op, block):
node = helper.make_node(
'Sum', inputs=op.input('X'), outputs=op.output('Out'))
return node
def floor(self, op, block):
node = helper.make_node(
'Floor', inputs=op.input('X'), outputs=op.output('Out'))
return node
def uniform_random_batch_size_like(self, op, block):
node = helper.make_node(
'RandomUniformLike',
inputs=op.input('Input'),
outputs=op.output('Out'),
high=op.attr('max'),
dtype=self.paddle_onnx_dtype_map[op.attr('dtype')],
low=op.attr('min'),
seed=float(op.attr('seed')), )
return node
def depthwise_conv2d(self, op, block):
return self.conv2d(op, block)
def relu6(self, op, block):
threshold = op.attr('threshold')
node = helper.make_node(
'Clip',
inputs=[op.input('X')[0]],
outputs=op.output('Out'),
max=threshold,
min=0.0)
return [node]
def shape(self, op, block):
node = helper.make_node(
'Shape', inputs=op.input('Input'), outputs=op.output('Out'))
return node
def split(self, op, block):
sections = op.attr('sections')
if len(sections) > 0:
node = helper.make_node(
'Split',
inputs=op.input('X'),
outputs=op.output('Out'),
axis=op.attr('axis'),
split=sections)
else:
node = helper.make_node(
'Split',
inputs=op.input('X'),
outputs=op.output('Out'),
axis=op.attr('axis'))
return node
def slice(self, op, block):
axes = op.attr('axes')
starts = op.attr('starts')
ends = op.attr('ends')
node = helper.make_node(
"Slice",
inputs=[op.input('Input')[0]],
outputs=op.output('Out'),
axes=axes,
starts=starts,
ends=ends)
return [node]
def fill_constant(self, op, block):
value = op.attr('value')
dtype = op.attr('dtype')
shape = op.attr('shape')
value = np.ones(shape) * value
if dtype == 2:
value = value.astype('int32')
node = helper.make_node(
'Constant',
inputs=[],
outputs=op.output('Out'),
value=helper.make_tensor(
name=op.output('Out')[0],
data_type=self.paddle_onnx_dtype_map[dtype],
dims=shape,
vals=value.tolist()))
return node
def transpose2(self, op, block):
node = helper.make_node(
'Transpose',
inputs=op.input('X'),
outputs=op.output('Out'),
perm=op.attr('axis'))
return node
def flatten2(self, op, block):
node = helper.make_node(
'Flatten',
inputs=op.input('X'),
outputs=op.output('Out'),
axis=op.attr('axis'))
return node
def reshape2(self, op, block):
input_names = op.input_names
if len(op.input('ShapeTensor')) > 1:
cast_shape_nodes = list()
cast_shape_names = list()
for i in range(len(op.input('ShapeTensor'))):
dim = op.input('ShapeTensor')[i]
temp_name = self.get_name(op.type, 'shape.cast')
node = helper.make_node(
'Cast',
inputs=[dim],
outputs=[temp_name],
to=onnx_pb.TensorProto.INT64)
cast_shape_nodes.append(node)
cast_shape_names.append(temp_name)
temp_name = self.get_name(op.type, 'shape.concat')
shape_node = helper.make_node(
'Concat', inputs=cast_shape_names, outputs=[temp_name], axis=-1)
node = helper.make_node(
'Reshape',
inputs=[op.input('X')[0], temp_name],
outputs=op.output('Out'))
return cast_shape_nodes + [shape_node, node]
elif len(op.input('ShapeTensor')) == 1:
temp_name = self.get_name(op.type, 'shape.cast')
cast_shape_node = helper.make_node(
'Cast',
inputs=op.input('ShapeTensor'),
outputs=[temp_name],
to=onnx_pb.TensorProto.INT64)
node = helper.make_node(
'Reshape',
inputs=[op.input('X')[0], temp_name],
outputs=op.output('Out'))
return [cast_shape_node, node]
elif op.attr('shape') is not None and len(op.attr('shape')) > 0:
shape_name = self.get_name(op.type, 'shape')
shape_node = self.make_constant_node(shape_name,
onnx_pb.TensorProto.INT64,
op.attr('shape'))
reshape_node = helper.make_node(
'Reshape',
inputs=[op.input('X')[0], shape_name],
outputs=op.output('Out'))
return [shape_node, reshape_node]
def dropout(self, op, block):
dropout_mode = op.attr('dropout_implementation')
dropout_prob = op.attr('dropout_prob')
if dropout_mode == 'upscale_in_train':
node = helper.make_node(
'Identity', inputs=op.input('X'), outputs=op.output('Out'))
return node
elif dropout_mode == 'downgrade_in_infer':
scale_name = self.get_name(op.type, 'scale')
scale_node = self.make_constant_node(
scale_name, onnx_pb.TensorProto.FLOAT, 1 - dropout_prob)
node = helper.make_node(
"Mul",
inputs=[op.input('X')[0], scale_name],
outputs=op.output('Out'))
return [scale_node, node]
else:
raise Exception("Unexpected situation happend")
def reduce_mean(self, op, block):
node = helper.make_node(
'ReduceMean',
inputs=op.input('X'),
outputs=op.output('Out'),
axes=op.attr('dim'),
keepdims=op.attr('keep_dim'))
return node
def bilinear_interp(self, op, block):
input_names = op.input_names
input_shape = block.vars[op.input('X')[0]].shape
if op.attr('align_corners') or op.attr('align_mode') == 0:
raise Exception(
"Resize in onnx(opset<=10) only support coordinate_transformation_mode: 'asymmetric', Try converting with --onnx_opset 11"
)
if ('OutSize' in input_names and len(op.input('OutSize')) > 0) or (
'SizeTensor' in input_names and
len(op.input('SizeTensor')) > 0):
node_list = list()
shape_name0 = self.get_name(op.type, 'shape')
shape_node0 = helper.make_node(
'Shape', inputs=op.input('X'), outputs=[shape_name0])
starts_name = self.get_name(op.type, 'slice.starts')
starts_node = self.make_constant_node(
starts_name, onnx_pb.TensorProto.INT64, [0])
ends_name = self.get_name(op.type, 'slice.ends')
ends_node = self.make_constant_node(ends_name,
onnx_pb.TensorProto.INT64, [2])
shape_name1 = self.get_name(op.type, 'shape')
shape_node1 = helper.make_node(
'Slice',
inputs=[shape_name0, starts_name, ends_name],
outputs=[shape_name1])
node_list.extend([shape_node0, starts_node, ends_node, shape_node1])
if 'OutSize' in input_names and len(op.input('OutSize')) > 0:
cast_shape_name = self.get_name(op.type, "shape.cast")
cast_shape_node = helper.make_node(
'Cast',
inputs=op.input('OutSize'),
outputs=[cast_shape_name],
to=onnx_pb.TensorProto.INT64)
node_list.append(cast_shape_node)
else:
concat_shape_name = self.get_name(
op.type, op.output('Out')[0] + "shape.concat")
concat_shape_node = helper.make_node(
"Concat",
inputs=op.input('SizeTensor'),
outputs=[concat_shape_name],
axis=0)
cast_shape_name = self.get_name(op.type, "shape.cast")
cast_shape_node = helper.make_node(
'Cast',
inputs=[concat_shape_name],
outputs=[cast_shape_name],
to=onnx_pb.TensorProto.INT64)
node_list.extend([concat_shape_node, cast_shape_node])
shape_name2 = self.get_name(op.type, "shape.concat")
shape_node2 = helper.make_node(
'Concat',
inputs=[shape_name1, cast_shape_name],
outputs=[shape_name2],
axis=0)
node_list.append(shape_node2)
cast_shape_name2 = self.get_name(op.type, "shape.cast")
cast_shape_node2 = helper.make_node(
'Cast',
inputs=[shape_name2],
outputs=[cast_shape_name2],
to=onnx_pb.TensorProto.FLOAT)
node_list.append(cast_shape_node2)
cast_shape_name0 = self.get_name(op.type, "shape.cast")
cast_shape_node0 = helper.make_node(
'Cast',
inputs=[shape_name0],
outputs=[cast_shape_name0],
to=onnx_pb.TensorProto.FLOAT)
node_list.append(cast_shape_node0)
outputs_h_w_scales = op.output('Out')[0] + "@out_hw_scales"
node_h_w_scales = helper.make_node(
'Div',
inputs=[cast_shape_name2, cast_shape_name0],
outputs=[outputs_h_w_scales])
node_list.append(node_h_w_scales)
result_node = helper.make_node(
'Resize',
inputs=[op.input('X')[0], outputs_h_w_scales],
outputs=op.output('Out'),
mode='linear')
node_list.extend([result_node])
return node_list
elif 'Scale' in input_names and len(op.input('Scale')) > 0:
node = helper.make_node(
'Resize',
inputs=[op.input('X')[0], op.input('Scale')[0]],
outputs=op.output('Out'),
mode='linear')
else:
out_shape = [op.attr('out_h'), op.attr('out_w')]
scale = op.attr('scale')
if out_shape.count(-1) > 0:
scale_name = self.get_name(op.type, 'scale')
scale_node = self.make_constant_node(scale_name,
onnx_pb.TensorProto.FLOAT,
[1, 1, scale, scale])
node = helper.make_node(
'Resize',
inputs=[op.input('X')[0], scale_name],
outputs=op.output('Out'),
mode='linear')
return [scale_node, node]
else:
raise Exception("Unexpected situation happend")
return node
def nearest_interp(self, op, block):
input_names = op.input_names
if op.attr('align_corners'):
raise Exception(
"Resize in onnx(opset<=10) only support coordinate_transformation_mode: 'asymmetric', Try converting with --onnx_opset 11"
)
if 'OutSize' in input_names and len(op.input('OutSize')) > 0:
node_list = list()
shape_name0 = self.get_name(op.type, 'shape')
shape_node0 = helper.make_node(
'Shape', inputs=op.input('X'), outputs=[shape_name0])
starts_name = self.get_name(op.type, 'slice.starts')
starts_node = self.make_constant_node(
starts_name, onnx_pb.TensorProto.INT64, [0])
ends_name = self.get_name(op.type, 'slice.ends')
ends_node = self.make_constant_node(ends_name,
onnx_pb.TensorProto.INT64, [2])
shape_name1 = self.get_name(op.type, 'shape')
shape_node1 = helper.make_node(
'Slice',
inputs=[shape_name0, starts_name, ends_name],
outputs=[shape_name1])
node_list.extend([shape_node0, starts_node, ends_node, shape_node1])
if 'OutSize' in input_names and len(op.input('OutSize')) > 0:
cast_shape_name = self.get_name(op.type, "shape.cast")
cast_shape_node = helper.make_node(
'Cast',
inputs=op.input('OutSize'),
outputs=[cast_shape_name],
to=onnx_pb.TensorProto.INT64)
node_list.append(cast_shape_node)
else:
concat_shape_name = self.get_name(
op.type, op.output('Out')[0] + "shape.concat")
concat_shape_node = helper.make_node(
"Concat",
inputs=op.input('SizeTensor'),
outputs=[concat_shape_name],
axis=0)
cast_shape_name = self.get_name(op.type, "shape.cast")
cast_shape_node = helper.make_node(
'Cast',
inputs=[concat_shape_name],
outputs=[cast_shape_name],
to=onnx_pb.TensorProto.INT64)
node_list.extend([concat_shape_node, cast_shape_node])
shape_name2 = self.get_name(op.type, "shape.concat")
shape_node2 = helper.make_node(
'Concat',
inputs=[shape_name1, cast_shape_name],
outputs=[shape_name2],
axis=0)
node_list.append(shape_node2)
cast_shape_name2 = self.get_name(op.type, "shape.cast")
cast_shape_node2 = helper.make_node(
'Cast',
inputs=[shape_name2],
outputs=[cast_shape_name2],
to=onnx_pb.TensorProto.FLOAT)
node_list.append(cast_shape_node2)
cast_shape_name0 = self.get_name(op.type, "shape.cast")
cast_shape_node0 = helper.make_node(
'Cast',
inputs=[shape_name0],
outputs=[cast_shape_name0],
to=onnx_pb.TensorProto.FLOAT)
node_list.append(cast_shape_node0)
outputs_h_w_scales = op.output('Out')[0] + "@out_hw_scales"
node_h_w_scales = helper.make_node(
'Div',
inputs=[cast_shape_name2, cast_shape_name0],
outputs=[outputs_h_w_scales])
node_list.append(node_h_w_scales)
result_node = helper.make_node(
'Resize',
inputs=[op.input('X')[0], outputs_h_w_scales],
outputs=op.output('Out'),
mode='linear')
node_list.extend([result_node])
return node_list
elif 'Scale' in input_names and len(op.input('Scale')) > 0:
node = helper.make_node(
'Resize',
inputs=[op.input('X')[0], op.input('Scale')[0]],
outputs=op.output('Out'),
mode='nearest')
else:
out_shape = [op.attr('out_h'), op.attr('out_w')]
scale = op.attr('scale')
if out_shape.count(-1) > 0:
scale_name = self.get_name(op.type, 'scale')
scale_node = self.make_constant_node(scale_name,
onnx_pb.TensorProto.FLOAT,
[1, 1, scale, scale])
node = helper.make_node(
'Resize',
inputs=[op.input('X')[0], scale_name],
outputs=op.output('Out'),
mode='nearest')
return [scale_node, node]
else:
raise Exception("Unexpected situation happend")
return node
def hard_sigmoid(self, op, block):
slope = op.attr('slope')
offset = op.attr('offset')
node = helper.make_node(
'HardSigmoid',
inputs=op.input('X'),
outputs=op.output('Out'),
alpha=slope,
beta=offset)
return node
def swish(self, op, block):
beta = op.attr('beta')
beta_name = self.get_name(op.type, 'beta')
beta_node = onnx.helper.make_node(
'Constant',
name=beta_name,
inputs=[],
outputs=[beta_name],
value=onnx.helper.make_tensor(
name=beta_name,
data_type=onnx.TensorProto.FLOAT,
dims=(),
vals=[beta]))
beta_x_name = self.get_name(op.type, 'beta_x')
beta_x_node = onnx.helper.make_node(
'Mul',
name=beta_x_name,
inputs=[op.input('X')[0], beta_name],
outputs=[beta_x_name])
sigmoid_name = self.get_name(op.type, 'sigmoid')
sigmoid_node = onnx.helper.make_node(
'Sigmoid',
name=sigmoid_name,
inputs=[beta_x_name],
outputs=[sigmoid_name])
swish_node = onnx.helper.make_node(
'Mul',
inputs=[op.input('X')[0], sigmoid_name],
outputs=op.output('Out'))
return [beta_node, beta_x_node, sigmoid_node, swish_node]
def hard_swish(self, op, block):
scale_name = self.get_name(op.type, 'scale')
offset_name = self.get_name(op.type, 'offset')
scale_node = self.make_constant_node(scale_name,
onnx_pb.TensorProto.FLOAT,
op.attr('scale'))
offset_node = self.make_constant_node(offset_name,
onnx_pb.TensorProto.FLOAT,
op.attr('offset'))
name0 = self.get_name(op.type, 'add')
node0 = helper.make_node(
'Add', inputs=[op.input('X')[0], offset_name], outputs=[name0])
name1 = self.get_name(op.type, 'relu')
min_value = op.attr('min')
max_value = op.attr('max')
node1 = helper.make_node(
'Clip',
inputs=[name0],
outputs=[name1],
max=max_value,
min=min_value)
name2 = self.get_name(op.type, 'mul')
node2 = helper.make_node(
'Mul', inputs=[op.input('X')[0], name1], outputs=[name2])
node3 = helper.make_node(
'Div', inputs=[name2, scale_name], outputs=op.output('Out'))
return [scale_node, offset_node, node0, node1, node2, node3]
def elementwise_mul(self, op, block):
axis = op.attr('axis')
x_shape = block.var(op.input('X')[0]).shape
y_shape = block.var(op.input('Y')[0]).shape
if len(y_shape) == 1 and axis == 1:
shape_name = self.get_name(op.type, 'shape')
shape_value = [1] * len(x_shape)
shape_value[axis] = y_shape[0]
shape_node = self.make_constant_node(
shape_name, onnx_pb.TensorProto.INT64, shape_value)
temp_value = self.get_name(op.type, 'temp')
y_node = helper.make_node(
'Reshape',
inputs=[op.input('Y')[0], shape_name],
outputs=[temp_value])
node = helper.make_node(
'Mul',
inputs=[op.input('X')[0], temp_value],
outputs=op.output('Out'))
return [shape_node, y_node, node]
elif axis == -1 or axis == (len(x_shape) - 1
) or len(x_shape) == len(y_shape):
node = helper.make_node(
'Mul',
inputs=[op.input('X')[0], op.input('Y')[0]],
outputs=op.output('Out'))
return node
else:
raise Exception("Unexpected situation happend in elementwise_mul")
return node
def feed(self, op, block):
name = op.output('Out')[0]
var = block.var(name)
tensor_info = helper.make_tensor_value_info(
name=name,
shape=var.shape,
elem_type=self.paddle_onnx_dtype_map[var.dtype])
return tensor_info
def fetch(self, op, block):
name = op.input('X')[0]
var = block.var(name)
tensor_info = helper.make_tensor_value_info(
name=name,
shape=var.shape,
elem_type=self.paddle_onnx_dtype_map[var.dtype])
return tensor_info
def unsqueeze2(self, op, block):
node = helper.make_node(
'Unsqueeze',
inputs=op.input('X'),
outputs=op.output('Out'),
axes=op.attr('axes'))
return node
def cast(self, op, block):
node = helper.make_node(
'Cast',
inputs=op.input('X'),
outputs=op.output('Out'),
to=self.paddle_onnx_dtype_map[op.attr('out_dtype')])
return node
def arg_max(self, op, block):
node = helper.make_node(
'ArgMax',
inputs=op.input('X'),
outputs=op.output('Out'),
axis=op.attr('axis'),
keepdims=0)
return node
def reciprocal(self, op, block):
inputs = op.input(op.input_names[0])
outputs = op.output(op.output_names[0])
node = helper.make_node('Reciprocal', inputs=inputs, outputs=outputs)
return node
def im2sequence(self, op, block):
from .paddle_custom_layer.im2sequence import im2sequence
return im2sequence(op, block)
def yolo_box(self, op, block):
from .paddle_custom_layer.yolo_box import yolo_box
return yolo_box(op, block)
def multiclass_nms(self, op, block):
from .paddle_custom_layer.multiclass_nms import multiclass_nms
return multiclass_nms(op, block)
def box_coder(self, op, block):
from .paddle_custom_layer.box_coder import box_coder
return box_coder(op, block)
def prior_box(self, op, block):
from .paddle_custom_layer.prior_box import prior_box
return prior_box(op, block)
x2paddle/x2paddle/op_mapper/paddle2onnx/opset9/paddle_custom_layer/box_coder.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import sys
import math
import onnx
import warnings
import numpy as np
from functools import partial
from onnx import TensorProto
from onnx.helper import make_node, make_tensor
from onnx import onnx_pb
from paddle.fluid.executor import _fetch_var as fetch_var
from onnx import helper
import paddle.fluid as fluid
import paddle.fluid.core as core
def box_coder(op, block):
"""
In this function, we will use the decode the prior box to target box,
we just use the decode mode to transform this op.
"""
node_list = []
input_names = op.input_names
prior_var = block.var(op.input('PriorBox')[0])
t_size = block.var(op.input('TargetBox')[0]).shape
p_size = prior_var.shape
# get the outout_name
result_name = op.output('OutputBox')[0]
# n is size of batch, m is boxes num of targe_boxes
n = t_size[0]
m = t_size[0]
axis = int(op.attr('axis'))
#norm
norm = bool(op.attr('box_normalized'))
name_slice_x1 = op.output('OutputBox')[0] + "@x1"
name_slice_y1 = op.output('OutputBox')[0] + "@y1"
name_slice_x2 = op.output('OutputBox')[0] + "@x2"
name_slice_y2 = op.output('OutputBox')[0] + "@y2"
#make onnx tensor to save the intermeidate reslut
name_slice_indices = [[op.output('OutputBox')[0] + "@slice_" + str(i)]
for i in range(1, 3)]
node_slice_indices = [None for i in range(1, 3)]
# create the range(0, 4) const data to slice
for i in range(1, 3):
node = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_slice_indices[i - 1],
value=onnx.helper.make_tensor(
name=name_slice_indices[i - 1][0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=(),
vals=[i]))
node_list.append(node)
# make node split data
name_box_split = [
name_slice_x1, name_slice_y1, name_slice_x2, name_slice_y2
]
split_shape = list(p_size)
split_shape[-1] = 1
node_split_prior_node = onnx.helper.make_node(
'Split', inputs=op.input('PriorBox'), outputs=name_box_split, axis=1)
node_list.append(node_split_prior_node)
# make node get centor node for decode
final_outputs_vars = []
if not norm:
name_centor_w_tmp = [op.output('OutputBox')[0] + "@centor_w_tmp"]
name_centor_h_tmp = [op.output('OutputBox')[0] + "@centor_h_tmp"]
node_centor_w_tmp = None
node_centor_h_tmp = None
name_centor_tmp_list = [name_centor_w_tmp, name_centor_h_tmp]
node_centor_tmp_list = [node_centor_w_tmp, node_centor_h_tmp]
count = 2
for (name, node) in zip(name_centor_tmp_list, node_centor_tmp_list):
node = onnx.helper.make_node('Add',
inputs=[op.output('OutputBox')[0] + "@slice_" + str(1)]\
+ [name_box_split[count]],
outputs=name)
node_list.append(node)
count = count + 1
if not norm:
inputs_sub = [[name_centor_w_tmp[0], name_box_split[0]],
[name_centor_h_tmp[0], name_box_split[1]]]
else:
inputs_sub = [[name_box_split[2], name_box_split[0]],
[name_box_split[3], name_box_split[1]]]
outputs_sub = [result_name + "@pb_w", result_name + "@pb_h"]
for i in range(0, 2):
node = onnx.helper.make_node(
'Sub', inputs=inputs_sub[i], outputs=[outputs_sub[i]])
node_list.append(node)
# according to prior_box height and weight to get centor x, y
name_half_value = [result_name + "@half_value"]
node_half_value = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_half_value,
value=onnx.helper.make_tensor(
name=name_slice_indices[i][0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=(),
vals=[0.5]))
node_list.append(node_half_value)
outputs_half_wh = [[result_name + "@pb_w_half"],
[result_name + "@pb_h_half"]]
inputs_half_wh = [[result_name + "@pb_w", name_half_value[0]],
[result_name + "@pb_h", name_half_value[0]]]
for i in range(0, 2):
node = onnx.helper.make_node(
'Mul', inputs=inputs_half_wh[i], outputs=outputs_half_wh[i])
node_list.append(node)
inputs_centor_xy = [[outputs_half_wh[0][0], name_slice_x1],
[outputs_half_wh[1][0], name_slice_y1]]
outputs_centor_xy = [[result_name + "@pb_x"], [result_name + "@pb_y"]]
# final calc the centor x ,y
for i in range(0, 2):
node = onnx.helper.make_node(
'Add', inputs=inputs_centor_xy[i], outputs=outputs_centor_xy[i])
node_list.append(node)
# reshape the data
shape = (1, split_shape[0]) if axis == 0 else (split_shape[0], 1)
# need to reshape the data
inputs_transpose_pb = [
[result_name + "@pb_w"],
[result_name + "@pb_h"],
[result_name + "@pb_x"],
[result_name + "@pb_y"],
]
outputs_transpose_pb = [
[result_name + "@pb_w_transpose"],
[result_name + "@pb_h_transpose"],
[result_name + "@pb_x_transpose"],
[result_name + "@pb_y_transpose"],
]
if axis == 0:
name_reshape_pb = [result_name + "@pb_transpose"]
# reshape the data
for i in range(0, 4):
node = onnx.helper.make_node(
'Transpose',
inputs=inputs_transpose_pb[i],
outputs=outputs_transpose_pb[i])
node_list.append(node)
# decoder the box according to the target_box and variacne
name_variance_raw = [result_name + "@variance_raw"]
name_variance_unsqueeze = [result_name + "@variance_unsqueeze"]
shape = []
# make node to extend the data
var_split_axis = 0
var_split_inputs_name = []
if 'PriorBoxVar' in input_names and len(op.input('PriorBoxVar')) > 0:
if axis == 1:
raise Exception(
"The op box_coder has variable do not support aixs broadcast")
prior_variance_var = block.var(op.input('PriorBoxVar')[0])
axes = []
var_split_inputs_name = [result_name + "@variance_split"]
node = onnx.helper.make_node(
'Transpose',
inputs=op.input('PriorBoxVar'),
outputs=var_split_inputs_name)
node_list.append(node)
var_split_axis = 0
else:
variances = [1.0, 1.0, 1.0, 1.0]
if 'variance' in op.attr and len(op.attr('variance')) > 0:
variances = [float(var) for var in op.attr('variance')]
node_variance_create = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_variance_raw,
value=onnx.helper.make_tensor(
name=name_variance_raw[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=[len(variances)],
vals=variances))
node_list.append(node_variance_create)
var_split_axis = 0
var_split_inputs_name = name_variance_raw
# decode the result
outputs_split_variance = [
result_name + "@variance_split" + str(i) for i in range(0, 4)
]
outputs_split_targebox = [
result_name + "@targebox_split" + str(i) for i in range(0, 4)
]
node_split_var = onnx.helper.make_node(
'Split',
inputs=var_split_inputs_name,
outputs=outputs_split_variance,
axis=var_split_axis)
node_split_target = onnx.helper.make_node(
'Split',
inputs=op.input('TargetBox'),
outputs=outputs_split_targebox,
axis=2)
node_list.extend([node_split_var, node_split_target])
outputs_squeeze_targebox = [
result_name + "@targebox_squeeze" + str(i) for i in range(0, 4)
]
for (input_name, output_name) in zip(outputs_split_targebox,
outputs_squeeze_targebox):
node = onnx.helper.make_node(
'Squeeze', inputs=[input_name], outputs=[output_name], axes=[2])
node_list.append(node)
output_shape_step1 = list(t_size)[:-1]
inputs_tb_step1 = [
[outputs_squeeze_targebox[0], outputs_split_variance[0]],
[outputs_squeeze_targebox[1], outputs_split_variance[1]],
[outputs_squeeze_targebox[2], outputs_split_variance[2]],
[outputs_squeeze_targebox[3], outputs_split_variance[3]]
]
outputs_tb_step1 = [[result_name + "@decode_x_step1"],
[result_name + "@decode_y_step1"],
[result_name + "@decode_w_step1"],
[result_name + "@decode_h_step1"]]
for input_step1, output_step_1 in zip(inputs_tb_step1, outputs_tb_step1):
node = onnx.helper.make_node(
'Mul', inputs=input_step1, outputs=output_step_1)
node_list.append(node)
if axis == 0:
inputs_tbxy_step2 = [
[outputs_tb_step1[0][0], outputs_transpose_pb[0][0]],
[outputs_tb_step1[1][0], outputs_transpose_pb[1][0]]
]
else:
inputs_tbxy_step2 = [
[outputs_tb_step1[0][0], inputs_transpose_pb[0][0]],
[outputs_tb_step1[1][0], inputs_transpose_pb[1][0]]
]
outputs_tbxy_step2 = [[result_name + "@decode_x_step2"],
[result_name + "@decode_y_step2"]]
for input_step2, output_step_2 in zip(inputs_tbxy_step2,
outputs_tbxy_step2):
node = onnx.helper.make_node(
'Mul', inputs=input_step2, outputs=output_step_2)
node_list.append(node)
if axis == 0:
inputs_tbxy_step3 = [
[outputs_tbxy_step2[0][0], outputs_transpose_pb[2][0]],
[outputs_tbxy_step2[1][0], outputs_transpose_pb[3][0]]
]
else:
inputs_tbxy_step3 = [
[outputs_tbxy_step2[0][0], inputs_transpose_pb[2][0]],
[outputs_tbxy_step2[1][0], inputs_transpose_pb[3][0]]
]
outputs_tbxy_step3 = [[result_name + "@decode_x_step3"],
[result_name + "@decode_y_step3"]]
for input_step3, output_step_3 in zip(inputs_tbxy_step3,
outputs_tbxy_step3):
node = onnx.helper.make_node(
'Add', inputs=input_step3, outputs=output_step_3)
node_list.append(node)
# deal with width & height
inputs_tbwh_step2 = [outputs_tb_step1[2], outputs_tb_step1[3]]
outputs_tbwh_step2 = [[result_name + "@decode_w_step2"],
[result_name + "@decode_h_step2"]]
for input_name, output_name in zip(inputs_tbwh_step2, outputs_tbwh_step2):
node = onnx.helper.make_node(
'Exp', inputs=input_name, outputs=output_name)
node_list.append(node)
if axis == 0:
inputs_tbwh_step3 = [
[outputs_tbwh_step2[0][0], outputs_transpose_pb[0][0]],
[outputs_tbwh_step2[1][0], outputs_transpose_pb[1][0]]
]
else:
inputs_tbwh_step3 = [
[outputs_tbwh_step2[0][0], inputs_transpose_pb[0][0]],
[outputs_tbwh_step2[1][0], inputs_transpose_pb[1][0]]
]
outputs_tbwh_step3 = [[result_name + "@decode_w_step3"],
[result_name + "@decode_h_step3"]]
for input_name, output_name in zip(inputs_tbwh_step3, outputs_tbwh_step3):
node = onnx.helper.make_node(
'Mul', inputs=input_name, outputs=output_name)
node_list.append(node)
# final step to calc the result, and concat the result to output
# return the output box, [(x1, y1), (x2, y2)]
inputs_half_tbwh_step4 = [
[outputs_tbwh_step3[0][0], result_name + "@slice_2"],
[outputs_tbwh_step3[1][0], result_name + "@slice_2"]
]
outputs_half_tbwh_step4 = [[result_name + "@decode_half_w_step4"],
[result_name + "@decode_half_h_step4"]]
for inputs_name, outputs_name in zip(inputs_half_tbwh_step4,
outputs_half_tbwh_step4):
node = onnx.helper.make_node(
'Div', inputs=inputs_name, outputs=outputs_name)
node_list.append(node)
inputs_output_point1 = [
[outputs_tbxy_step3[0][0], outputs_half_tbwh_step4[0][0]],
[outputs_tbxy_step3[1][0], outputs_half_tbwh_step4[1][0]]
]
outputs_output_point1 = [[result_name + "@ouput_x1"],
[result_name + "@output_y1"]]
for input_name, output_name in zip(inputs_output_point1,
outputs_output_point1):
node = onnx.helper.make_node(
'Sub', inputs=input_name, outputs=output_name)
node_list.append(node)
inputs_output_point2 = [
[outputs_tbxy_step3[0][0], outputs_half_tbwh_step4[0][0]],
[outputs_tbxy_step3[1][0], outputs_half_tbwh_step4[1][0]]
]
outputs_output_point2 = [[result_name + "@ouput_x2"],
[result_name + "@output_y2"]]
for input_name, output_name in zip(inputs_output_point2,
outputs_output_point2):
node = onnx.helper.make_node(
'Add', inputs=input_name, outputs=output_name)
node_list.append(node)
if not norm:
inputs_unnorm_point2 = [
[outputs_output_point2[0][0], result_name + "@slice_1"],
[outputs_output_point2[1][0], result_name + "@slice_1"]
]
outputs_unnorm_point2 = [[result_name + "@ouput_unnorm_x2"],
[result_name + "@ouput_unnorm_y2"]]
for input_name, output_name in zip(inputs_unnorm_point2,
outputs_unnorm_point2):
node = onnx.helper.make_node(
'Sub', inputs=input_name, outputs=output_name)
node_list.append(node)
outputs_output_point2 = outputs_unnorm_point2
outputs_output_point1.extend(outputs_output_point2)
ouputs_points_unsqueeze = [[result_name + "@points_unsqueeze_x1"],
[result_name + "points_unsqueeze_y1"],
[result_name + "points_unsqueeze_x2"],
[result_name + "points_unsqueeze_y2"]]
for input_name, output_name in zip(outputs_output_point1,
ouputs_points_unsqueeze):
node = onnx.helper.make_node(
'Unsqueeze',
inputs=input_name,
outputs=output_name,
axes=[len(output_shape_step1)])
node_list.append(node)
outputs_points_unsqueeze_list = [
output[0] for output in ouputs_points_unsqueeze
]
node_point_final = onnx.helper.make_node(
'Concat',
inputs=outputs_points_unsqueeze_list,
outputs=op.output('OutputBox'),
axis=len(output_shape_step1))
node_list.append(node_point_final)
return node_list
x2paddle/x2paddle/op_mapper/paddle2onnx/opset9/paddle_custom_layer/im2sequence.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import onnx
import numpy as np
from onnx import onnx_pb, helper
im2seq_counter = 0
def im2sequence(op, block):
global im2sequence_counter
n, c, h, w = block.var(op.input('X')[0]).shape
assert h > 0 and w > 0, "Only supported fixed input shape for im2sequence operator."
stride_h, stride_w = op.attr('strides')
paddings = op.attr('paddings')
assert op.attr(
'out_stride'
) != 1, "Only out_stride==1 is supported for im2sequence operator."
h = h + paddings[0] + paddings[1]
w = w + paddings[1] + paddings[2]
kernel_h, kernel_w = op.attr('kernels')
out_h = 1 + (h - kernel_h + stride_h - 1) // stride_h
out_w = 1 + (w - kernel_w + stride_w - 1) // stride_w
h_steps = list()
for i in range(out_h):
h_steps.append([i * stride_h, i * stride_h + kernel_h])
w_steps = list()
for i in range(out_w):
w_steps.append([i * stride_w, i * stride_w + kernel_w])
nodes = list()
slice_blocks = list()
for i in range(out_h):
for j in range(out_w):
starts_name = "im2sequence.starts.{}.{}.{}".format(im2seq_counter,
i, j)
starts_tensor = helper.make_tensor(
name=starts_name,
data_type=onnx_pb.TensorProto.INT64,
dims=[4],
vals=[0, 0, h_steps[i][0], w_steps[j][0]])
ends_name = "im2sequence.ends.{}.{}.{}".format(im2seq_counter, i, j)
ends_tensor = helper.make_tensor(
name=ends_name,
data_type=onnx_pb.TensorProto.INT64,
dims=[4],
vals=[999999, 999999, h_steps[i][1], w_steps[j][1]])
starts_node = helper.make_node(
'Constant',
inputs=[],
outputs=[starts_name],
value=starts_tensor)
ends_node = helper.make_node(
'Constant', inputs=[], outputs=[ends_name], value=ends_tensor)
nodes.extend([starts_node, ends_node])
slice_block_name = "im2sequence.slice.{}.{}.{}".format(
im2seq_counter, i, j)
slice_block_node = helper.make_node(
'Slice',
inputs=[op.input('X')[0], starts_name, ends_name],
outputs=[slice_block_name])
flatten_block_name = "im2sequence.flatten.{}.{}.{}".format(
im2seq_counter, i, j)
flatten_block_node = helper.make_node(
"Flatten",
inputs=[slice_block_name],
outputs=[flatten_block_name],
axis=0)
nodes.extend([slice_block_node, flatten_block_node])
slice_blocks.append(flatten_block_name)
concat_block_name = "im2sequence.concat_block.{}".format(im2seq_counter)
# concat_block_node = helper.make_node("Concat", inputs=slice_blocks, outputs=[concat_block_name], axis=0)
concat_block_node = helper.make_node(
"Concat", inputs=slice_blocks, outputs=op.output('Out'), axis=0)
nodes.append(concat_block_node)
print("\n\n==========Importance Notice===========")
print(
"Since im2sequence operator is used in your paddlepaddle model, the translated onnx model only support input data with batch_size=1."
)
print("======================================\n")
return nodes
x2paddle/x2paddle/op_mapper/paddle2onnx/opset9/paddle_custom_layer/multiclass_nms.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
import sys
import os
import numpy as np
import paddle.fluid.core as core
import paddle.fluid as fluid
import onnx
import logging
from onnx import helper, onnx_pb
def multiclass_nms(op, block):
"""
Convert the paddle multiclass_nms to onnx op.
This op is get the select boxes from origin boxes.
"""
inputs = dict()
outputs = dict()
attrs = dict()
for name in op.input_names:
inputs[name] = op.input(name)
for name in op.output_names:
outputs[name] = op.output(name)
for name in op.attr_names:
attrs[name] = op.attr(name)
result_name = outputs['Out'][0]
background = attrs['background_label']
normalized = attrs['normalized']
if normalized == False:
logging.warn(
"The parameter normalized of multiclass_nms OP of Paddle is False, which has diff with ONNX." \
" Please set normalized=True in multiclass_nms of Paddle, see doc Q4 in https://github.com/PaddlePaddle/X2Paddle/blob/develop/FAQ.md")
#convert the paddle attribute to onnx tensor
name_score_threshold = [outputs['Out'][0] + "@score_threshold"]
name_iou_threshold = [outputs['Out'][0] + "@iou_threshold"]
name_keep_top_k = [outputs['Out'][0] + '@keep_top_k']
name_keep_top_k_2D = [outputs['Out'][0] + '@keep_top_k_1D']
node_score_threshold = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_score_threshold,
value=onnx.helper.make_tensor(
name=name_score_threshold[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=(),
vals=[float(attrs['score_threshold'])]))
node_iou_threshold = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_iou_threshold,
value=onnx.helper.make_tensor(
name=name_iou_threshold[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=(),
vals=[float(attrs['nms_threshold'])]))
node_keep_top_k = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_keep_top_k,
value=onnx.helper.make_tensor(
name=name_keep_top_k[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=(),
vals=[np.int64(attrs['keep_top_k'])]))
node_keep_top_k_2D = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_keep_top_k_2D,
value=onnx.helper.make_tensor(
name=name_keep_top_k_2D[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=[1, 1],
vals=[np.int64(attrs['keep_top_k'])]))
# the paddle data format is x1,y1,x2,y2
kwargs = {'center_point_box': 0}
name_select_nms = [outputs['Out'][0] + "@select_index"]
node_select_nms= onnx.helper.make_node(
'NonMaxSuppression',
inputs=inputs['BBoxes'] + inputs['Scores'] + name_keep_top_k +\
name_iou_threshold + name_score_threshold,
outputs=name_select_nms)
# step 1 nodes select the nms class
node_list = [
node_score_threshold, node_iou_threshold, node_keep_top_k,
node_keep_top_k_2D, node_select_nms
]
# create some const value to use
name_const_value = [result_name+"@const_0",
result_name+"@const_1",\
result_name+"@const_2",\
result_name+"@const_-1"]
value_const_value = [0, 1, 2, -1]
for name, value in zip(name_const_value, value_const_value):
node = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=[name],
value=onnx.helper.make_tensor(
name=name + "@const",
data_type=onnx.TensorProto.INT64,
dims=[1],
vals=[value]))
node_list.append(node)
# In this code block, we will deocde the raw score data, reshape N * C * M to 1 * N*C*M
# and the same time, decode the select indices to 1 * D, gather the select_indices
outputs_gather_1 = [result_name + "@gather_1"]
node_gather_1 = onnx.helper.make_node(
'Gather',
inputs=name_select_nms + [result_name + "@const_1"],
outputs=outputs_gather_1,
axis=1)
node_list.append(node_gather_1)
outputs_squeeze_gather_1 = [result_name + "@sequeeze_gather_1"]
node_squeeze_gather_1 = onnx.helper.make_node(
'Squeeze',
inputs=outputs_gather_1,
outputs=outputs_squeeze_gather_1,
axes=[1])
node_list.append(node_squeeze_gather_1)
outputs_gather_2 = [result_name + "@gather_2"]
node_gather_2 = onnx.helper.make_node(
'Gather',
inputs=name_select_nms + [result_name + "@const_2"],
outputs=outputs_gather_2,
axis=1)
node_list.append(node_gather_2)
#slice the class is not 0
if background == 0:
outputs_nonzero = [result_name + "@nonzero"]
node_nonzero = onnx.helper.make_node(
'NonZero', inputs=outputs_squeeze_gather_1, outputs=outputs_nonzero)
node_list.append(node_nonzero)
else:
name_thresh = [result_name + "@thresh"]
node_thresh = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_thresh,
value=onnx.helper.make_tensor(
name=name_thresh[0] + "@const",
data_type=onnx.TensorProto.INT32,
dims=[1],
vals=[-1]))
node_list.append(node_thresh)
outputs_cast = [result_name + "@cast"]
node_cast = onnx.helper.make_node(
'Cast', inputs=outputs_squeeze_gather_1, outputs=outputs_cast, to=6)
node_list.append(node_cast)
outputs_greater = [result_name + "@greater"]
node_greater = onnx.helper.make_node(
'Greater',
inputs=outputs_cast + name_thresh,
outputs=outputs_greater)
node_list.append(node_greater)
outputs_nonzero = [result_name + "@nonzero"]
node_nonzero = onnx.helper.make_node(
'NonZero', inputs=outputs_greater, outputs=outputs_nonzero)
node_list.append(node_nonzero)
outputs_gather_1_nonzero = [result_name + "@gather_1_nonzero"]
node_gather_1_nonzero = onnx.helper.make_node(
'Gather',
inputs=outputs_gather_1 + outputs_nonzero,
outputs=outputs_gather_1_nonzero,
axis=0)
node_list.append(node_gather_1_nonzero)
outputs_gather_2_nonzero = [result_name + "@gather_2_nonzero"]
node_gather_2_nonzero = onnx.helper.make_node(
'Gather',
inputs=outputs_gather_2 + outputs_nonzero,
outputs=outputs_gather_2_nonzero,
axis=0)
node_list.append(node_gather_2_nonzero)
# reshape scores N * C * M to (N*C*M) * 1
outputs_reshape_scores_rank1 = [result_name + "@reshape_scores_rank1"]
node_reshape_scores_rank1 = onnx.helper.make_node(
"Reshape",
inputs=inputs['Scores'] + [result_name + "@const_-1"],
outputs=outputs_reshape_scores_rank1)
node_list.append(node_reshape_scores_rank1)
# get the shape of scores
outputs_shape_scores = [result_name + "@shape_scores"]
node_shape_scores = onnx.helper.make_node(
'Shape', inputs=inputs['Scores'], outputs=outputs_shape_scores)
node_list.append(node_shape_scores)
# gather the index: 2 shape of scores
outputs_gather_scores_dim1 = [result_name + "@gather_scores_dim1"]
node_gather_scores_dim1 = onnx.helper.make_node(
'Gather',
inputs=outputs_shape_scores + [result_name + "@const_2"],
outputs=outputs_gather_scores_dim1,
axis=0)
node_list.append(node_gather_scores_dim1)
# mul class * M
outputs_mul_classnum_boxnum = [result_name + "@mul_classnum_boxnum"]
node_mul_classnum_boxnum = onnx.helper.make_node(
'Mul',
inputs=outputs_gather_1_nonzero + outputs_gather_scores_dim1,
outputs=outputs_mul_classnum_boxnum)
node_list.append(node_mul_classnum_boxnum)
# add class * M * index
outputs_add_class_M_index = [result_name + "@add_class_M_index"]
node_add_class_M_index = onnx.helper.make_node(
'Add',
inputs=outputs_mul_classnum_boxnum + outputs_gather_2_nonzero,
outputs=outputs_add_class_M_index)
node_list.append(node_add_class_M_index)
# Squeeze the indices to 1 dim
outputs_squeeze_select_index = [result_name + "@squeeze_select_index"]
node_squeeze_select_index = onnx.helper.make_node(
'Squeeze',
inputs=outputs_add_class_M_index,
outputs=outputs_squeeze_select_index,
axes=[0, 2])
node_list.append(node_squeeze_select_index)
# gather the data from flatten scores
outputs_gather_select_scores = [result_name + "@gather_select_scores"]
node_gather_select_scores = onnx.helper.make_node('Gather',
inputs=outputs_reshape_scores_rank1 + \
outputs_squeeze_select_index,
outputs=outputs_gather_select_scores,
axis=0)
node_list.append(node_gather_select_scores)
# get nums to input TopK
outputs_shape_select_num = [result_name + "@shape_select_num"]
node_shape_select_num = onnx.helper.make_node(
'Shape',
inputs=outputs_gather_select_scores,
outputs=outputs_shape_select_num)
node_list.append(node_shape_select_num)
outputs_gather_select_num = [result_name + "@gather_select_num"]
node_gather_select_num = onnx.helper.make_node(
'Gather',
inputs=outputs_shape_select_num + [result_name + "@const_0"],
outputs=outputs_gather_select_num,
axis=0)
node_list.append(node_gather_select_num)
outputs_unsqueeze_select_num = [result_name + "@unsqueeze_select_num"]
node_unsqueeze_select_num = onnx.helper.make_node(
'Unsqueeze',
inputs=outputs_gather_select_num,
outputs=outputs_unsqueeze_select_num,
axes=[0])
node_list.append(node_unsqueeze_select_num)
outputs_concat_topK_select_num = [result_name + "@conat_topK_select_num"]
node_conat_topK_select_num = onnx.helper.make_node(
'Concat',
inputs=outputs_unsqueeze_select_num + name_keep_top_k_2D,
outputs=outputs_concat_topK_select_num,
axis=0)
node_list.append(node_conat_topK_select_num)
outputs_cast_concat_topK_select_num = [
result_name + "@concat_topK_select_num"
]
node_outputs_cast_concat_topK_select_num = onnx.helper.make_node(
'Cast',
inputs=outputs_concat_topK_select_num,
outputs=outputs_cast_concat_topK_select_num,
to=6)
node_list.append(node_outputs_cast_concat_topK_select_num)
# get min(topK, num_select)
outputs_compare_topk_num_select = [result_name + "@compare_topk_num_select"]
node_compare_topk_num_select = onnx.helper.make_node(
'ReduceMin',
inputs=outputs_cast_concat_topK_select_num,
outputs=outputs_compare_topk_num_select,
keepdims=0)
node_list.append(node_compare_topk_num_select)
# unsqueeze the indices to 1D tensor
outputs_unsqueeze_topk_select_indices = [
result_name + "@unsqueeze_topk_select_indices"
]
node_unsqueeze_topk_select_indices = onnx.helper.make_node(
'Unsqueeze',
inputs=outputs_compare_topk_num_select,
outputs=outputs_unsqueeze_topk_select_indices,
axes=[0])
node_list.append(node_unsqueeze_topk_select_indices)
# cast the indices to INT64
outputs_cast_topk_indices = [result_name + "@cast_topk_indices"]
node_cast_topk_indices = onnx.helper.make_node(
'Cast',
inputs=outputs_unsqueeze_topk_select_indices,
outputs=outputs_cast_topk_indices,
to=7)
node_list.append(node_cast_topk_indices)
# select topk scores indices
outputs_topk_select_topk_indices = [result_name + "@topk_select_topk_values",\
result_name + "@topk_select_topk_indices"]
node_topk_select_topk_indices = onnx.helper.make_node(
'TopK',
inputs=outputs_gather_select_scores + outputs_cast_topk_indices,
outputs=outputs_topk_select_topk_indices)
node_list.append(node_topk_select_topk_indices)
# gather topk label, scores, boxes
outputs_gather_topk_scores = [result_name + "@gather_topk_scores"]
node_gather_topk_scores = onnx.helper.make_node(
'Gather',
inputs=outputs_gather_select_scores +
[outputs_topk_select_topk_indices[1]],
outputs=outputs_gather_topk_scores,
axis=0)
node_list.append(node_gather_topk_scores)
outputs_gather_topk_class = [result_name + "@gather_topk_class"]
node_gather_topk_class = onnx.helper.make_node(
'Gather',
inputs=outputs_gather_1_nonzero +
[outputs_topk_select_topk_indices[1]],
outputs=outputs_gather_topk_class,
axis=1)
node_list.append(node_gather_topk_class)
# gather the boxes need to gather the boxes id, then get boxes
outputs_gather_topk_boxes_id = [result_name + "@gather_topk_boxes_id"]
node_gather_topk_boxes_id = onnx.helper.make_node(
'Gather',
inputs=outputs_gather_2_nonzero +
[outputs_topk_select_topk_indices[1]],
outputs=outputs_gather_topk_boxes_id,
axis=1)
node_list.append(node_gather_topk_boxes_id)
# squeeze the gather_topk_boxes_id to 1 dim
outputs_squeeze_topk_boxes_id = [result_name + "@squeeze_topk_boxes_id"]
node_squeeze_topk_boxes_id = onnx.helper.make_node(
'Squeeze',
inputs=outputs_gather_topk_boxes_id,
outputs=outputs_squeeze_topk_boxes_id,
axes=[0, 2])
node_list.append(node_squeeze_topk_boxes_id)
outputs_gather_select_boxes = [result_name + "@gather_select_boxes"]
node_gather_select_boxes = onnx.helper.make_node(
'Gather',
inputs=inputs['BBoxes'] + outputs_squeeze_topk_boxes_id,
outputs=outputs_gather_select_boxes,
axis=1)
node_list.append(node_gather_select_boxes)
# concat the final result
# before concat need to cast the class to float
outputs_cast_topk_class = [result_name + "@cast_topk_class"]
node_cast_topk_class = onnx.helper.make_node(
'Cast',
inputs=outputs_gather_topk_class,
outputs=outputs_cast_topk_class,
to=1)
node_list.append(node_cast_topk_class)
outputs_unsqueeze_topk_scores = [result_name + "@unsqueeze_topk_scores"]
node_unsqueeze_topk_scores = onnx.helper.make_node(
'Unsqueeze',
inputs=outputs_gather_topk_scores,
outputs=outputs_unsqueeze_topk_scores,
axes=[0, 2])
node_list.append(node_unsqueeze_topk_scores)
inputs_concat_final_results = outputs_cast_topk_class + outputs_unsqueeze_topk_scores +\
outputs_gather_select_boxes
outputs_sort_by_socre_results = [result_name + "@concat_topk_scores"]
node_sort_by_socre_results = onnx.helper.make_node(
'Concat',
inputs=inputs_concat_final_results,
outputs=outputs_sort_by_socre_results,
axis=2)
node_list.append(node_sort_by_socre_results)
# select topk classes indices
outputs_squeeze_cast_topk_class = [result_name + "@squeeze_cast_topk_class"]
node_squeeze_cast_topk_class = onnx.helper.make_node(
'Squeeze',
inputs=outputs_cast_topk_class,
outputs=outputs_squeeze_cast_topk_class,
axes=[0, 2])
node_list.append(node_squeeze_cast_topk_class)
outputs_neg_squeeze_cast_topk_class = [
result_name + "@neg_squeeze_cast_topk_class"
]
node_neg_squeeze_cast_topk_class = onnx.helper.make_node(
'Neg',
inputs=outputs_squeeze_cast_topk_class,
outputs=outputs_neg_squeeze_cast_topk_class)
node_list.append(node_neg_squeeze_cast_topk_class)
outputs_topk_select_classes_indices = [result_name + "@topk_select_topk_classes_scores",\
result_name + "@topk_select_topk_classes_indices"]
node_topk_select_topk_indices = onnx.helper.make_node(
'TopK',
inputs=outputs_neg_squeeze_cast_topk_class + outputs_cast_topk_indices,
outputs=outputs_topk_select_classes_indices)
node_list.append(node_topk_select_topk_indices)
outputs_concat_final_results = outputs['Out']
node_concat_final_results = onnx.helper.make_node(
'Gather',
inputs=outputs_sort_by_socre_results +
[outputs_topk_select_classes_indices[1]],
outputs=outputs_concat_final_results,
axis=1)
node_list.append(node_concat_final_results)
return node_list
x2paddle/x2paddle/op_mapper/paddle2onnx/opset9/paddle_custom_layer/prior_box.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import sys
import math
import onnx
import warnings
import numpy as np
from functools import partial
from onnx import TensorProto
from onnx.helper import make_node, make_tensor
from onnx import onnx_pb
from paddle.fluid.executor import _fetch_var as fetch_var
from onnx import helper
import paddle.fluid as fluid
import paddle.fluid.core as core
def ExpandAspectRations(input_aspect_ratior, flip):
expsilon = 1e-6
output_ratios = [1.0]
for input_ratio in input_aspect_ratior:
already_exis = False
for output_ratio in output_ratios:
if abs(input_ratio - output_ratio) < expsilon:
already_exis = True
break
if already_exis == False:
output_ratios.append(input_ratio)
if flip:
output_ratios.append(1.0 / input_ratio)
return output_ratios
def prior_box(op, block):
"""
In this function, use the attribute to get the prior box, because we do not use
the image data and feature map, wo could the python code to create the varaible,
and to create the onnx tensor as output.
"""
flip = bool(op.attr('flip'))
clip = bool(op.attr('clip'))
min_max_aspect_ratios_order = bool(op.attr('min_max_aspect_ratios_order'))
min_sizes = [float(size) for size in op.attr('min_sizes')]
max_sizes = [float(size) for size in op.attr('max_sizes')]
if isinstance(op.attr('aspect_ratios'), list):
aspect_ratios = [float(ratio) for ratio in op.attr('aspect_ratios')]
else:
aspect_ratios = [float(op.attr('aspect_ratios'))]
variances = [float(var) for var in op.attr('variances')]
# set min_max_aspect_ratios_order = false
output_ratios = ExpandAspectRations(aspect_ratios, flip)
step_w = float(op.attr('step_w'))
step_h = float(op.attr('step_h'))
offset = float(op.attr('offset'))
input_shape = block.var(op.input('Input')[0]).shape
image_shape = block.var(op.input('Image')[0]).shape
img_width = image_shape[3]
img_height = image_shape[2]
feature_width = input_shape[3]
feature_height = input_shape[2]
step_width = 1.0
step_height = 1.0
if step_w == 0.0 or step_h == 0.0:
step_w = float(img_width / feature_width)
step_h = float(img_height / feature_height)
num_priors = len(output_ratios) * len(min_sizes)
if len(max_sizes) > 0:
num_priors += len(max_sizes)
out_dim = (feature_height, feature_width, num_priors, 4)
out_boxes = np.zeros(out_dim).astype('float32')
out_var = np.zeros(out_dim).astype('float32')
idx = 0
for h in range(feature_height):
for w in range(feature_width):
c_x = (w + offset) * step_w
c_y = (h + offset) * step_h
idx = 0
for s in range(len(min_sizes)):
min_size = min_sizes[s]
if not min_max_aspect_ratios_order:
# rest of priors
for r in range(len(output_ratios)):
ar = output_ratios[r]
c_w = min_size * math.sqrt(ar) / 2
c_h = (min_size / math.sqrt(ar)) / 2
out_boxes[h, w, idx, :] = [
(c_x - c_w) / img_width, (c_y - c_h) / img_height,
(c_x + c_w) / img_width, (c_y + c_h) / img_height
]
idx += 1
if len(max_sizes) > 0:
max_size = max_sizes[s]
# second prior: aspect_ratio = 1,
c_w = c_h = math.sqrt(min_size * max_size) / 2
out_boxes[h, w, idx, :] = [
(c_x - c_w) / img_width, (c_y - c_h) / img_height,
(c_x + c_w) / img_width, (c_y + c_h) / img_height
]
idx += 1
else:
c_w = c_h = min_size / 2.
out_boxes[h, w, idx, :] = [
(c_x - c_w) / img_width, (c_y - c_h) / img_height,
(c_x + c_w) / img_width, (c_y + c_h) / img_height
]
idx += 1
if len(max_sizes) > 0:
max_size = max_sizes[s]
# second prior: aspect_ratio = 1,
c_w = c_h = math.sqrt(min_size * max_size) / 2
out_boxes[h, w, idx, :] = [
(c_x - c_w) / img_width, (c_y - c_h) / img_height,
(c_x + c_w) / img_width, (c_y + c_h) / img_height
]
idx += 1
# rest of priors
for r in range(len(output_ratios)):
ar = output_ratios[r]
if abs(ar - 1.) < 1e-6:
continue
c_w = min_size * math.sqrt(ar) / 2
c_h = (min_size / math.sqrt(ar)) / 2
out_boxes[h, w, idx, :] = [
(c_x - c_w) / img_width, (c_y - c_h) / img_height,
(c_x + c_w) / img_width, (c_y + c_h) / img_height
]
idx += 1
if clip:
out_boxes = np.clip(out_boxes, 0.0, 1.0)
# set the variance.
out_var = np.tile(variances, (feature_height, feature_width, num_priors, 1))
#make node that
node_boxes = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=op.output('Boxes'),
value=onnx.helper.make_tensor(
name=op.output('Boxes')[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=out_boxes.shape,
vals=out_boxes.flatten()))
node_vars = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=op.output('Variances'),
value=onnx.helper.make_tensor(
name=op.output('Variances')[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=out_var.shape,
vals=out_var.flatten()))
return [node_boxes, node_vars]
x2paddle/x2paddle/op_mapper/paddle2onnx/opset9/paddle_custom_layer/yolo_box.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import onnx
import numpy as np
from onnx import onnx_pb, helper
MAX_FLOAT32 = np.asarray(
[255, 255, 127, 127], dtype=np.uint8).view(np.float32)[0]
def get_old_name(arg, name_prefix=''):
prefix_index = arg.find(name_prefix)
if prefix_index != -1:
last_prefix = arg[len(name_prefix):]
else:
last_prefix = arg
idx = last_prefix.find('@')
if idx != -1:
last_prefix = last_prefix[:idx]
return name_prefix + last_prefix
def is_static_shape(shape):
if len(shape) > 1 and shape.count(-1) > 1:
raise Exception(
"Converting this model to ONNX need with static input shape, please fix input shape of this model, see doc Q5 in https://github.com/PaddlePaddle/X2Paddle/blob/develop/FAQ.md."
)
def yolo_box(op, block):
inputs = dict()
outputs = dict()
attrs = dict()
for name in op.input_names:
inputs[name] = op.input(name)
for name in op.output_names:
outputs[name] = op.output(name)
for name in op.attr_names:
attrs[name] = op.attr(name)
model_name = outputs['Boxes'][0]
input_shape = block.vars[get_old_name(inputs['X'][0])].shape
is_static_shape(input_shape)
image_size = inputs['ImgSize']
input_height = input_shape[2]
input_width = input_shape[3]
class_num = attrs['class_num']
anchors = attrs['anchors']
num_anchors = int(len(anchors)) // 2
downsample_ratio = attrs['downsample_ratio']
input_size = input_height * downsample_ratio
conf_thresh = attrs['conf_thresh']
conf_thresh_mat = np.ones([num_anchors * input_height *
input_width]) * conf_thresh
node_list = []
im_outputs = []
x_shape = [1, num_anchors, 5 + class_num, input_height, input_width]
name_x_shape = [model_name + "@x_shape"]
node_x_shape = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_x_shape,
value=onnx.helper.make_tensor(
name=name_x_shape[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=[5],
vals=x_shape))
node_list.append(node_x_shape)
outputs_x_reshape = [model_name + "@reshape"]
node_x_reshape = onnx.helper.make_node(
'Reshape', inputs=inputs['X'] + name_x_shape, outputs=outputs_x_reshape)
node_list.append(node_x_reshape)
outputs_x_transpose = [model_name + "@x_transpose"]
node_x_transpose = onnx.helper.make_node(
'Transpose',
inputs=outputs_x_reshape,
outputs=outputs_x_transpose,
perm=[0, 1, 3, 4, 2])
node_list.append(node_x_transpose)
range_x = []
range_y = []
for i in range(0, input_width):
range_x.append(i)
for j in range(0, input_height):
range_y.append(j)
name_range_x = [model_name + "@range_x"]
node_range_x = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_range_x,
value=onnx.helper.make_tensor(
name=name_range_x[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=[input_width],
vals=range_x))
node_list.append(node_range_x)
name_range_y = [model_name + "@range_y"]
node_range_y = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_range_y,
value=onnx.helper.make_tensor(
name=name_range_y[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=[input_height],
vals=range_y))
node_list.append(node_range_y)
range_x_new_shape = [1, input_width]
range_y_new_shape = [input_height, 1]
name_range_x_new_shape = [model_name + "@range_x_new_shape"]
node_range_x_new_shape = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_range_x_new_shape,
value=onnx.helper.make_tensor(
name=name_range_x_new_shape[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=[len(range_x_new_shape)],
vals=range_x_new_shape))
node_list.append(node_range_x_new_shape)
name_range_y_new_shape = [model_name + "@range_y_new_shape"]
node_range_y_new_shape = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_range_y_new_shape,
value=onnx.helper.make_tensor(
name=name_range_y_new_shape[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=[len(range_y_new_shape)],
vals=range_y_new_shape))
node_list.append(node_range_y_new_shape)
outputs_range_x_reshape = [model_name + "@range_x_reshape"]
node_range_x_reshape = onnx.helper.make_node(
'Reshape',
inputs=name_range_x + name_range_x_new_shape,
outputs=outputs_range_x_reshape)
node_list.append(node_range_x_reshape)
outputs_range_y_reshape = [model_name + "@range_y_reshape"]
node_range_y_reshape = onnx.helper.make_node(
'Reshape',
inputs=name_range_y + name_range_y_new_shape,
outputs=outputs_range_y_reshape)
node_list.append(node_range_y_reshape)
outputs_grid_x = [model_name + "@grid_x"]
node_grid_x = onnx.helper.make_node(
"Tile",
inputs=outputs_range_x_reshape + name_range_y_new_shape,
outputs=outputs_grid_x)
node_list.append(node_grid_x)
outputs_grid_y = [model_name + "@grid_y"]
node_grid_y = onnx.helper.make_node(
"Tile",
inputs=outputs_range_y_reshape + name_range_x_new_shape,
outputs=outputs_grid_y)
node_list.append(node_grid_y)
outputs_box_x = [model_name + "@box_x"]
outputs_box_y = [model_name + "@box_y"]
outputs_box_w = [model_name + "@box_w"]
outputs_box_h = [model_name + "@box_h"]
outputs_conf = [model_name + "@conf"]
outputs_prob = [model_name + "@prob"]
node_split_input = onnx.helper.make_node(
"Split",
inputs=outputs_x_transpose,
outputs=outputs_box_x + outputs_box_y + outputs_box_w\
+ outputs_box_h + outputs_conf + outputs_prob,
axis=-1,
split=[1, 1, 1, 1, 1, class_num])
node_list.append(node_split_input)
outputs_box_x_sigmoid = [model_name + "@box_x_sigmoid"]
outputs_box_y_sigmoid = [model_name + "@box_y_sigmoid"]
node_box_x_sigmoid = onnx.helper.make_node(
"Sigmoid", inputs=outputs_box_x, outputs=outputs_box_x_sigmoid)
node_list.append(node_box_x_sigmoid)
node_box_y_sigmoid = onnx.helper.make_node(
"Sigmoid", inputs=outputs_box_y, outputs=outputs_box_y_sigmoid)
node_list.append(node_box_y_sigmoid)
outputs_box_x_squeeze = [model_name + "@box_x_squeeze"]
outputs_box_y_squeeze = [model_name + "@box_y_squeeze"]
node_box_x_squeeze = onnx.helper.make_node(
'Squeeze',
inputs=outputs_box_x_sigmoid,
outputs=outputs_box_x_squeeze,
axes=[4])
node_list.append(node_box_x_squeeze)
node_box_y_squeeze = onnx.helper.make_node(
'Squeeze',
inputs=outputs_box_y_sigmoid,
outputs=outputs_box_y_squeeze,
axes=[4])
node_list.append(node_box_y_squeeze)
outputs_box_x_add_grid = [model_name + "@box_x_add_grid"]
outputs_box_y_add_grid = [model_name + "@box_y_add_grid"]
node_box_x_add_grid = onnx.helper.make_node(
"Add",
inputs=outputs_grid_x + outputs_box_x_squeeze,
outputs=outputs_box_x_add_grid)
node_list.append(node_box_x_add_grid)
node_box_y_add_grid = onnx.helper.make_node(
"Add",
inputs=outputs_grid_y + outputs_box_y_squeeze,
outputs=outputs_box_y_add_grid)
node_list.append(node_box_y_add_grid)
name_input_h = [model_name + "@input_h"]
name_input_w = [model_name + "@input_w"]
node_input_h = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_input_h,
value=onnx.helper.make_tensor(
name=name_input_w[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=(),
vals=[input_height]))
node_list.append(node_input_h)
node_input_w = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_input_w,
value=onnx.helper.make_tensor(
name=name_input_w[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=(),
vals=[input_width]))
node_list.append(node_input_w)
outputs_box_x_encode = [model_name + "@box_x_encode"]
outputs_box_y_encode = [model_name + "@box_y_encode"]
node_box_x_encode = onnx.helper.make_node(
'Div',
inputs=outputs_box_x_add_grid + name_input_w,
outputs=outputs_box_x_encode)
node_list.append(node_box_x_encode)
node_box_y_encode = onnx.helper.make_node(
'Div',
inputs=outputs_box_y_add_grid + name_input_h,
outputs=outputs_box_y_encode)
node_list.append(node_box_y_encode)
name_anchor_tensor = [model_name + "@anchor_tensor"]
node_anchor_tensor = onnx.helper.make_node(
"Constant",
inputs=[],
outputs=name_anchor_tensor,
value=onnx.helper.make_tensor(
name=name_anchor_tensor[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=[len(anchors)],
vals=anchors))
node_list.append(node_anchor_tensor)
anchor_shape = [int(num_anchors), 2]
name_anchor_shape = [model_name + "@anchor_shape"]
node_anchor_shape = onnx.helper.make_node(
"Constant",
inputs=[],
outputs=name_anchor_shape,
value=onnx.helper.make_tensor(
name=name_anchor_shape[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=[2],
vals=anchor_shape))
node_list.append(node_anchor_shape)
outputs_anchor_tensor_reshape = [model_name + "@anchor_tensor_reshape"]
node_anchor_tensor_reshape = onnx.helper.make_node(
"Reshape",
inputs=name_anchor_tensor + name_anchor_shape,
outputs=outputs_anchor_tensor_reshape)
node_list.append(node_anchor_tensor_reshape)
name_input_size = [model_name + "@input_size"]
node_input_size = onnx.helper.make_node(
"Constant",
inputs=[],
outputs=name_input_size,
value=onnx.helper.make_tensor(
name=name_input_size[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=(),
vals=[input_size]))
node_list.append(node_input_size)
outputs_anchors_div_input_size = [model_name + "@anchors_div_input_size"]
node_anchors_div_input_size = onnx.helper.make_node(
"Div",
inputs=outputs_anchor_tensor_reshape + name_input_size,
outputs=outputs_anchors_div_input_size)
node_list.append(node_anchors_div_input_size)
outputs_anchor_w = [model_name + "@anchor_w"]
outputs_anchor_h = [model_name + "@anchor_h"]
node_anchor_split = onnx.helper.make_node(
'Split',
inputs=outputs_anchors_div_input_size,
outputs=outputs_anchor_w + outputs_anchor_h,
axis=1,
split=[1, 1])
node_list.append(node_anchor_split)
new_anchor_shape = [1, int(num_anchors), 1, 1]
name_new_anchor_shape = [model_name + "@new_anchor_shape"]
node_new_anchor_shape = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_new_anchor_shape,
value=onnx.helper.make_tensor(
name=name_new_anchor_shape[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=[len(new_anchor_shape)],
vals=new_anchor_shape))
node_list.append(node_new_anchor_shape)
outputs_anchor_w_reshape = [model_name + "@anchor_w_reshape"]
outputs_anchor_h_reshape = [model_name + "@anchor_h_reshape"]
node_anchor_w_reshape = onnx.helper.make_node(
'Reshape',
inputs=outputs_anchor_w + name_new_anchor_shape,
outputs=outputs_anchor_w_reshape)
node_list.append(node_anchor_w_reshape)
node_anchor_h_reshape = onnx.helper.make_node(
'Reshape',
inputs=outputs_anchor_h + name_new_anchor_shape,
outputs=outputs_anchor_h_reshape)
node_list.append(node_anchor_h_reshape)
outputs_box_w_squeeze = [model_name + "@box_w_squeeze"]
node_box_w_squeeze = onnx.helper.make_node(
'Squeeze',
inputs=outputs_box_w,
outputs=outputs_box_w_squeeze,
axes=[4])
node_list.append(node_box_w_squeeze)
outputs_box_h_squeeze = [model_name + "@box_h_squeeze"]
node_box_h_squeeze = onnx.helper.make_node(
'Squeeze',
inputs=outputs_box_h,
outputs=outputs_box_h_squeeze,
axes=[4])
node_list.append(node_box_h_squeeze)
outputs_box_w_exp = [model_name + "@box_w_exp"]
node_box_w_exp = onnx.helper.make_node(
"Exp", inputs=outputs_box_w_squeeze, outputs=outputs_box_w_exp)
node_list.append(node_box_w_exp)
outputs_box_h_exp = [model_name + "@box_h_exp"]
node_box_h_exp = onnx.helper.make_node(
"Exp", inputs=outputs_box_h_squeeze, outputs=outputs_box_h_exp)
node_list.append(node_box_h_exp)
outputs_box_w_encode = [model_name + "box_w_encode"]
outputs_box_h_encode = [model_name + "box_h_encode"]
node_box_w_encode = onnx.helper.make_node(
'Mul',
inputs=outputs_box_w_exp + outputs_anchor_w_reshape,
outputs=outputs_box_w_encode)
node_list.append(node_box_w_encode)
node_box_h_encode = onnx.helper.make_node(
'Mul',
inputs=outputs_box_h_exp + outputs_anchor_h_reshape,
outputs=outputs_box_h_encode)
node_list.append(node_box_h_encode)
outputs_conf_sigmoid = [model_name + "@conf_sigmoid"]
node_conf_sigmoid = onnx.helper.make_node(
'Sigmoid', inputs=outputs_conf, outputs=outputs_conf_sigmoid)
node_list.append(node_conf_sigmoid)
name_conf_thresh = [model_name + "@conf_thresh"]
node_conf_thresh = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_conf_thresh,
value=onnx.helper.make_tensor(
name=name_conf_thresh[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=[num_anchors * input_height * input_width],
vals=conf_thresh_mat))
node_list.append(node_conf_thresh)
conf_shape = [1, int(num_anchors), input_height, input_width, 1]
name_conf_shape = [model_name + "@conf_shape"]
node_conf_shape = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_conf_shape,
value=onnx.helper.make_tensor(
name=name_conf_shape[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=[len(conf_shape)],
vals=conf_shape))
node_list.append(node_conf_shape)
outputs_conf_thresh_reshape = [model_name + "@conf_thresh_reshape"]
node_conf_thresh_reshape = onnx.helper.make_node(
'Reshape',
inputs=name_conf_thresh + name_conf_shape,
outputs=outputs_conf_thresh_reshape)
node_list.append(node_conf_thresh_reshape)
outputs_conf_sub = [model_name + "@conf_sub"]
node_conf_sub = onnx.helper.make_node(
'Sub',
inputs=outputs_conf_sigmoid + outputs_conf_thresh_reshape,
outputs=outputs_conf_sub)
node_list.append(node_conf_sub)
outputs_conf_clip = [model_name + "@conf_clip"]
node_conf_clip = onnx.helper.make_node(
'Clip', inputs=outputs_conf_sub, outputs=outputs_conf_clip)
node_list.append(node_conf_clip)
zeros = [0]
name_zeros = [model_name + "@zeros"]
node_zeros = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_zeros,
value=onnx.helper.make_tensor(
name=name_zeros[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=(),
vals=zeros))
node_list.append(node_zeros)
outputs_conf_clip_bool = [model_name + "@conf_clip_bool"]
node_conf_clip_bool = onnx.helper.make_node(
'Greater',
inputs=outputs_conf_clip + name_zeros,
outputs=outputs_conf_clip_bool)
node_list.append(node_conf_clip_bool)
outputs_conf_clip_cast = [model_name + "@conf_clip_cast"]
node_conf_clip_cast = onnx.helper.make_node(
'Cast',
inputs=outputs_conf_clip_bool,
outputs=outputs_conf_clip_cast,
to=1)
node_list.append(node_conf_clip_cast)
outputs_conf_set_zero = [model_name + "@conf_set_zero"]
node_conf_set_zero = onnx.helper.make_node(
'Mul',
inputs=outputs_conf_sigmoid + outputs_conf_clip_cast,
outputs=outputs_conf_set_zero)
node_list.append(node_conf_set_zero)
outputs_prob_sigmoid = [model_name + "@prob_sigmoid"]
node_prob_sigmoid = onnx.helper.make_node(
'Sigmoid', inputs=outputs_prob, outputs=outputs_prob_sigmoid)
node_list.append(node_prob_sigmoid)
new_shape = [1, int(num_anchors), input_height, input_width, 1]
name_new_shape = [model_name + "@new_shape"]
node_new_shape = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_new_shape,
value=onnx.helper.make_tensor(
name=name_new_shape[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=[len(new_shape)],
vals=new_shape))
node_list.append(node_new_shape)
outputs_conf_new_shape = [model_name + "@_conf_new_shape"]
node_conf_new_shape = onnx.helper.make_node(
'Reshape',
inputs=outputs_conf_set_zero + name_new_shape,
outputs=outputs_conf_new_shape)
node_list.append(node_conf_new_shape)
outputs_score = [model_name + "@score"]
node_score = onnx.helper.make_node(
'Mul',
inputs=outputs_prob_sigmoid + outputs_conf_new_shape,
outputs=outputs_score)
node_list.append(node_score)
outputs_conf_bool = [model_name + "@conf_bool"]
node_conf_bool = onnx.helper.make_node(
'Greater',
inputs=outputs_conf_new_shape + name_zeros,
outputs=outputs_conf_bool)
node_list.append(node_conf_bool)
outputs_box_x_new_shape = [model_name + "@box_x_new_shape"]
node_box_x_new_shape = onnx.helper.make_node(
'Reshape',
inputs=outputs_box_x_encode + name_new_shape,
outputs=outputs_box_x_new_shape)
node_list.append(node_box_x_new_shape)
outputs_box_y_new_shape = [model_name + "@box_y_new_shape"]
node_box_y_new_shape = onnx.helper.make_node(
'Reshape',
inputs=outputs_box_y_encode + name_new_shape,
outputs=outputs_box_y_new_shape)
node_list.append(node_box_y_new_shape)
outputs_box_w_new_shape = [model_name + "@box_w_new_shape"]
node_box_w_new_shape = onnx.helper.make_node(
'Reshape',
inputs=outputs_box_w_encode + name_new_shape,
outputs=outputs_box_w_new_shape)
node_list.append(node_box_w_new_shape)
outputs_box_h_new_shape = [model_name + "@box_h_new_shape"]
node_box_h_new_shape = onnx.helper.make_node(
'Reshape',
inputs=outputs_box_h_encode + name_new_shape,
outputs=outputs_box_h_new_shape)
node_list.append(node_box_h_new_shape)
outputs_pred_box = [model_name + "@pred_box"]
node_pred_box = onnx.helper.make_node(
'Concat',
inputs=outputs_box_x_new_shape + outputs_box_y_new_shape + \
outputs_box_w_new_shape + outputs_box_h_new_shape,
outputs=outputs_pred_box,
axis=4)
node_list.append(node_pred_box)
outputs_conf_cast = [model_name + "conf_cast"]
node_conf_cast = onnx.helper.make_node(
'Cast', inputs=outputs_conf_bool, outputs=outputs_conf_cast, to=1)
node_list.append(node_conf_cast)
outputs_pred_box_mul_conf = [model_name + "@pred_box_mul_conf"]
node_pred_box_mul_conf = onnx.helper.make_node(
'Mul',
inputs=outputs_pred_box + outputs_conf_cast,
outputs=outputs_pred_box_mul_conf)
node_list.append(node_pred_box_mul_conf)
box_shape = [1, int(num_anchors) * input_height * input_width, 4]
name_box_shape = [model_name + "@box_shape"]
node_box_shape = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_box_shape,
value=onnx.helper.make_tensor(
name=name_box_shape[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=[len(box_shape)],
vals=box_shape))
node_list.append(node_box_shape)
outputs_pred_box_new_shape = [model_name + "@pred_box_new_shape"]
node_pred_box_new_shape = onnx.helper.make_node(
'Reshape',
inputs=outputs_pred_box_mul_conf + name_box_shape,
outputs=outputs_pred_box_new_shape)
node_list.append(node_pred_box_new_shape)
outputs_pred_box_x = [model_name + "@_pred_box_x"]
outputs_pred_box_y = [model_name + "@_pred_box_y"]
outputs_pred_box_w = [model_name + "@_pred_box_w"]
outputs_pred_box_h = [model_name + "@_pred_box_h"]
node_pred_box_split = onnx.helper.make_node(
'Split',
inputs=outputs_pred_box_new_shape,
outputs=outputs_pred_box_x + outputs_pred_box_y + outputs_pred_box_w +
outputs_pred_box_h,
axis=2)
node_list.append(node_pred_box_split)
name_number_two = [model_name + "@number_two"]
node_number_two = onnx.helper.make_node(
"Constant",
inputs=[],
outputs=name_number_two,
value=onnx.helper.make_tensor(
name=name_number_two[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=(),
vals=[2]))
node_list.append(node_number_two)
outputs_half_w = [model_name + "@half_w"]
node_half_w = onnx.helper.make_node(
"Div",
inputs=outputs_pred_box_w + name_number_two,
outputs=outputs_half_w)
node_list.append(node_half_w)
outputs_half_h = [model_name + "@half_h"]
node_half_h = onnx.helper.make_node(
"Div",
inputs=outputs_pred_box_h + name_number_two,
outputs=outputs_half_h)
node_list.append(node_half_h)
outputs_pred_box_x1 = [model_name + "@pred_box_x1"]
node_pred_box_x1 = onnx.helper.make_node(
'Sub',
inputs=outputs_pred_box_x + outputs_half_w,
outputs=outputs_pred_box_x1)
node_list.append(node_pred_box_x1)
outputs_pred_box_y1 = [model_name + "@pred_box_y1"]
node_pred_box_y1 = onnx.helper.make_node(
'Sub',
inputs=outputs_pred_box_y + outputs_half_h,
outputs=outputs_pred_box_y1)
node_list.append(node_pred_box_y1)
outputs_pred_box_x2 = [model_name + "@pred_box_x2"]
node_pred_box_x2 = onnx.helper.make_node(
'Add',
inputs=outputs_pred_box_x + outputs_half_w,
outputs=outputs_pred_box_x2)
node_list.append(node_pred_box_x2)
outputs_pred_box_y2 = [model_name + "@pred_box_y2"]
node_pred_box_y2 = onnx.helper.make_node(
'Add',
inputs=outputs_pred_box_y + outputs_half_h,
outputs=outputs_pred_box_y2)
node_list.append(node_pred_box_y2)
outputs_sqeeze_image_size = [model_name + "@sqeeze_image_size"]
node_sqeeze_image_size = onnx.helper.make_node(
"Squeeze",
axes=[0],
inputs=image_size,
outputs=outputs_sqeeze_image_size)
node_list.append(node_sqeeze_image_size)
output_img_height = [model_name + "@img_height"]
output_img_width = [model_name + "@img_width"]
node_image_size_split = onnx.helper.make_node(
"Split",
inputs=outputs_sqeeze_image_size,
outputs=output_img_height + output_img_width,
axis=-1,
split=[1, 1])
node_list.append(node_image_size_split)
output_img_width_cast = [model_name + "@img_width_cast"]
node_img_width_cast = onnx.helper.make_node(
'Cast', inputs=output_img_width, outputs=output_img_width_cast, to=1)
node_list.append(node_img_width_cast)
output_img_height_cast = [model_name + "@img_height_cast"]
node_img_height_cast = onnx.helper.make_node(
'Cast', inputs=output_img_height, outputs=output_img_height_cast, to=1)
node_list.append(node_img_height_cast)
outputs_pred_box_x1_decode = [model_name + "@pred_box_x1_decode"]
outputs_pred_box_y1_decode = [model_name + "@pred_box_y1_decode"]
outputs_pred_box_x2_decode = [model_name + "@pred_box_x2_decode"]
outputs_pred_box_y2_decode = [model_name + "@pred_box_y2_decode"]
node_pred_box_x1_decode = onnx.helper.make_node(
'Mul',
inputs=outputs_pred_box_x1 + output_img_width_cast,
outputs=outputs_pred_box_x1_decode)
node_list.append(node_pred_box_x1_decode)
node_pred_box_y1_decode = onnx.helper.make_node(
'Mul',
inputs=outputs_pred_box_y1 + output_img_height_cast,
outputs=outputs_pred_box_y1_decode)
node_list.append(node_pred_box_y1_decode)
node_pred_box_x2_decode = onnx.helper.make_node(
'Mul',
inputs=outputs_pred_box_x2 + output_img_width_cast,
outputs=outputs_pred_box_x2_decode)
node_list.append(node_pred_box_x2_decode)
node_pred_box_y2_decode = onnx.helper.make_node(
'Mul',
inputs=outputs_pred_box_y2 + output_img_height_cast,
outputs=outputs_pred_box_y2_decode)
node_list.append(node_pred_box_y2_decode)
name_number_one = [model_name + "@one"]
node_number_one = onnx.helper.make_node(
'Constant',
inputs=[],
outputs=name_number_one,
value=onnx.helper.make_tensor(
name=name_number_one[0] + "@const",
data_type=onnx.TensorProto.FLOAT,
dims=(),
vals=[1]))
node_list.append(node_number_one)
output_new_img_height = [model_name + "@new_img_height"]
node_new_img_height = onnx.helper.make_node(
'Sub',
inputs=output_img_height_cast + name_number_one,
outputs=output_new_img_height)
node_list.append(node_new_img_height)
output_new_img_width = [model_name + "@new_img_width"]
node_new_img_width = onnx.helper.make_node(
'Sub',
inputs=output_img_width_cast + name_number_one,
outputs=output_new_img_width)
node_list.append(node_new_img_width)
outputs_pred_box_x2_sub_w = [model_name + "@pred_box_x2_sub_w"]
node_pred_box_x2_sub_w = onnx.helper.make_node(
'Sub',
inputs=outputs_pred_box_x2_decode + output_new_img_width,
outputs=outputs_pred_box_x2_sub_w)
node_list.append(node_pred_box_x2_sub_w)
outputs_pred_box_y2_sub_h = [model_name + "@pred_box_y2_sub_h"]
node_pred_box_y2_sub_h = onnx.helper.make_node(
'Sub',
inputs=outputs_pred_box_y2_decode + output_new_img_height,
outputs=outputs_pred_box_y2_sub_h)
node_list.append(node_pred_box_y2_sub_h)
outputs_pred_box_x1_clip = [model_name + "@pred_box_x1_clip"]
outputs_pred_box_y1_clip = [model_name + "@pred_box_y1_clip"]
outputs_pred_box_x2_clip = [model_name + "@pred_box_x2_clip"]
outputs_pred_box_y2_clip = [model_name + "@pred_box_y2_clip"]
node_pred_box_x1_clip = onnx.helper.make_node(
'Clip',
inputs=outputs_pred_box_x1_decode,
outputs=outputs_pred_box_x1_clip,
min=0.0,
max=float(MAX_FLOAT32))
node_list.append(node_pred_box_x1_clip)
node_pred_box_y1_clip = onnx.helper.make_node(
'Clip',
inputs=outputs_pred_box_y1_decode,
outputs=outputs_pred_box_y1_clip,
min=0.0,
max=float(MAX_FLOAT32))
node_list.append(node_pred_box_y1_clip)
node_pred_box_x2_clip = onnx.helper.make_node(
'Clip',
inputs=outputs_pred_box_x2_sub_w,
outputs=outputs_pred_box_x2_clip,
min=0.0,
max=float(MAX_FLOAT32))
node_list.append(node_pred_box_x2_clip)
node_pred_box_y2_clip = onnx.helper.make_node(
'Clip',
inputs=outputs_pred_box_y2_sub_h,
outputs=outputs_pred_box_y2_clip,
min=0.0,
max=float(MAX_FLOAT32))
node_list.append(node_pred_box_y2_clip)
outputs_pred_box_x2_res = [model_name + "@box_x2_res"]
node_pred_box_x2_res = onnx.helper.make_node(
'Sub',
inputs=outputs_pred_box_x2_decode + outputs_pred_box_x2_clip,
outputs=outputs_pred_box_x2_res)
node_list.append(node_pred_box_x2_res)
outputs_pred_box_y2_res = [model_name + "@box_y2_res"]
node_pred_box_y2_res = onnx.helper.make_node(
'Sub',
inputs=outputs_pred_box_y2_decode + outputs_pred_box_y2_clip,
outputs=outputs_pred_box_y2_res)
node_list.append(node_pred_box_y2_res)
node_pred_box_result = onnx.helper.make_node(
'Concat',
inputs=outputs_pred_box_x1_clip + outputs_pred_box_y1_clip +
outputs_pred_box_x2_res + outputs_pred_box_y2_res,
outputs=outputs['Boxes'],
axis=-1)
node_list.append(node_pred_box_result)
score_shape = [1, input_height * input_width * int(num_anchors), class_num]
name_score_shape = [model_name + "@score_shape"]
node_score_shape = onnx.helper.make_node(
"Constant",
inputs=[],
outputs=name_score_shape,
value=onnx.helper.make_tensor(
name=name_score_shape[0] + "@const",
data_type=onnx.TensorProto.INT64,
dims=[len(score_shape)],
vals=score_shape))
node_list.append(node_score_shape)
node_score_new_shape = onnx.helper.make_node(
'Reshape',
inputs=outputs_score + name_score_shape,
outputs=outputs['Scores'])
node_list.append(node_score_new_shape)
return node_list
x2paddle/x2paddle/op_mapper/paddle2onnx/paddle_op_mapper.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
import sys
import x2paddle
import os
import numpy as np
import paddle.fluid.core as core
import paddle.fluid as fluid
import onnx
from onnx import helper, onnx_pb
from x2paddle.op_mapper.paddle2onnx.opset9.opset import OpSet9
from x2paddle.op_mapper.paddle2onnx.opset10.opset import OpSet10
from x2paddle.op_mapper.paddle2onnx.opset11.opset import OpSet11
class PaddleOpMapper(object):
def __init__(self):
self.support_opsets = [9, 10, 11]
self.default_opset = 10
self.name_counter = dict()
self.op_set = None
def convert(self, program, save_dir, scope=None, opset_version=10):
self.op_set = self.create_opset(opset_version)
weight_nodes = self.op_set.convert_weights(program, scope=scope)
op_nodes = list()
input_nodes = list()
output_nodes = list()
unsupported_ops = set()
print("Translating PaddlePaddle to ONNX...\n")
for block in program.blocks:
for i, op in enumerate(block.ops):
sys.stdout.write("\rTotal:{}, Current:{} : {} ".format(
len(block.ops), i + 1, op.type))
sys.stdout.flush()
if not hasattr(self.op_set, op.type):
unsupported_ops.add(op.type)
continue
if len(unsupported_ops) > 0:
continue
node = getattr(self.op_set, op.type)(op, block)
if op.type == 'feed':
print(node.name)
input_nodes.append(node)
elif op.type == 'fetch':
output_nodes.append(node)
else:
if isinstance(node, list):
op_nodes = op_nodes + node
else:
op_nodes.append(node)
if len(unsupported_ops) > 0:
print("\nThere's {} ops are not supported yet".format(
len(unsupported_ops)))
for op in unsupported_ops:
print("=========== {} ===========".format(op))
return
graph = helper.make_graph(
nodes=weight_nodes + op_nodes,
name='onnx_model_from_paddle',
initializer=[],
inputs=input_nodes,
outputs=output_nodes)
opset_imports = [helper.make_opsetid("", opset_version)]
model = helper.make_model(
graph, producer_name='X2Paddle', opset_imports=opset_imports)
onnx.checker.check_model(model)
if not os.path.isdir(save_dir):
os.makedirs(save_dir)
with open(os.path.join(save_dir, 'x2paddle_model.onnx'), 'wb') as f:
f.write(model.SerializeToString())
print("\nTranslated model saved in {}".format(
os.path.join(save_dir, 'x2paddle_model.onnx')))
def create_opset(self, opset_version=10):
run_opset = self.default_opset
opset = ''
if opset_version in self.support_opsets:
run_opset = opset_version
else:
for support_opset_version in self.support_opsets:
if support_opset_version < opset_version:
run_opset = support_opset_version
else:
break
print(
'Now, onnx2paddle support convert onnx model opset_verison {},'
'opset_verison of your onnx model is {}, automatically treated as op_set: {}.'
.format(self.support_opsets, opset_version, run_opset))
opset = 'OpSet' + str(run_opset)
return eval(opset)()
x2paddle/x2paddle/op_mapper/tf_op_mapper.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from x2paddle.decoder.tf_decoder import TFGraph
from x2paddle.core.op_mapper import OpMapper
from x2paddle.core.util import *
import inspect
import numpy
import sys
# compute padding size for SAME mode
def get_same_padding(in_size, kernel_size, stride):
new_size = int(math.ceil(in_size * 1.0 / stride))
pad_size = (new_size - 1) * stride + kernel_size - in_size
if pad_size < 0:
pad_size = 0
pad0 = int(pad_size / 2)
pad1 = pad_size - pad0
return [pad0, pad1]
def nhwc_dim_to_nchw(node, dim):
tf_data_format = list(node.tf_data_format)
pd_data_format = list(node.pd_data_format)
if isinstance(dim, list):
for i in range(len(dim)):
char = tf_data_format[dim[i]]
dim[i] = pd_data_format.index(char)
else:
char = tf_data_format[dim]
dim = pd_data_format.index(char)
return dim
if dim < 0:
dim += 4
if dim > 0:
dim = (dim + 1) % 4 + int((dim + 1) / 4)
return dim
class TFOpMapper(OpMapper):
directly_map_ops = {
'Relu': ['relu'],
'Relu6': ['relu6'],
'Shape': ['shape'],
'Abs': ['abs'],
'Sigmoid': ['sigmoid'],
'Exp': ['exp'],
'Rsqrt': ['rsqrt'],
'swish_f32': ['swish'],
'Tanh': ['tanh'],
'LeakyRelu': ['leaky_relu', {
'alpha': 'alpha'
}]
}
elementwise_ops = {
'Add': 'elementwise_add',
'RealDiv': 'elementwise_div',
'Sub': 'elementwise_sub',
'Maximum': 'elementwise_max',
'Mul': 'elementwise_mul',
'FloorDiv': 'elementwise_floordiv'
}
def __init__(self, decoder):
super(TFOpMapper, self).__init__()
self.decoder = decoder
self.graph = decoder.tf_graph
self.batch_node = None
self.weights = dict()
self.omit_nodes = list()
self.used_custom_layers = dict()
not_placeholder = list()
for name in self.graph.input_nodes:
if self.graph.get_node(name).layer_type != "Placeholder" \
and self.graph.get_node(name).layer_type != "OneShotIterator":
not_placeholder.append(name)
for name in not_placeholder:
idx = self.graph.input_nodes.index(name)
del self.graph.input_nodes[idx]
sys.stderr.write("Total nodes: {}\n".format(len(self.graph.topo_sort)))
unsupported_ops = set()
for i, node_name in enumerate(self.graph.topo_sort):
sys.stderr.write("\rConverting node {} ... ".format(i + 1))
node = self.graph.get_node(node_name)
op = node.layer_type
if op in self.directly_map_ops:
if len(unsupported_ops) > 0:
continue
self.directly_map(node)
elif op in self.elementwise_ops:
if len(unsupported_ops) > 0:
continue
self.elementwise_map(node)
elif hasattr(self, op):
if len(unsupported_ops) > 0:
continue
func = getattr(self, op)
func(node)
else:
unsupported_ops.add(op)
if len(unsupported_ops) > 0:
sys.stderr.write("=========={} Ops are not supported yet======\n".
format(len(unsupported_ops)))
for op in unsupported_ops:
sys.stderr.write("========== {} ==========\n".format(op))
sys.exit(-1)
sys.stderr.write('\nDone!\n')
def add_omit_nodes(self, in_node_name, out_node_name):
in_node = self.graph.get_node(in_node_name)
out_node = self.graph.get_node(out_node_name)
index = in_node.outputs.index(out_node_name)
# del in_node.outputs[index]
index = out_node.inputs.index(in_node_name)
# del out_node.inputs[index]
self.omit_nodes.append(in_node.layer_name)
def directly_map(self, node):
assert node.layer_type in self.directly_map_ops
op_info = self.directly_map_ops[node.layer_type]
input = self.graph.get_node(node.layer.input[0], copy=True)
attr = dict()
for param in op_info[1:]:
tf_param_name = list(param.keys())[0]
pd_param_name = list(param.values())[0]
tf_param = node.get_attr(tf_param_name)
attr[pd_param_name] = tf_param
node.fluid_code.add_layer(
op_info[0], inputs=input, output=node, param_attr=attr)
def elementwise_map(self, node):
assert node.layer_type in self.elementwise_ops
op_type = self.elementwise_ops[node.layer_type]
x = self.graph.get_node(node.layer.input[0], copy=True)
y = self.graph.get_node(node.layer.input[1], copy=True)
x_shape = x.out_shapes[0]
y_shape = y.out_shapes[0]
if len(x_shape) == 0:
x_shape = [1]
if len(y_shape) == 0:
y_shape = [1]
# incomplement broadcasting support for paddle
x_input = x
y_input = y
if len(x_shape) < len(y_shape):
unrevertable_ops = [
"elementwise_sub", "elementwise_div", "elementwise_floordiv",
"elementwise_mod", "elementwise_pow"
]
if op_type not in unrevertable_ops:
x_input = y
y_input = x
x_shape = y.out_shapes[0]
if len(x_shape) == 0:
x_shape = [1]
y_shape = x.out_shapes[0]
if len(y_shape) == 0:
y_shape = [1]
else:
if len(x_shape) == 1 and len(y_shape) == 4 and x_shape[
0] == y_shape[-1] and y_shape.count(-1) < 1:
shape = [1, x_shape[0], 1, 1]
attr = {"shape": shape}
node.fluid_code.add_layer(
"reshape",
inputs=x_input,
output="reshape_x",
param_attr=attr)
if y_shape[0] != 1:
attr = {"expand_times": [y_shape[0], 1, 1, 1]}
node.fluid_code.add_layer(
"expand",
inputs="reshape_x",
output="reshape_x",
param_attr=attr)
inputs = {"x": "reshape_x", "y": y_input}
node.fluid_code.add_layer(
op_type, inputs=inputs, output=node, param_attr=None)
return
else:
raise Exception("Unexpected situation happend")
if len(x_shape) == 4 and len(y_shape) == 1:
if x_input.tf_data_format == "NHWC":
axis = 1
else:
axis = -1
attr = {"axis": axis}
inputs = {"x": x_input, "y": y_input}
node.fluid_code.add_layer(
op_type, inputs=inputs, output=node, param_attr=attr)
return
is_sub_seq = True
for i in range(len(y_shape)):
index = -1 * i - 1
if y_shape[index] != x_shape[index]:
is_sub_seq = False
if not is_sub_seq:
if x_shape.count(-1) > 2:
x_shape = self.decoder.infer_tensor_shape(x_input)
if y_shape.count(-1) > 2:
y_shape = self.decoder.infer_tensor_shape(y_input)
x_expand_times = [1] * len(x_shape)
y_expand_times = [1] * len(y_shape)
x_need_expand = False
y_need_expand = False
for i in range(len(y_shape)):
index = -1 * i - 1
if y_shape[index] != x_shape[index]:
if y_shape[index] == 1:
y_expand_times[index] = x_shape[index]
y_need_expand = True
elif x_shape[index] == 1:
x_expand_times[index] = y_shape[index]
x_need_expand = True
else:
raise Exception("Unexpected situation happend")
if x_need_expand:
if len(x_expand_times) == 3 and x.tf_data_format == "NHWC":
x_expand_times = [x_expand_times[i] for i in [2, 0, 1]]
if len(x_expand_times) == 4 and x.tf_data_format == "NHWC":
x_expand_times = [x_expand_times[i] for i in [0, 3, 1, 2]]
attr = {"expand_times": x_expand_times}
node.fluid_code.add_layer(
"expand", inputs=x_input, output="x_tmp", param_attr=attr)
x_input = "x_tmp"
if y_need_expand:
if len(y_expand_times) == 3 and y.tf_data_format == "NHWC":
y_expand_times = [y_expand_times[i] for i in [2, 0, 1]]
if len(y_expand_times) == 4 and y.tf_data_format == "NHWC":
y_expand_times = [y_expand_times[i] for i in [0, 3, 1, 2]]
attr = {"expand_times": y_expand_times}
node.fluid_code.add_layer(
"expand", inputs=y_input, output="y_tmp", param_attr=attr)
y_input = "y_tmp"
inputs = {"x": x_input, "y": y_input}
node.fluid_code.add_layer(
op_type, inputs=inputs, output=node, param_attr=None)
def Placeholder(self, node):
shape = node.out_shapes[0]
assert len(shape) != 0, "Unknown shape of input nodes[{}].".format(
node.layer_name)
if node.tf_data_format == "NHWC" and len(shape) == 4:
shape = [shape[i] for i in [0, 3, 1, 2]]
elif node.tf_data_format == "NCHW" and len(shape) == 4:
self.graph.data_format_propagation(node)
dtype = node.dtype
attr = {
'dtype': string(dtype),
'shape': shape,
'name': string(node.layer_name),
'append_batch_size': False
}
if shape[0] < 0:
self.batch_node = node
node.fluid_code.add_layer(
"data", inputs=None, output=node, param_attr=attr)
def OneShotIterator(self, node):
return self.Placeholder(node)
def Const(self, node):
shape = node.out_shapes[0]
dtype = node.dtype
value = node.value
initializer = "Constant(0.0)"
if len(shape) == 0:
assert value.size == 1, "Unexpected situation happend"
shape = [1]
initializer = "Constant({})".format(value)
self.weights[node.layer_name] = node.value
if node.tf_data_format == "NHWC":
if len(shape) == 4:
shape = [shape[i] for i in [0, 3, 1, 2]]
if len(shape) == 3:
shape = [shape[i] for i in [2, 0, 1]]
self.weights[node.layer_name] = numpy.transpose(node.value,
(2, 0, 1))
elif node.tf_data_format == "NCHW":
if len(shape) == 4:
self.graph.data_format_propagation(node)
attr = {
'dtype': string(dtype),
'shape': shape,
'name': string(node.layer_name),
'default_initializer': initializer
}
node.fluid_code.add_layer(
"create_parameter", inputs=None, output=node, param_attr=attr)
def Transpose(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
perm = self.graph.get_node(node.layer.input[1], copy=True)
assert perm.layer_type == "Const", "Perm of transpose OP should be Const"
del self.weights[perm.layer_name.replace('/', '_')]
perm.fluid_code.clear()
perm = perm.value.tolist()
if perm == [0, 3, 1, 2] and input.data_format == "NHWC":
input_name = input.layer_name
if hasattr(input, "index"):
input_name = input_name + "[{}]".format(input.index)
node.fluid_code.add_layer("{} = {}").format(node.layer_name,
input_name)
node.tf_data_format = "NCHW"
self.graph.data_format_propagation(node)
elif perm == [0, 2, 3, 1] and input.tf_data_format == "NCHW":
input_name = input.layer_name
if hasattr(input, "index"):
input_name = input_name + "[{}]".format(input.index)
node.fluid_code.add_layer("{} = {}").format(node.layer_name,
input_name)
node.tf_data_format = "NHWC"
self.graph.data_format_propagation(node)
elif len(input.out_shapes[0]) > 4:
tf_data_format = list(input.tf_data_format)
pd_data_format = list(input.pd_data_format)
new_perm = [i for i in range(len(perm))]
for i in range(len(perm)):
char0 = tf_data_format[i]
char1 = tf_data_format[perm[i]]
index0 = pd_data_format.index(char0)
index1 = pd_data_format.index(char1)
new_perm[index0] = index1
node.tf_data_format = [tf_data_format[i] for i in perm]
node.pd_data_format = [pd_data_format[i] for i in perm]
attr = {'perm': new_perm}
node.fluid_code.add_layer(
"transpose", inputs=input, output=node, param_attr=attr)
elif len(node.out_shapes[0]) != 4:
attr = {'perm': perm}
node.fluid_code.add_layer(
"transpose", inputs=input, output=node, param_attr=attr)
else:
raise Exception("Unexpected situation happend in Transpose OP")
def MaxPool(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
in_shape = input.out_shapes[0]
if in_shape.count(-1) > 2:
in_shape = self.decoder.infer_tensor(input).shape
k_size = node.get_attr("ksize")
strides = node.get_attr("strides")
data_format = node.get_attr("data_format").decode()
pad_mode = node.get_attr("padding").decode()
channel_first = data_format == "NCHW"
if not channel_first:
in_shape = [in_shape[i] for i in [0, 3, 1, 2]]
strides = [strides[i] for i in [0, 3, 1, 2]]
k_size = [k_size[i] for i in [0, 3, 1, 2]]
else:
self.graph.data_format_propagation(node)
attr = {
"pool_size": k_size[2:4],
"pool_type": string("max"),
"pool_padding": string(pad_mode),
"pool_stride": strides[2:4]
}
node.fluid_code.add_layer(
"pool2d", inputs=input, output=node, param_attr=attr)
def Conv2D(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
kernel = self.graph.get_node(node.layer.input[1], copy=True)
assert kernel.layer_type == "Const", "Kernel of Conv2D should be Const"
self.add_omit_nodes(kernel.layer_name, node.layer_name)
in_shape = input.out_shapes[0]
if in_shape.count(-1) > 2:
in_shape = self.decoder.infer_tensor(input).shape
k_size = kernel.out_shapes[0]
if k_size.count(-1) > 2:
k_size = self.decoder.infer_tensor(kernel).shape
strides = node.get_attr("strides")
dilations = node.get_attr("dilations")
data_format = node.get_attr("data_format").decode()
pad_mode = node.get_attr("padding").decode()
channel_first = data_format == "NCHW"
self.weights[kernel.layer_name.replace('/', '_')] = numpy.transpose(
kernel.value, (3, 2, 0, 1))
if not channel_first:
in_shape = [in_shape[i] for i in [0, 3, 1, 2]]
strides = [strides[i] for i in [0, 3, 1, 2]]
dilations = [dilations[i] for i in [0, 3, 1, 2]]
else:
self.graph.data_format_propagation(node)
attr = {
"bias_attr": False,
"param_attr": string(kernel.layer_name),
"num_filters": k_size[3],
"filter_size": k_size[0:2],
"stride": strides[2:4],
"dilation": dilations[2:4],
"padding": string(pad_mode)
}
node.fluid_code.add_layer(
"conv2d", inputs=input, output=node, param_attr=attr)
def BiasAdd(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
bias = self.graph.get_node(node.layer.input[1], copy=True)
axis = -1
if input.tf_data_format == "NHWC" and len(input.out_shapes[0]) == 4:
axis = 1
inputs = {"x": input, "y": bias}
attr = {"axis": axis}
node.fluid_code.add_layer(
"elementwise_add", inputs=inputs, output=node, param_attr=attr)
def FusedBatchNorm(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
gamma = self.graph.get_node(node.layer.input[1], copy=True)
beta = self.graph.get_node(node.layer.input[2], copy=True)
moving_mean = self.graph.get_node(node.layer.input[3], copy=True)
moving_var = self.graph.get_node(node.layer.input[4], copy=True)
data_format = node.get_attr("data_format").decode()
channel_first = data_format == "NCHW"
assert gamma.layer_type == "Const"
assert beta.layer_type == "Const"
assert moving_mean.layer_type == "Const"
assert moving_var.layer_type == "Const"
self.add_omit_nodes(gamma.layer_name, node.layer_name)
self.add_omit_nodes(beta.layer_name, node.layer_name)
self.add_omit_nodes(moving_mean.layer_name, node.layer_name)
self.add_omit_nodes(moving_var.layer_name, node.layer_name)
if channel_first:
self.data_format_propagation(node)
attr = {
"epsilon": node.get_attr("epsilon"),
"param_attr": string(gamma.layer_name),
"bias_attr": string(beta.layer_name),
"moving_mean_name": string(moving_mean.layer_name),
"moving_variance_name": string(moving_var.layer_name),
"is_test": True
}
node.fluid_code.add_layer(
"batch_norm", inputs=input, output=node, param_attr=attr)
def FusedBatchNormV3(self, node):
return self.FusedBatchNorm(node)
def DepthwiseConv2dNative(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
kernel = self.graph.get_node(node.layer.input[1], copy=True)
assert kernel.layer_type == "Const", "Kernel of DepthwiseConv2DNative should be Const"
self.add_omit_nodes(kernel.layer_name, node.layer_name)
in_shape = input.out_shapes[0]
if in_shape.count(-1) > 2:
in_shape = self.decoder.infer_tensor(input).shape
k_size = kernel.out_shapes[0]
if k_size.count(-1) > 2:
k_size = self.decoder.infer_tensor(kernel).shape
strides = node.get_attr("strides")
dilations = node.get_attr("dilations")
data_format = node.get_attr("data_format").decode()
pad_mode = node.get_attr("padding").decode()
channel_first = data_format == "NCHW"
self.weights[kernel.layer_name.replace('/', '_')] = numpy.transpose(
kernel.value, (2, 3, 0, 1))
if not channel_first:
in_shape = [in_shape[i] for i in [0, 3, 1, 2]]
strides = [strides[i] for i in [0, 3, 1, 2]]
dilations = [dilations[i] for i in [0, 3, 1, 2]]
else:
self.data_format_propagation(node)
attr = {
"bias_attr": False,
"param_attr": string(kernel.layer_name),
"num_filters": in_shape[1],
"filter_size": k_size[0:2],
"stride": strides[2:4],
"dilation": dilations[2:4],
"groups": k_size[3] * in_shape[1],
"use_cudnn": False,
"padding": string(pad_mode)
}
node.fluid_code.add_layer(
"conv2d", inputs=input, output=node, param_attr=attr)
def Reshape(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
param = self.graph.get_node(node.layer.input[1], copy=True)
is_variable = False
if param.layer_type == "Const":
attr = {"shape": param.value.tolist()}
self.add_omit_nodes(param.layer_name, node.layer_name)
else:
# Here is a trick method to solove tensor parameter in tensorflow
shape = self.decoder.infer_shape_tensor(param, node.out_shapes[0])
if shape.count(-1) <= 1:
attr = {"shape": shape}
self.add_omit_nodes(param.layer_name, node.layer_name)
elif shape.count(-1) == 2 and shape[0] == -1:
shape[0] = 0
attr = {"shape": shape}
self.add_omit_nodes(param.layer_name, node.layer_name)
else:
assert len(param.out_shapes[
0]) == 1, "Unexpected situation of shape parameter"
attr = {"shape": [-1]}
node.fluid_code.add_layer(
"reshape",
inputs=param,
output="shape_param",
param_attr=attr)
attr = {"num_or_sections": param.out_shapes[0][0], "dim": 0}
node.fluid_code.add_layer(
"split", inputs="shape_param", output=node, param_attr=attr)
new_param = "["
for i in range(param.out_shapes[0][0]):
new_param += (node.layer_name + "[{}]".format(i) + ", ")
new_param = new_param.strip(", ") + "]"
attr = {"shape": new_param}
is_variable = True
# to change [192, -1]->[-1, 192], allways put -1 in the first dimension
# optimization for Paddle-Lite
in_shape = input.out_shapes[0]
if not is_variable and in_shape.count(-1) < 1:
total_size = 1
for i in range(len(in_shape)):
total_size *= in_shape[i]
for i in range(len(attr["shape"])):
if attr["shape"][i] == 0:
attr["shape"][i] = in_shape[i]
if attr["shape"][i] != -1:
total_size /= attr["shape"][i]
if attr["shape"].count(-1) > 0:
index = attr["shape"].index(-1)
attr["shape"][index] = int(total_size)
attr["shape"][0] = -1
if len(input.out_shapes[0]) == 4 and node.tf_data_format == "NHWC":
if len(attr["shape"]) < 3:
perm = {"perm": [0, 2, 3, 1]}
node.fluid_code.add_layer(
"transpose", inputs=input, output=node, param_attr=perm)
node.fluid_code.add_layer(
"reshape", inputs=node, output=node, param_attr=attr)
return
if len(attr["shape"]) == 4 and node.tf_data_format == "NHWC":
input_shape = self.decoder.infer_tensor(input).shape
if input_shape[1] == attr["shape"][1]:
attr["shape"] = [attr["shape"][i] for i in [0, 3, 1, 2]]
else:
perm = {"perm": [0, 2, 3, 1]}
node.fluid_code.add_layer(
"transpose", inputs=input, output=node, param_attr=perm)
node.fluid_code.add_layer(
"reshape", inputs=node, output=node, param_attr=attr)
perm = {"perm": [0, 3, 1, 2]}
node.fluid_code.add_layer(
"transpose", inputs=node, output=node, param_attr=perm)
return
if len(attr["shape"]) == 5:
attr["shape"] = [attr["shape"][i] for i in [0, 1, 4, 2, 3]]
node.fluid_code.add_layer(
"reshape", inputs=input, output=node, param_attr=attr)
def AvgPool(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
in_shape = input.out_shapes[0]
if in_shape.count(-1) > 2:
in_shape = self.decoder.infer_tensor(input).shape
k_size = node.get_attr("ksize")
strides = node.get_attr("strides")
data_format = node.get_attr("data_format").decode()
pad_mode = node.get_attr("padding").decode()
channel_first = data_format == "NCHW"
if not channel_first:
in_shape = [in_shape[i] for i in [0, 3, 1, 2]]
strides = [strides[i] for i in [0, 3, 1, 2]]
k_size = [k_size[i] for i in [0, 3, 1, 2]]
else:
self.graph.data_format_propagation(node)
attr = {
"pool_size": k_size[2:4],
"pool_type": string("avg"),
"pool_stride": strides[2:4],
"pool_padding": string(pad_mode)
}
node.fluid_code.add_layer(
"pool2d", inputs=input, output=node, param_attr=attr)
def SplitV(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
num_sections = self.graph.get_node(node.layer.input[1], copy=True)
dim = self.graph.get_node(node.layer.input[2], copy=True)
assert num_sections.layer_type == "Const"
assert dim.layer_type == "Const"
self.add_omit_nodes(num_sections.layer_name, node.layer_name)
self.add_omit_nodes(dim.layer_name, node.layer_name)
dim = dim.value
if input.tf_data_format == "NHWC" and len(input.out_shapes[0]) == 4:
dim = nhwc_dim_to_nchw(input, dim)
attr = {
"num_or_sections": num_sections.value.tolist(),
"dim": dim.value
}
node.fluid_code.add_layer(
"split", inputs=input, output=node, param_attr=attr)
def ConcatV2(self, node):
inputs = [
self.graph.get_node(
name, copy=True) for name in node.layer.input[:-1]
]
axis = self.graph.get_node(node.layer.input[-1], copy=True)
assert axis.layer_type == "Const"
self.add_omit_nodes(axis.layer_name, node.layer_name)
axis = axis.value
if inputs[0].tf_data_format == "NHWC" and len(inputs[0].out_shapes[
0]) == 4:
axis = nhwc_dim_to_nchw(inputs[0], axis)
attr = {"axis": axis}
node.fluid_code.add_layer(
"concat", inputs=inputs, output=node, param_attr=attr)
def Tile(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
expand_times = self.graph.get_node(node.layer.input[1], copy=True)
self.add_omit_nodes(expand_times.layer_name, node.layer_name)
if expand_times.layer_type == "Const":
expand_times = expand_times.value.tolist()
else:
expand_times = self.decoder.infer_shape_tensor(expand_times)
if input.tf_data_format == "NHWC":
if len(input.out_shapes[0]) == 4:
expand_times = [expand_times[i] for i in [0, 3, 1, 2]]
elif len(input.out_shapes[0]) == 3:
expand_times = [expand_times[i] for i in [2, 0, 1]]
for i in range(len(expand_times)):
if expand_times[i] < 0:
expand_times[i] = 1
attr = {"expand_times": expand_times}
node.fluid_code.add_layer(
"expand", inputs=input, output=node, param_attr=attr)
def Pack(self, node):
inputs = [
self.graph.get_node(
name, copy=True) for name in node.layer.input
]
axis = node.get_attr("axis")
if inputs[0].tf_data_format == "NHWC" and len(inputs[0].out_shapes[
0]) == 4:
tf_data_format = list(inputs[0].tf_data_format)
tf_data_format.insert(axis, str(len(tf_data_format)))
axis = nhwc_dim_to_nchw(inputs[0], axis)
pd_data_format = list(inputs[0].pd_data_format)
pd_data_format.insert(axis, str(len(pd_data_format)))
node.tf_data_format = "".join(tf_data_format)
node.pd_data_format = "".join(pd_data_format)
attr = {"axis": axis}
node.fluid_code.add_layer(
"stack", inputs=inputs, output=node, param_attr=attr)
def Pad(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
paddings = self.graph.get_node(node.layer.input[1], copy=True)
assert paddings.layer_type == "Const", "Padding should be Const"
self.add_omit_nodes(paddings.layer_name, node.layer_name)
paddings = paddings.value.flatten().tolist()
if input.tf_data_format == "NHWC" and len(input.out_shapes[0]) == 4:
paddings = [paddings[i] for i in [0, 1, 6, 7, 2, 3, 4, 5]]
pad_op = "pad"
if len(input.out_shapes[0]) == 4:
if paddings[0] + paddings[1] + paddings[2] + paddings[3] == 0:
paddings = paddings[4:]
pad_op = "pad2d"
attr = {"paddings": paddings}
node.fluid_code.add_layer(
pad_op, inputs=input, output=node, param_attr=attr)
def MirrorPad(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
paddings = self.graph.get_node(node.layer.input[1], copy=True)
assert paddings.layer_type == "Const", "Padding should be Const"
self.add_omit_nodes(paddings.layer_name, node.layer_name)
paddings = paddings.value.flatten().tolist()
mode = node.get_attr("mode").decode()
assert mode == "REFLECT", "Only support 'REFLECT` mode in MirrorPad"
if input.tf_data_format == "NHWC" and len(input.out_shapes[0]) == 4:
paddings = [paddings[i] for i in [0, 1, 6, 7, 2, 3, 4, 5]]
pad_op = "pad"
if len(input.out_shapes[0]) == 4:
if paddings[0] + paddings[1] + paddings[2] + paddings[3] == 0:
paddings = paddings[4:]
pad_op = "pad2d"
attr = {"paddings": paddings, "mode": string("reflect")}
node.fluid_code.add_layer(
pad_op, inputs=input, output=node, param_attr=attr)
def Range(self, node):
start = self.graph.get_node(node.layer.input[0], copy=True)
limit = self.graph.get_node(node.layer.input[1], copy=True)
delta = self.graph.get_node(node.layer.input[2], copy=True)
self.add_omit_nodes(start.layer_name, node.layer_name)
self.add_omit_nodes(limit.layer_name, node.layer_name)
self.add_omit_nodes(delta.layer_name, node.layer_name)
if start.layer_type == "Const":
start = start.value
else:
start = self.decoder.infer_tensor(start)
if limit.layer_type == "Const":
limit = limit.value
else:
limit = self.decoder.infer_tensor(limit)
if delta.layer_type == "Const":
delta = delta.value
else:
delta = self.decoder.infer_tensor(delta)
inputs = {"start": start, "end": limit, "step": delta}
attr = {"dtype": string(node.dtype)}
node.fluid_code.add_layer(
"range", inputs=inputs, output=node, param_attr=attr)
def Mean(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
reduce_idx = self.graph.get_node(node.layer.input[1], copy=True)
assert reduce_idx.layer_type == "Const", "Only support Const parameter[reduce_idx]"
dims = reduce_idx.value.tolist()
keep_dims = node.get_attr("keep_dims")
if input.tf_data_format == "NHWC" and len(input.out_shapes[0]) == 4:
for i in range(len(dims)):
dims[i] = nhwc_dim_to_nchw(input, dims[i])
attr = {"dim": dims, "keep_dim": keep_dims}
node.fluid_code.add_layer(
"reduce_mean", inputs=input, output=node, param_attr=attr)
def MatMul(self, node):
x = self.graph.get_node(node.layer.input[0], copy=True)
y = self.graph.get_node(node.layer.input[1], copy=True)
transpose_a = node.get_attr('transpose_a')
transpose_b = node.get_attr('transpose_b')
inputs = {"x": x, "y": y}
# fix paddle shape infer problem
# should be removed after paddle 1.6
if x.out_shapes[0][-1] < 0 and y.out_shapes[0][0] > 0:
shape = x.out_shapes[0]
shape[-1] = y.out_shapes[0][0]
attr = {"shape": shape}
node.fluid_code.add_layer(
"reshape", inputs=x, output=x, param_attr=attr)
attr = {"transpose_x": transpose_a, "transpose_y": transpose_b}
node.fluid_code.add_layer(
"matmul", inputs=inputs, output=node, param_attr=attr)
def ArgMax(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
axis = self.graph.get_node(node.layer.input[1], copy=True)
assert axis.layer_type == "Const", "ArgMax only support Const parameter"
self.add_omit_nodes(axis.layer_name, node.layer_name)
axis = axis.value
if input.tf_data_format == "NHWC" and len(input.out_shapes[0]) == 4:
axis = nhwc_dim_to_nchw(input, axis)
attr = {"axis": axis}
node.fluid_code.add_layer(
"argmax", inputs=input, output=node, param_attr=attr)
def StridedSlice(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
begin = self.graph.get_node(node.layer.input[1], copy=True)
end = self.graph.get_node(node.layer.input[2], copy=True)
strides = self.graph.get_node(node.layer.input[3], copy=True)
assert begin.layer_type == "Const"
assert end.layer_type == "Const"
assert strides.layer_type == "Const"
self.add_omit_nodes(begin.layer_name, node.layer_name)
self.add_omit_nodes(end.layer_name, node.layer_name)
self.add_omit_nodes(strides.layer_name, node.layer_name)
strides = strides.value.tolist()
assert len(set(strides)) == 1 and strides[0] == 1
begin = begin.value.tolist()
end = end.value.tolist()
if input.tf_data_format == "NHWC" and len(input.out_shapes[0]) == 4:
begin = [begin[i] for i in [0, 3, 1, 2]]
end = [end[i] for i in [0, 3, 1, 2]]
for i in range(len(end)):
if end[i] == 0:
end[i] = 999999
attr = {
"axes": [i for i in range(len(strides))],
"starts": begin,
"ends": end
}
shrink_axis_mask = node.get_attr('shrink_axis_mask')
squeeze_dims = list()
for i in range(len(begin)):
x = shrink_axis_mask >> i & 1
if x == 1:
squeeze_dims.append(i)
node.fluid_code.add_layer(
"slice", inputs=input, output=node, param_attr=attr)
if shrink_axis_mask > 0 and len(input.out_shapes[0]) == 5:
attr = {"axes": squeeze_dims}
node.fluid_code.add_layer(
"squeeze", inputs=node, output=node, param_attr=attr)
def Slice(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
begin = self.graph.get_node(node.layer.input[1], copy=True)
size = self.graph.get_node(node.layer.input[2], copy=True)
self.add_omit_nodes(begin.layer_name, node.layer_name)
self.add_omit_nodes(size.layer_name, node.layer_name)
if begin.layer_type == "Const":
begin = begin.value.tolist()
else:
begin = self.decoder.infer_tensor(begin).tolist()
if size.layer_type == "const":
size = size.value.tolist()
else:
size = self.decoder.infer_tensor(size).tolist()
if input.tf_data_format == "NHWC" and len(input.out_shapes[0]) == 4:
size = [size[i] for i in [0, 3, 1, 2]]
begin = [begin[i] for i in [0, 3, 1, 2]]
for i in range(len(size)):
if size[i] < 0:
size[i] = 99999999
else:
size[i] = size[i] + begin[i]
attr = {
"axes": [i for i in range(len(size))],
"starts": begin,
"ends": size
}
node.fluid_code.add_layer(
"slice", inputs=input, output=node, param_attr=attr)
def Conv2DBackpropInput(self, node):
out_shape = self.graph.get_node(node.layer.input[0], copy=True)
kernel = self.graph.get_node(node.layer.input[1], copy=True)
input = self.graph.get_node(node.layer.input[2], copy=True)
assert kernel.layer_type == "Const", "Kernel of Conv2DBackpropInput should be Const"
self.add_omit_nodes(kernel.layer_name, node.layer_name)
self.add_omit_nodes(out_shape.layer_name, node.layer_name)
if out_shape.layer_type == "Const":
out_shape = out_shape.value.tolist()
else:
out_shape = self.decoder.infer_shape_tensor(out_shape,
node.out_shapes[0])
in_shape = input.out_shapes[0]
if in_shape.count(-1) > 2:
in_shape = self.decoder.infer_tensor(input).shape
k_size = kernel.out_shapes[0]
if k_size.count(-1) > 2:
k_size = self.decoder.infer_tensor(kernel).shape
pad_mode = node.get_attr("padding").decode()
strides = node.get_attr("strides")
dilations = node.get_attr("dilations")
data_format = node.get_attr("data_format").decode()
channel_first = data_format == "NCHW"
self.weights[kernel.layer_name.replace('/', '_')] = numpy.transpose(
kernel.value, (3, 2, 0, 1))
if not channel_first:
in_shape = [in_shape[i] for i in [0, 3, 1, 2]]
strides = [strides[i] for i in [0, 3, 1, 2]]
dilations = [dilations[i] for i in [0, 3, 1, 2]]
else:
self.data_format_propagation(node)
attr = {
"bias_attr": False,
"param_attr": string(kernel.layer_name),
"num_filters": k_size[2],
"filter_size": k_size[0:2],
"stride": strides[2:4],
"dilation": dilations[2:4],
"padding": string(pad_mode),
"output_size": out_shape[1:3]
}
node.fluid_code.add_layer(
"conv2d_transpose", inputs=input, output=node, param_attr=attr)
def Max(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
reduce_idx = self.graph.get_node(node.layer.input[1], copy=True)
assert reduce_idx.layer_type == "Const", "Only support Const parameter[reduce_idx]"
keep_dims = node.get_attr("keep_dims")
dim = reduce_idx.value.tolist()
if input.tf_data_format == "NHWC" and len(input.out_shapes[0]) == 4:
dim = nhwc_dim_to_nchw(input, dim)
attr = {"dim": dim, "keep_dim": keep_dims}
node.fluid_code.add_layer(
"reduce_max", inputs=input, output=node, param_attr=attr)
def Sum(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
reduce_idx = self.graph.get_node(node.layer.input[1], copy=True)
assert reduce_idx.layer_type == "Const", "Only support Const parameter[reduce_idx]"
keep_dims = node.get_attr("keep_dims")
dim = reduce_idx.value.tolist()
if input.tf_data_format == "NHWC" and len(input.out_shapes[0]) == 4:
dim = nhwc_dim_to_nchw(input, dim)
attr = {"dim": dim, "keep_dim": keep_dims}
node.fluid_code.add_layer(
"reduce_sum", inputs=input, output=node, param_attr=attr)
def Cast(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
dtype = node.dtype_map[node.get_attr('DstT')]
attr = {"dtype": string(dtype)}
node.fluid_code.add_layer(
"cast", inputs=input, output=node, param_attr=attr)
def Split(self, node):
dim = self.graph.get_node(node.layer.input[0], copy=True)
input = self.graph.get_node(node.layer.input[1], copy=True)
self.add_omit_nodes(dim.layer_name, node.layer_name)
num_split = node.get_attr('num_split')
dim = dim.value
if input.tf_data_format == "NHWC" and len(input.out_shapes[0]) == 4:
dim = nhwc_dim_to_nchw(input, dim)
attr = {"num_or_sections": num_split, "dim": dim}
node.fluid_code.add_layer(
"split", inputs=input, output=node, param_attr=attr)
def Squeeze(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
squeeze_dims = node.get_attr('squeeze_dims')
if input.tf_data_format == "NHWC" and len(input.out_shapes[0]) == 4:
for i in range(len(squeeze_dims)):
squeeze_dims[i] = nhwc_dim_to_nchw(input, squeeze_dims[i])
attr = {"axes": squeeze_dims}
node.fluid_code.add_layer(
"squeeze", inputs=input, output=node, param_attr=attr)
def Softmax(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
axis = node.get_attr("axis")
if axis is None:
axis = -1 + len(input.out_shapes[0])
if input.tf_data_format == "NHWC" and len(input.out_shapes[0]) == 4:
axis = nhwc_dim_to_nchw(input, axis)
attr = {"axis": axis}
node.fluid_code.add_layer(
"softmax", inputs=input, output=node, param_attr=attr)
def ResizeNearestNeighbor(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
resize_shape = self.graph.get_node(node.layer.input[1], copy=True)
self.add_omit_nodes(resize_shape.layer_name, node.layer_name)
if resize_shape.layer_type == "Const":
resize_shape = resize_shape.value.tolist()
else:
resize_shape = self.decoder.infer_shape_tensor(resize_shape,
node.out_shapes[0])
align_corners = node.get_attr("align_corners")
attr = {"align_corners": align_corners, "out_shape": resize_shape}
node.fluid_code.add_layer(
"resize_nearest", inputs=input, output=node, param_attr=attr)
def ResizeBilinear(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
resize_shape = self.graph.get_node(node.layer.input[1], copy=True)
self.add_omit_nodes(resize_shape.layer_name, node.layer_name)
if resize_shape.layer_type == "Const":
resize_shape = resize_shape.value.tolist()
else:
resize_shape = self.decoder.infer_shape_tensor(resize_shape,
node.out_shapes[0])
align_corners = node.get_attr("align_corners")
attr = {
"align_corners": align_corners,
"out_shape": resize_shape,
"align_mode": 1
}
node.fluid_code.add_layer(
"resize_bilinear", inputs=input, output=node, param_attr=attr)
def GreaterEqual(self, node):
x = self.graph.get_node(node.layer.input[0], copy=True)
y = self.graph.get_node(node.layer.input[1], copy=True)
inputs = {"x": x, "y": y}
node.fluid_code.add_layer(
"greater_equal", inputs=inputs, output=node, param_attr=None)
def RandomUniform(self, node):
shape = self.graph.get_node(node.layer.input[0], copy=True)
self.add_omit_nodes(shape.layer_name, node.layer_name)
if shape.layer_type == "Const":
shape = shape.value.tolist()
else:
shape = self.decoder.infer_shape_tensor(shape)
if len(shape) == 4 and node.tf_data_format == "NHWC":
shape = [shape[i] for i in [0, 3, 1, 2]]
attr = {"shape": shape, "min": 0.0, "max": 0.9999}
if shape[0] < 0:
input = self.batch_node
node.fluid_code.add_layer(
"uniform_random_batch_size_like",
inputs=input,
output=node,
param_attr=attr)
else:
node.fluid_code.add_layer(
"uniform_random", inputs=None, output=node, param_attr=attr)
def SquaredDifference(self, node):
x = self.graph.get_node(node.layer.input[0], copy=True)
y = self.graph.get_node(node.layer.input[1], copy=True)
inputs = {"x": x, "y": y}
node.fluid_code.add_layer(
"elementwise_sub", inputs=inputs, output=node, param_attr=None)
inputs = {"x": node, "y": node}
node.fluid_code.add_layer(
"elementwise_mul", inputs=inputs, output=node, param_attr=None)
x2paddle/x2paddle/op_mapper/tf_op_mapper_nhwc.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from x2paddle.decoder.tf_decoder import TFGraph
from x2paddle.core.op_mapper import OpMapper
from x2paddle.core.util import *
import math
import inspect
import numpy
import sys
# compute padding size for SAME mode
def get_same_padding(in_size, kernel_size, stride):
new_size = int(math.ceil(in_size * 1.0 / stride))
pad_size = (new_size - 1) * stride + kernel_size - in_size
if pad_size < 0:
pad_size = 0
pad0 = int(pad_size / 2)
pad1 = pad_size - pad0
return [pad0, pad1]
class TFOpMapperNHWC(OpMapper):
directly_map_ops = {
'Relu': ['relu'],
'Relu6': ['relu6'],
'Shape': ['shape'],
'Abs': ['abs'],
'Sigmoid': ['sigmoid'],
'Exp': ['exp'],
'Rsqrt': ['rsqrt'],
'Sqrt': ['sqrt'],
'swish_f32': ['swish'],
'Tanh': ['tanh'],
'Softplus': ['softplus'],
'LeakyRelu': ['leaky_relu', {
'alpha': 'alpha'
}],
'Floor': ['floor'],
'Erf': ['erf']
}
elementwise_ops = {
'Add': 'elementwise_add',
'AddV2': 'elementwise_add',
'RealDiv': 'elementwise_div',
'Sub': 'elementwise_sub',
'Maximum': 'elementwise_max',
'Minimum': 'elementwise_min',
'LessEqual': 'less_equal',
'Mul': 'elementwise_mul',
'FloorDiv': 'elementwise_floordiv'
}
def __init__(self, decoder):
super(TFOpMapperNHWC, self).__init__()
self.decoder = decoder
self.graph = decoder.tf_graph
self.weights = dict()
self.batch_node = None
self.omit_nodes = list()
self.used_custom_layers = dict()
not_placeholder = list()
for name in self.graph.input_nodes:
if self.graph.get_node(
name).layer_type != "Placeholder" and self.graph.get_node(
name
).layer_type != "OneShotIterator" and self.graph.get_node(
name).layer_type != "IteratorV2":
not_placeholder.append(name)
for name in not_placeholder:
idx = self.graph.input_nodes.index(name)
del self.graph.input_nodes[idx]
unsupported_ops = set()
sys.stderr.write("Total nodes: {}\n".format(len(self.graph.topo_sort)))
for i, node_name in enumerate(self.graph.topo_sort):
sys.stderr.write("\rConverting node {} ... ".format(i + 1))
node = self.graph.get_node(node_name)
op = node.layer_type
if op in self.directly_map_ops:
if len(unsupported_ops) > 0:
continue
self.directly_map(node)
elif op in self.elementwise_ops:
if len(unsupported_ops) > 0:
continue
self.elementwise_map(node)
elif hasattr(self, op):
if len(unsupported_ops) > 0:
continue
func = getattr(self, op)
try:
func(node)
except Exception as e:
unsupported_ops.add(op)
print(e)
else:
unsupported_ops.add(op)
if len(unsupported_ops) > 0:
print("========= {} OPs are not supported yet ===========".format(
len(unsupported_ops)))
for op in unsupported_ops:
print("========== {} ============".format(op))
sys.exit(-1)
sys.stderr.write("\nDone!\n")
def add_omit_nodes(self, in_node_name, out_node_name):
in_node = self.graph.get_node(in_node_name)
out_node = self.graph.get_node(out_node_name)
index = in_node.outputs.index(out_node_name)
del in_node.outputs[index]
index = out_node.inputs.index(in_node_name)
del out_node.inputs[index]
self.omit_nodes.append(in_node.layer_name)
def directly_map(self, node):
assert node.layer_type in self.directly_map_ops
op_info = self.directly_map_ops[node.layer_type]
input = self.graph.get_node(node.layer.input[0], copy=True)
attr = dict()
for param in op_info[1:]:
tf_param_name = list(param.keys())[0]
pd_param_name = list(param.values())[0]
tf_param = node.get_attr(tf_param_name)
attr[pd_param_name] = tf_param
if len(input.out_shapes[0]) == 4 and op_info[0] != 'shape':
attr1 = {"perm": [0, 3, 1, 2]}
node.fluid_code.add_layer(
'transpose', inputs=input, output=node, param_attr=attr1)
input = node
node.fluid_code.add_layer(
op_info[0], inputs=input, output=node, param_attr=attr)
input = node
attr2 = {"perm": [0, 2, 3, 1]}
node.fluid_code.add_layer(
'transpose', inputs=input, output=node, param_attr=attr2)
else:
node.fluid_code.add_layer(
op_info[0], inputs=input, output=node, param_attr=attr)
def elementwise_map(self, node):
assert node.layer_type in self.elementwise_ops
op_type = self.elementwise_ops[node.layer_type]
x = self.graph.get_node(node.layer.input[0], copy=True)
y = self.graph.get_node(node.layer.input[1], copy=True)
inputs = {"x": x, "y": y}
node.fluid_code.add_layer(
op_type, inputs=inputs, output=node, param_attr=None)
def Placeholder(self, node):
shape = node.out_shapes[0]
assert len(shape) != 0, "Unknown shape of input nodes[{}].".format(
node.layer_name)
dtype = node.dtype
if shape[0] < 0:
self.batch_node = node
attr = {
'dtype': string(dtype),
'shape': shape,
'name': string(node.layer_name),
'append_batch_size': False
}
node.fluid_code.add_layer(
"data", inputs=None, output=node, param_attr=attr)
def Const(self, node):
shape = node.out_shapes[0]
dtype = node.dtype
value = node.value
initializer = "Constant(0.0)"
if len(shape) == 0:
assert value.size == 1, "Unexpected situation happend"
shape = [1]
initializer = "Constant({})".format(value)
self.weights[node.layer_name] = node.value
attr = {
'dtype': string(dtype),
'shape': shape,
'name': string(node.layer_name),
'default_initializer': initializer
}
node.fluid_code.add_layer(
"create_parameter", inputs=None, output=node, param_attr=attr)
def Transpose(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
perm = self.graph.get_node(node.layer.input[1], copy=True)
assert perm.layer_type == "Const", "Perm of transpose OP should be Const"
del self.weights[perm.layer_name.replace('/', '_')]
perm.fluid_code.clear()
perm = perm.value.tolist()
attr = {'perm': perm}
node.fluid_code.add_layer(
"transpose", inputs=input, output=node, param_attr=attr)
def Fill(self, node):
dims = self.graph.get_node(node.layer.input[0], copy=True)
input_value = self.graph.get_node(node.layer.input[1], copy=True)
assert input_value.layer_type == "Const", "Value of fill OP should be Const"
self.add_omit_nodes(input_value.layer_name, node.layer_name)
input_value = input_value.value
input_dtype = string(input_value.dtype)
attr = {'value': input_value, 'dtype': input_dtype}
node.fluid_code.add_layer(
"fill_constant", inputs=dims, output=node, param_attr=attr)
def DepthToSpace(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
block_size = node.get_attr("block_size")
data_format = node.get_attr("data_format").decode()
if data_format == "NHWC":
attr = {"perm": [0, 3, 1, 2]}
node.fluid_code.add_layer(
"transpose", inputs=input, output=input, param_attr=attr)
n, h, w, c = input.out_shapes[0]
attr = {'shape': [0, block_size * block_size, -1, h, w]}
node.fluid_code.add_layer(
"reshape", inputs=input, output=input, param_attr=attr)
attr = {'perm': [0, 2, 1, 3, 4]}
node.fluid_code.add_layer(
"transpose", inputs=input, output=input, param_attr=attr)
attr = {'shape': [0, c, h, w]}
node.fluid_code.add_layer(
"reshape", inputs=input, output=input, param_attr=attr)
attr = {'upscale_factor': block_size}
node.fluid_code.add_layer(
"pixel_shuffle", inputs=input, output=node, param_attr=attr)
if data_format == "NHWC":
attr = {"perm": [0, 2, 3, 1]}
node.fluid_code.add_layer(
"transpose", inputs=node, output=node, param_attr=attr)
def MaxPool(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
k_size = node.get_attr("ksize")
strides = node.get_attr("strides")
data_format = node.get_attr("data_format").decode()
pad_mode = node.get_attr("padding").decode()
channel_first = data_format == "NCHW"
if not channel_first:
attr = {"perm": [0, 3, 1, 2]}
node.fluid_code.add_layer(
"transpose", inputs=input, output=node, param_attr=attr)
strides = [strides[i] for i in [0, 3, 1, 2]]
k_size = [k_size[i] for i in [0, 3, 1, 2]]
input = node
attr = {
"pool_size": k_size[2:4],
"pool_type": string("max"),
"pool_stride": strides[2:4],
"pool_padding": string(pad_mode)
}
node.fluid_code.add_layer(
"pool2d", inputs=input, output=node, param_attr=attr)
if not channel_first:
attr = {"perm": [0, 2, 3, 1]}
node.fluid_code.add_layer(
"transpose", inputs=node, output=node, param_attr=attr)
def Conv2D(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
kernel = self.graph.get_node(node.layer.input[1], copy=True)
self.add_omit_nodes(kernel.layer_name, node.layer_name)
k_size = kernel.out_shapes[0]
strides = node.get_attr("strides")
dilations = node.get_attr("dilations")
data_format = node.get_attr("data_format").decode()
pad_mode = node.get_attr("padding").decode()
channel_first = data_format == "NCHW"
if data_format == "NHWC":
n, h, w, c = input.out_shapes[0]
else:
n, c, h, w = input.out_shapes[0]
if kernel.layer_type == 'Const':
kernel_value = kernel.value
kernel_weight_name = kernel.layer_name.replace('/', '_')
else:
kernel_value = self.decoder.infer_tensor(kernel)
if kernel.layer_type == 'Split':
kernel_weight_name = "{}_{}_kernel".format(node.layer_name,
kernel.layer_name)
else:
kernel_weight_name = kernel.layer_name.replace('/', '_')
self.weights[kernel_weight_name] = numpy.transpose(kernel_value,
(3, 2, 0, 1))
if not channel_first:
strides = [strides[i] for i in [0, 3, 1, 2]]
dilations = [dilations[i] for i in [0, 3, 1, 2]]
attr = {"perm": [0, 3, 1, 2]}
node.fluid_code.add_layer(
"transpose", inputs=input, output=node, param_attr=attr)
input = node
attr = {
"bias_attr": False,
"param_attr": string(kernel_weight_name),
"num_filters": k_size[3],
"filter_size": k_size[0:2],
"stride": strides[2:4],
"dilation": dilations[2:4],
"padding": string(pad_mode)
}
if hasattr(node, 'dilation') and attr['dilation'] == [1, 1]:
if len(node.dilation) == 1:
attr['dilation'] = [1, node.dilation[0]]
if c == -1:
reshape_attr = {"shape": [0, k_size[2], 0, 0]}
node.fluid_code.add_layer(
"reshape", inputs=input, output=input, param_attr=reshape_attr)
node.fluid_code.add_layer(
"conv2d", inputs=input, output=node, param_attr=attr)
if not channel_first:
attr = {"perm": [0, 2, 3, 1]}
node.fluid_code.add_layer(
"transpose", inputs=node, output=node, param_attr=attr)
def BiasAdd(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
bias = self.graph.get_node(node.layer.input[1], copy=True)
inputs = {"x": input, "y": bias}
node.fluid_code.add_layer(
"elementwise_add", inputs=inputs, output=node, param_attr=None)
def FusedBatchNorm(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
gamma = self.graph.get_node(node.layer.input[1], copy=True)
beta = self.graph.get_node(node.layer.input[2], copy=True)
moving_mean = self.graph.get_node(node.layer.input[3], copy=True)
moving_var = self.graph.get_node(node.layer.input[4], copy=True)
data_format = node.get_attr("data_format").decode()
channel_first = data_format == "NCHW"
assert gamma.layer_type == "Const"
assert beta.layer_type == "Const"
assert moving_mean.layer_type == "Const"
assert moving_var.layer_type == "Const"
self.add_omit_nodes(gamma.layer_name, node.layer_name)
self.add_omit_nodes(beta.layer_name, node.layer_name)
self.add_omit_nodes(moving_mean.layer_name, node.layer_name)
self.add_omit_nodes(moving_var.layer_name, node.layer_name)
if not channel_first:
attr = {"perm": [0, 3, 1, 2]}
node.fluid_code.add_layer(
"transpose", inputs=input, output=node, param_attr=attr)
input = node
attr = {
"epsilon": node.get_attr("epsilon"),
"param_attr": string(gamma.layer_name),
"bias_attr": string(beta.layer_name),
"moving_mean_name": string(moving_mean.layer_name),
"moving_variance_name": string(moving_var.layer_name),
"is_test": True
}
node.fluid_code.add_layer(
"batch_norm", inputs=input, output=node, param_attr=attr)
if not channel_first:
attr = {"perm": [0, 2, 3, 1]}
node.fluid_code.add_layer(
"transpose", inputs=node, output=node, param_attr=attr)
def DepthwiseConv2dNative(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
kernel = self.graph.get_node(node.layer.input[1], copy=True)
assert kernel.layer_type == "Const", "Kernel of DepthwiseConv2DNative should be Const"
self.add_omit_nodes(kernel.layer_name, node.layer_name)
in_shape = input.out_shapes[0]
k_size = kernel.out_shapes[0]
strides = node.get_attr("strides")
dilations = node.get_attr("dilations")
data_format = node.get_attr("data_format").decode()
pad_mode = node.get_attr("padding").decode()
channel_first = data_format == "NCHW"
self.weights[kernel.layer_name.replace('/', '_')] = numpy.transpose(
kernel.value, (2, 3, 0, 1))
if not channel_first:
in_shape = [in_shape[i] for i in [0, 3, 1, 2]]
strides = [strides[i] for i in [0, 3, 1, 2]]
dilations = [dilations[i] for i in [0, 3, 1, 2]]
attr = {"perm": [0, 3, 1, 2]}
node.fluid_code.add_layer(
"transpose", inputs=input, output=node, param_attr=attr)
input = node
attr = {
"bias_attr": False,
"param_attr": string(kernel.layer_name),
"num_filters": in_shape[1],
"filter_size": k_size[0:2],
"stride": strides[2:4],
"dilation": dilations[2:4],
"groups": k_size[3] * in_shape[1],
"use_cudnn": False,
"padding": string(pad_mode)
}
node.fluid_code.add_layer(
"conv2d", inputs=input, output=node, param_attr=attr)
if not channel_first:
attr = {"perm": [0, 2, 3, 1]}
node.fluid_code.add_layer(
"transpose", inputs=node, output=node, param_attr=attr)
def Reshape(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
param = self.graph.get_node(node.layer.input[1], copy=True)
if param.layer_type == "Const":
self.add_omit_nodes(param.layer_name, node.layer_name)
shape = param.value.tolist()
else:
shape = param
inputs = {"x": input, "shape": shape}
node.fluid_code.add_layer(
"reshape", inputs=inputs, output=node, param_attr=None)
if param.layer_type != "Const":
out_shape = numpy.array(node.out_shapes[0])
if (out_shape > 0).any():
out_shape[out_shape < 0] = 0
attr = {'shape': out_shape.tolist()}
node.fluid_code.add_layer(
"reshape", inputs=node, output=node, param_attr=attr)
def AvgPool(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
k_size = node.get_attr("ksize")
strides = node.get_attr("strides")
data_format = node.get_attr("data_format").decode()
pad_mode = node.get_attr("padding").decode()
channel_first = data_format == "NCHW"
if not channel_first:
strides = [strides[i] for i in [0, 3, 1, 2]]
k_size = [k_size[i] for i in [0, 3, 1, 2]]
attr = {"perm": [0, 3, 1, 2]}
node.fluid_code.add_layer(
"transpose", inputs=input, output=node, param_attr=attr)
input = node
attr = {
"pool_size": k_size[2:4],
"pool_type": string("avg"),
"pool_stride": strides[2:4],
"pool_padding": string(pad_mode)
}
node.fluid_code.add_layer(
"pool2d", inputs=input, output=node, param_attr=attr)
if not channel_first:
attr = {"perm": [0, 2, 3, 1]}
node.fluid_code.add_layer(
"transpose", inputs=node, output=node, param_attr=attr)
def SplitV(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
num_sections = self.graph.get_node(node.layer.input[1], copy=True)
dim = self.graph.get_node(node.layer.input[2], copy=True)
assert num_sections.layer_type == "Const"
assert dim.layer_type == "Const"
self.add_omit_nodes(num_sections.layer_name, node.layer_name)
self.add_omit_nodes(dim.layer_name, node.layer_name)
dim = dim.value
attr = {
"num_or_sections": num_sections.value.tolist(),
"dim": dim.value
}
node.fluid_code.add_layer(
"split", inputs=input, output=node, param_attr=attr)
def ConcatV2(self, node):
inputs = [
self.graph.get_node(
name, copy=True) for name in node.layer.input[:-1]
]
axis = self.graph.get_node(node.layer.input[-1], copy=True)
assert axis.layer_type == "Const"
self.add_omit_nodes(axis.layer_name, node.layer_name)
axis = axis.value
if axis < 0:
axis += len(inputs[0].out_shapes[0])
attr = {"axis": axis}
node.fluid_code.add_layer(
"concat", inputs=inputs, output=node, param_attr=attr)
def Tile(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
expand_times = self.graph.get_node(node.layer.input[1], copy=True)
if expand_times.layer_type == "Const":
self.add_omit_nodes(expand_times.layer_name, node.layer_name)
expand_times = expand_times.value.tolist()
else:
expand_times = expand_times
inputs = {"x": input, "expand_times": expand_times}
node.fluid_code.add_layer(
"expand", inputs=inputs, output=node, param_attr=None)
def Pack(self, node):
inputs = [
self.graph.get_node(
name, copy=True) for name in node.layer.input
]
reshape_shape = list()
for input_node in inputs:
k_size = input_node.out_shapes[0]
if len(k_size) and k_size[-1] != -1:
reshape_shape = [0] * len(k_size)
reshape_shape[-1] = k_size[-1]
break
if len(reshape_shape):
for i, input_node in enumerate(inputs):
node.fluid_code.add_layer(
"reshape",
inputs=input_node,
output='tmp_{}'.format(i),
param_attr={"shape": reshape_shape})
axis = node.get_attr("axis")
attr = {"axis": axis}
if len(reshape_shape):
inputs = ['tmp_{}'.format(i) for i in range(len(inputs))]
node.fluid_code.add_layer(
"stack", inputs=inputs, output=node, param_attr=attr)
def Pad(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
paddings = self.graph.get_node(node.layer.input[1], copy=True)
assert paddings.layer_type == "Const", "Padding should be Const"
self.add_omit_nodes(paddings.layer_name, node.layer_name)
paddings = paddings.value.flatten().tolist()
data_format = input.tf_data_format
if len(input.out_shapes[0]) == 4:
new_padding = None
if input.tf_data_format == "NHWC":
if paddings[0] + paddings[1] + paddings[6] + paddings[7] == 0:
new_padding = paddings[2:6]
else:
if paddings[0] + paddings[1] + paddings[2] + paddings[3] == 0:
new_padding = paddings[4:]
if new_padding is not None:
if input.tf_data_format == "NHWC":
attr = {"perm": [0, 3, 1, 2]}
node.fluid_code.add_layer(
"transpose", inputs=input, output=node, param_attr=attr)
input = node
attr = {"paddings": new_padding}
node.fluid_code.add_layer(
"pad2d", inputs=input, output=node, param_attr=attr)
if input.tf_data_format == "NHWC":
attr = {"perm": [0, 2, 3, 1]}
node.fluid_code.add_layer(
"transpose", inputs=node, output=node, param_attr=attr)
return
attr = {"paddings": paddings}
node.fluid_code.add_layer(
"pad", inputs=input, output=node, param_attr=attr)
def Range(self, node):
start = self.graph.get_node(node.layer.input[0], copy=True)
limit = self.graph.get_node(node.layer.input[1], copy=True)
delta = self.graph.get_node(node.layer.input[2], copy=True)
if start.layer_type == "Const":
self.add_omit_nodes(start.layer_name, node.layer_name)
start = start.value
if limit.layer_type == "Const":
self.add_omit_nodes(limit.layer_name, node.layer_name)
limit = limit.value
if delta.layer_type == "Const":
self.add_omit_nodes(delta.layer_name, node.layer_name)
delta = delta.value
dtype = node.dtype
inputs = {
"start": start,
"end": limit,
"step": delta,
}
attr = {"dtype": string(node.dtype)}
node.fluid_code.add_layer(
"range", inputs=inputs, output=node, param_attr=attr)
def Mean(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
reduce_idx = self.graph.get_node(node.layer.input[1], copy=True)
assert reduce_idx.layer_type == "Const", "Only support Const parameter[reduce_idx]"
dims = reduce_idx.value.tolist()
keep_dims = node.get_attr("keep_dims")
attr = {"dim": dims, "keep_dim": keep_dims}
node.fluid_code.add_layer(
"reduce_mean", inputs=input, output=node, param_attr=attr)
def MatMul(self, node):
x = self.graph.get_node(node.layer.input[0], copy=True)
y = self.graph.get_node(node.layer.input[1], copy=True)
transpose_a = node.get_attr('transpose_a')
transpose_b = node.get_attr('transpose_b')
inputs = {"x": x, "y": y}
# fix paddle shape infer problem
# should be removed after paddle 1.6
if x.out_shapes[0][-1] < 0 and y.out_shapes[0][0] > 0:
shape = x.out_shapes[0]
shape[-1] = y.out_shapes[0][0]
attr = {"shape": shape}
node.fluid_code.add_layer(
"reshape", inputs=x, output=x, param_attr=attr)
if transpose_a is None:
transpose_a = node.get_attr('adj_x')
if transpose_b is None:
transpose_b = node.get_attr('adj_y')
attr = {"transpose_x": transpose_a, "transpose_y": transpose_b}
node.fluid_code.add_layer(
"matmul", inputs=inputs, output=node, param_attr=attr)
def BatchMatMul(self, node):
return self.MatMul(node)
def BatchMatMulV2(self, node):
return self.MatMul(node)
def ArgMax(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
axis = self.graph.get_node(node.layer.input[1], copy=True)
assert axis.layer_type == "Const", "ArgMax only support Const parameter"
self.add_omit_nodes(axis.layer_name, node.layer_name)
axis = axis.value
attr = {"axis": axis}
node.fluid_code.add_layer(
"argmax", inputs=input, output=node, param_attr=attr)
def StridedSlice(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
begin = self.graph.get_node(node.layer.input[1], copy=True)
end = self.graph.get_node(node.layer.input[2], copy=True)
strides = self.graph.get_node(node.layer.input[3], copy=True)
assert begin.layer_type == "Const"
assert end.layer_type == "Const"
assert strides.layer_type == "Const"
self.add_omit_nodes(begin.layer_name, node.layer_name)
self.add_omit_nodes(end.layer_name, node.layer_name)
self.add_omit_nodes(strides.layer_name, node.layer_name)
strides = strides.value.tolist()
assert len(set(strides)) == 1 and strides[
0] == 1, "Only support strides be 1 in StridedSlice OP"
begin = begin.value.tolist()
end = end.value.tolist()
for i in range(len(end)):
if end[i] == 0:
end[i] = 999999
begin_mask = node.get_attr('begin_mask')
end_mask = node.get_attr('end_mask')
ellipsis_mask = node.get_attr('ellipsis_mask')
new_axis_mask = node.get_attr('new_axis_mask')
shrink_axis_mask = node.get_attr('shrink_axis_mask')
assert ellipsis_mask == 0, "(OP:{} Name:{})Only support ellipsis_mask be 0[now: {}] n StridedSlice OP".format(
node.layer_type, node.layer.name, ellipsis_mask)
# TODO codes without validation
# Use it carefully
new_begin = list()
new_end = list()
new_axes = list()
shrink_axes = list()
for i, item in enumerate(begin):
mask = (new_axis_mask >> i) & 1
if mask != 0:
new_axes.append(i)
continue
mask = (shrink_axis_mask >> i) & 1
if mask != 0:
shrink_axes.append(i)
mask = (begin_mask >> i) & 1
if mask != 0:
new_begin.append(0)
else:
new_begin.append(item)
mask = (end_mask >> i) & 1
if mask != 0:
new_end.append(999999)
else:
new_end.append(end[i])
attr = {
"axes": [i for i in range(len(new_begin))],
"starts": new_begin,
"ends": new_end
}
node.fluid_code.add_layer(
"slice", inputs=input, output=node, param_attr=attr)
if len(new_axes) > 0:
attr = {"axes": new_axes}
node.fluid_code.add_layer(
"unsqueeze", inputs=node, output=node, param_attr=attr)
if len(shrink_axes) > 0:
if len(input.out_shapes[0]) + len(new_axes) <= 1:
pass
else:
attr = {"axes": shrink_axes}
node.fluid_code.add_layer(
"squeeze", inputs=node, output=node, param_attr=attr)
def Slice(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
begin = self.graph.get_node(node.layer.input[1], copy=True)
size = self.graph.get_node(node.layer.input[2], copy=True)
if begin.layer_type == "Const":
self.add_omit_nodes(begin.layer_name, node.layer_name)
begin = begin.value.tolist()
else:
begin = self.decoder.infer_tensor(begin).tolist()
# shape = begin.out_shapes[0]
# attr = {"shape": shape}
# node.fluid_code.add_layer(
# "reshape", inputs=begin, output=begin, param_attr=attr)
if size.layer_type == "Const":
self.add_omit_nodes(size.layer_name, node.layer_name)
size = size.value.tolist()
else:
size = size
shape = size.out_shapes[0]
attr = {"shape": shape}
node.fluid_code.add_layer(
"reshape", inputs=size, output=size, param_attr=attr)
inputs = {"x": input, "offsets": begin, "shape": size}
node.fluid_code.add_layer(
"crop_tensor", inputs=inputs, output=node, param_attr=None)
def Conv2DBackpropInput(self, node):
out_shape = self.graph.get_node(node.layer.input[0], copy=True)
kernel = self.graph.get_node(node.layer.input[1], copy=True)
input = self.graph.get_node(node.layer.input[2], copy=True)
assert kernel.layer_type == "Const", "Kernel of Conv2DBackpropInput should be Const"
self.add_omit_nodes(kernel.layer_name, node.layer_name)
self.add_omit_nodes(out_shape.layer_name, node.layer_name)
if out_shape.layer_type == "Const":
out_shape = out_shape.value.tolist()
else:
out_shape = self.decoder.infer_shape_tensor(out_shape,
node.out_shapes[0])
in_shape = input.out_shapes[0]
if in_shape.count(-1) > 2:
in_shape = self.decoder.infer_tensor(input).shape
k_size = kernel.out_shapes[0]
if k_size.count(-1) > 2:
k_size = self.decoder.infer_tensor(kernel).shape
pad_mode = node.get_attr("padding").decode()
strides = node.get_attr("strides")
dilations = node.get_attr("dilations")
data_format = node.get_attr("data_format").decode()
channel_first = data_format == "NCHW"
self.weights[kernel.layer_name.replace('/', '_')] = numpy.transpose(
kernel.value, (3, 2, 0, 1))
if not channel_first:
in_shape = [in_shape[i] for i in [0, 3, 1, 2]]
strides = [strides[i] for i in [0, 3, 1, 2]]
dilations = [dilations[i] for i in [0, 3, 1, 2]]
attr = {"perm": [0, 3, 1, 2]}
node.fluid_code.add_layer(
"transpose", inputs=input, output=node, param_attr=attr)
input = node
else:
self.graph.data_format_propagation(node)
attr = {
"bias_attr": False,
"param_attr": string(kernel.layer_name),
"num_filters": k_size[2],
"filter_size": k_size[0:2],
"stride": strides[2:4],
"dilation": dilations[2:4],
"padding": string(pad_mode),
"output_size": out_shape[1:3]
}
node.fluid_code.add_layer(
"conv2d_transpose", inputs=input, output=node, param_attr=attr)
if not channel_first:
attr = {"perm": [0, 2, 3, 1]}
node.fluid_code.add_layer(
"transpose", inputs=node, output=node, param_attr=attr)
def Max(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
reduce_idx = self.graph.get_node(node.layer.input[1], copy=True)
assert reduce_idx.layer_type == "Const", "Only support Const parameter[reduce_idx]"
keep_dims = node.get_attr("keep_dims")
dim = reduce_idx.value.tolist()
attr = {"dim": dim, "keep_dim": keep_dims}
node.fluid_code.add_layer(
"reduce_max", inputs=input, output=node, param_attr=attr)
def Sum(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
reduce_idx = self.graph.get_node(node.layer.input[1], copy=True)
assert reduce_idx.layer_type == "Const", "Only support Const parameter[reduce_idx]"
keep_dims = node.get_attr("keep_dims")
dim = reduce_idx.value.tolist()
attr = {"dim": dim, "keep_dim": keep_dims}
node.fluid_code.add_layer(
"reduce_sum", inputs=input, output=node, param_attr=attr)
def Cast(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
dtype = node.dtype_map[node.get_attr('DstT')]
attr = {"dtype": string(dtype)}
node.fluid_code.add_layer(
"cast", inputs=input, output=node, param_attr=attr)
def Split(self, node):
dim = self.graph.get_node(node.layer.input[0], copy=True)
input = self.graph.get_node(node.layer.input[1], copy=True)
assert dim.layer_type == "Const"
self.add_omit_nodes(dim.layer_name, node.layer_name)
num_split = node.get_attr('num_split')
dim = dim.value
attr = {"num_or_sections": num_split, "dim": dim}
node.fluid_code.add_layer(
"split", inputs=input, output=node, param_attr=attr)
def Squeeze(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
squeeze_dims = node.get_attr('squeeze_dims')
attr = {"axes": squeeze_dims}
node.fluid_code.add_layer(
"squeeze", inputs=input, output=node, param_attr=attr)
def Softmax(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
axis = node.get_attr("axis")
attr = {"axis": axis}
node.fluid_code.add_layer(
"softmax", inputs=input, output=node, param_attr=attr)
def ResizeNearestNeighbor(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
resize_shape = self.graph.get_node(node.layer.input[1], copy=True)
if resize_shape.layer_type == "Const":
self.add_omit_nodes(resize_shape.layer_name, node.layer_name)
resize_shape = resize_shape.value.tolist()
else:
resize_shape = resize_shape
shape = resize_shape.out_shapes[0]
attr = {"shape": shape}
node.fluid_code.add_layer(
"reshape",
inputs=resize_shape,
output=resize_shape,
param_attr=attr)
align_corners = node.get_attr("align_corners")
attr = {"perm": [0, 3, 1, 2]}
node.fluid_code.add_layer(
"transpose", inputs=input, output=node, param_attr=attr)
inputs = {"input": node, "out_shape": resize_shape}
attr = {"align_corners": align_corners}
node.fluid_code.add_layer(
"resize_nearest", inputs=inputs, output=node, param_attr=attr)
attr = {"perm": [0, 2, 3, 1]}
node.fluid_code.add_layer(
"transpose", inputs=node, output=node, param_attr=attr)
def ResizeBilinear(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
resize_shape = self.graph.get_node(node.layer.input[1], copy=True)
if resize_shape.layer_type == "Const":
self.add_omit_nodes(resize_shape.layer_name, node.layer_name)
resize_shape = resize_shape.value.tolist()
else:
shape = resize_shape.out_shapes[0]
attr = {"shape": shape}
node.fluid_code.add_layer(
"reshape",
inputs=resize_shape,
output=resize_shape,
param_attr=attr)
align_corners = node.get_attr("align_corners")
attr = {"perm": [0, 3, 1, 2]}
node.fluid_code.add_layer(
"transpose", inputs=input, output=node, param_attr=attr)
inputs = {"input": node, "out_shape": resize_shape}
attr = {
#"out_shape": resize_shape,
"align_corners": align_corners,
"align_mode": 1
}
node.fluid_code.add_layer(
"resize_bilinear", inputs=inputs, output=node, param_attr=attr)
attr = {"perm": [0, 2, 3, 1]}
node.fluid_code.add_layer(
"transpose", inputs=node, output=node, param_attr=attr)
def GreaterEqual(self, node):
x = self.graph.get_node(node.layer.input[0], copy=True)
y = self.graph.get_node(node.layer.input[1], copy=True)
inputs = {"x": x, "y": y}
node.fluid_code.add_layer(
"greater_equal", inputs=inputs, output=node, param_attr=None)
def RandomUniform(self, node):
shape = self.graph.get_node(node.layer.input[0], copy=True)
if shape.layer_type == "Const":
self.add_omit_nodes(shape.layer_name, node.layer_name)
shape = shape.value.tolist()
else:
shape = shape
attr = {"min": 0.0, "max": 0.9999}
node.fluid_code.add_layer(
"uniform_random", inputs=shape, output=node, param_attr=attr)
def SquaredDifference(self, node):
x = self.graph.get_node(node.layer.input[0], copy=True)
y = self.graph.get_node(node.layer.input[1], copy=True)
inputs = {"x": x, "y": y}
node.fluid_code.add_layer(
"elementwise_sub", inputs=inputs, output=node, param_attr=None)
inputs = {"x": node, "y": node}
node.fluid_code.add_layer(
"elementwise_mul", inputs=inputs, output=node, param_attr=None)
def ExpandDims(self, node):
x = self.graph.get_node(node.layer.input[0], copy=True)
y = self.graph.get_node(node.layer.input[1], copy=True)
if y.layer_type == 'Const':
self.add_omit_nodes(y.layer_name, node.layer_name)
dim = y.value.tolist()
if not isinstance(dim, list):
dim = [dim]
attr = {'axes': dim}
else:
attr = {'axes': y}
node.fluid_code.add_layer(
"unsqueeze", inputs=x, output=node, param_attr=attr)
def BatchToSpaceND(self, node):
x = self.graph.get_node(node.layer.input[0], copy=True)
y = self.graph.get_node(node.layer.input[1], copy=True)
if hasattr(node, 'skip') and node.skip:
node.fluid_code.add_layer(
"=", inputs=x, output=node, param_attr=None)
else:
raise Exception("BatchToSpaceND is not supported")
def SpaceToBatchND(self, node):
x = self.graph.get_node(node.layer.input[0], copy=True)
y = self.graph.get_node(node.layer.input[1], copy=True)
if hasattr(node, 'skip') and node.skip:
node.fluid_code.add_layer(
"=", inputs=x, output=node, param_attr=None)
else:
raise Exception("SpaceToBatchND is not supported")
def OneHot(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
depth = self.graph.get_node(node.layer.input[1], copy=True)
on_value = self.graph.get_node(node.layer.input[2], copy=True)
off_value = self.graph.get_node(node.layer.input[3], copy=True)
assert depth.layer_type == 'Const', 'Parameter depth should be Const in OneHot'
assert on_value.layer_type == 'Const', 'Parameter on_value should be Const in OneHot'
assert off_value.layer_type == 'Const', 'Parameter off_value should be Const in OneHot'
self.add_omit_nodes(depth.layer_name, node.layer_name)
self.add_omit_nodes(on_value.layer_name, node.layer_name)
self.add_omit_nodes(off_value.layer_name, node.layer_name)
depth = depth.value
on_value = on_value.value
off_value = off_value.value
assert math.fabs(on_value -
1.0) < 1e-06, "on_value should be 1 in OneHot"
assert math.fabs(off_value -
0.0) < 1e-06, "off_value should be 0 in OneHot"
attr = {'depth': depth}
node.fluid_code.add_layer(
"one_hot",
inputs=input,
output=node,
param_attr=attr,
use_fluid=True)
def Pow(self, node):
x = self.graph.get_node(node.layer.input[0], copy=True)
factor = self.graph.get_node(node.layer.input[1], copy=True)
self.add_omit_nodes(factor.layer_name, node.layer_name)
if factor.layer_type == 'Const':
factor = factor.value.tolist()
else:
factor = self.decoder.infer_tensor(factor)
attr = {'factor': factor}
node.fluid_code.add_layer("pow", inputs=x, output=node, param_attr=attr)
def All(self, node):
input = self.graph.get_node(node.layer.input[0], copy=True)
reduce_idx = self.graph.get_node(node.layer.input[1], copy=True)
self.add_omit_nodes(reduce_idx.layer_name, node.layer_name)
assert reduce_idx.layer_type == "Const", "Only support Const parameter[reduce_idx]"
dims = reduce_idx.value.tolist()
keep_dims = node.get_attr("keep_dims")
attr = {"dim": dims, "keep_dim": keep_dims}
node.fluid_code.add_layer(
"reduce_all", inputs=input, output=node, param_attr=attr)
def GatherV2(self, node):
embeddings = self.graph.get_node(node.layer.input[0], copy=True)
index = self.graph.get_node(node.layer.input[1], copy=True)
axis = self.graph.get_node(node.layer.input[2], copy=True)
self.add_omit_nodes(axis.layer_name, node.layer_name)
assert axis.layer_type == 'Const', "Only support Const parameter[axis]"
axis = axis.value.tolist()
assert axis == 0, "Only support axis=0 in GatherV2 OP"
attr = {'overwrite': False}
embeddings_shape = embeddings.out_shapes[0][-1]
reshape_list = list()
reshape_name = index.layer_name
if len(index.out_shapes[0]) != 1:
reshape_list = index.out_shapes[0]
reshape_attr = {"shape": [-1]}
reshape_name = "{}_reshape".format(index.layer_name)
node.fluid_code.add_layer(
"reshape",
inputs=index,
output=reshape_name,
param_attr=reshape_attr)
inputs = {'input': embeddings, 'index': reshape_name}
node.fluid_code.add_layer(
"gather", inputs=inputs, output=node, param_attr=attr)
if len(index.out_shapes[0]) != 1:
reshape_attr = {"shape": reshape_list + [embeddings_shape]}
node.fluid_code.add_layer(
"reshape", inputs=node, output=node, param_attr=reshape_attr)
def OneShotIterator(self, node):
return self.Placeholder(node)
def IteratorV2(self, node):
dtype_map = {
1: "float32",
3: "int32",
4: "uint8",
9: "int64",
10: "bool"
}
shapes = node.out_shapes
dtypes = node.layer.attr['output_types'].list.type
node.fluid_code.add_note("{} = [0] * {}".format(node.layer_name,
len(shapes)))
for i, shape in enumerate(shapes):
attr = {
'dtype': string(dtype_map[dtypes[i]]),
'shape': shape,
'name': string("{}_{}".format(node.layer_name, i)),
'append_batch_size': False
}
output = "{}[{}]".format(node.layer_name, i)
node.fluid_code.add_layer(
"data", inputs=None, output=output, param_attr=attr)
x2paddle/x2paddle/optimizer/caffe_optimizer.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from x2paddle.decoder.caffe_decoder import CaffeGraph
from x2paddle.core.util import *
class CaffeOptimizer(object):
layers_with_act = ['Convolution', 'Deconvolution', 'InnerProduct']
activation_ops = ['ReLU', 'Sigmoid']
def __init__(self, mapper):
self.graph = mapper.graph
def merge_bn_scale(self):
for node_name in self.graph.topo_sort:
node = self.graph.get_node(node_name)
if node.layer_type == 'Scale':
parent_node = self.graph.get_bottom_node(node, idx=0)
if parent_node.layer_type == 'BatchNorm':
is_delete_node = True if len(
parent_node.outputs) == 1 else False
parent_fluid_layer = parent_node.fluid_code.layers[0]
input = parent_fluid_layer.inputs
parent_param_attr = parent_fluid_layer.param_attr
parent_param_attr['param_attr'] = string(node.layer_name +
'_scale')
parent_param_attr['bias_attr'] = string(node.layer_name +
'_offset')
if is_delete_node:
parent_node.fluid_code.clear()
node.fluid_code.clear()
node.fluid_code.add_layer(
"batch_norm",
inputs=input,
output=node,
param_attr=parent_param_attr)
def merge_op_activation(self):
for node_name in self.graph.topo_sort:
node = self.graph.get_node(node_name)
if node.layer_type in self.activation_ops:
parent_node = self.graph.get_bottom_node(node, idx=0)
if parent_node.layer_type in self.layers_with_act:
is_delete_node = True if len(
parent_node.outputs) == 1 else False
parent_fluid_layer = parent_node.fluid_code.layers[0]
input = parent_fluid_layer.inputs
parent_param_attr = parent_fluid_layer.param_attr
parent_param_attr['act'] = string(node.layer_type.lower())
op = parent_fluid_layer.op
if is_delete_node:
parent_node.fluid_code.clear()
node.fluid_code.clear()
node.fluid_code.add_layer(
op,
inputs=input,
output=node,
param_attr=parent_param_attr)
x2paddle/x2paddle/optimizer/onnx_optimizer.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# TODO useless node remove
class ONNXOptimizer(object):
def __init__(self, op_mapper):
self.op_mapper = op_mapper
self.graph = op_mapper.graph
def delete_redundance_code(self):
for node_name in self.graph.topo_sort:
if node_name in self.op_mapper.omit_nodes:
node = self.graph.get_node(node_name)
omit_freq = self.op_mapper.omit_nodes.count(node_name)
if len(node.outputs) <= omit_freq:
node.fluid_code.clear()
x2paddle/x2paddle/optimizer/tf_optimizer.py 0 → 100644
浏览文件 @ cf7f9b88
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# TODO useless node remove
from x2paddle.op_mapper.tf_op_mapper import TFOpMapper
from x2paddle.core.fluid_code import Layer
from x2paddle.core.util import *
import six
import numpy
import copy as cp
def exist_act(node):
for layer in node.fluid_code.layers:
if layer.param_attr is not None:
act = layer.param_attr.get("act", None)
if act is not None:
return True
return False
class TFOptimizer(object):
activation_ops = {
'Relu': 'relu',
'Sigmoid': 'sigmoid',
'Relu6': 'relu6',
'swish_f32': 'swish'
}
layers_with_act = [
'Conv2D', 'BiasAdd', 'DepthwiseConv2dNative', 'Conv2DBackpropInput',
'FusedBatchNorm', 'conv2d', 'elementwise_add', 'conv2d_transpose',
'batch_norm'
]
layers_with_bias = [
'Conv2D', 'DepthwiseConv2dNative', 'Conv2DBackpropInput', 'conv2d',
'conv2d_transpose'
]
def __init__(self, op_mapper):
self.op_mapper = op_mapper
self.graph = op_mapper.graph
def delete_redundance_code(self):
for node_name in self.graph.topo_sort:
if node_name in self.op_mapper.omit_nodes:
node = self.graph.get_node(node_name)
if node is None:
continue
omit_freq = self.op_mapper.omit_nodes.count(node_name)
if len(node.outputs) <= omit_freq:
node.fluid_code.clear()
# remove node from graph
input_names = node.inputs
output_names = node.outputs
for in_name in input_names:
in_node = self.graph.get_node(in_name)
index = in_node.outputs.index(node_name)
del in_node.outputs[index]
for out_name in output_names:
out_node = self.graph.get_node(out_name)
index = out_node.inputs.index(node_name)
del out_node.inputs[index]
del self.graph.node_map[node_name]
def strip_graph(self):
visited_nodes = set()
def visit(node_name):
if node_name in visited_nodes:
return
visited_nodes.add(node_name)
input_names = self.graph.get_node(node_name).inputs
for in_name in input_names:
visit(in_name)
for node_name in self.graph.output_nodes:
visit(node_name)
for i, node_name in enumerate(self.graph.topo_sort):
if node_name not in visited_nodes:
node = self.graph.get_node(node_name)
if node is None:
continue
input_names = node.inputs
output_names = node.outputs
for in_name in input_names:
in_node = self.graph.get_node(in_name)
index = in_node.outputs.index(node_name)
del in_node.outputs[index]
for out_name in output_names:
out_node = self.graph.get_node(out_name)
index = out_node.inputs.index(node_name)
del out_node.inputs[index]
del self.graph.node_map[node_name]
def optimize_elementwise_op(self):
elementwise_ops = [
'Sub', 'Add', 'RealDiv', 'Maximum', 'Mul', 'FloorDiv',
'GreaterEqual'
]
revertable_ops = ['Add', 'Mul']
for node_name in self.graph.topo_sort:
node = self.graph.get_node(node_name)
if node is None:
continue
if node.layer_type in elementwise_ops:
if len(node.fluid_code.layers) != 2:
continue
if node.fluid_code.layers[0].op != "expand":
continue
expand_out = node.fluid_code.layers[0].output
expand_in = node.fluid_code.layers[0].inputs
expand_times = node.fluid_code.layers[0].param_attr[
"expand_times"]
x = node.fluid_code.layers[1].inputs["x"]
y = node.fluid_code.layers[1].inputs["y"]
if isinstance(
x,
six.string_types) and node.layer_type in revertable_ops:
node.fluid_code.layers[1].inputs["y"] = x
node.fluid_code.layers[1].inputs["x"] = y
x = node.fluid_code.layers[1].inputs["x"]
y = expand_in
elif isinstance(y, six.string_types):
y = expand_in
else:
continue
x_shape = x.out_shapes[0]
y_shape = y.out_shapes[0]
if len(x_shape) != len(y_shape):
continue
if len(x_shape) == 4:
x_shape = [x_shape[i] for i in [0, 3, 1, 2]]
y_shape = [y_shape[i] for i in [0, 3, 1, 2]]
continue_flag = True
for i in range(len(x_shape)):
if y_shape[-1 * (i + 1)] == 1 and continue_flag:
expand_times[-1 * (i + 1)] = 1
else:
continue_flag = False
if expand_times.count(1) == len(expand_times):
node.fluid_code.layers[1].inputs["y"] = expand_in
del node.fluid_code.layers[0]
def merge_activation(self):
act_nodes = list()
for node_name in self.graph.topo_sort:
node = self.graph.get_node(node_name)
if node is None:
continue
if node.layer_type in self.activation_ops:
act_nodes.append(node_name)
for act_node_name in act_nodes:
node = self.graph.get_node(act_node_name)
input = self.graph.get_node(node.inputs[0])
if input.layer_type not in self.layers_with_act:
continue
if len(input.fluid_code.layers) == 0:
continue
if 'act' in input.fluid_code.layers[
-1].param_attr and input.fluid_code.layers[-1].param_attr[
'act'] is not None:
continue
if len(input.outputs) != 1:
continue
index = -1
for i in range(len(input.fluid_code.layers)):
if input.fluid_code.layers[i].op in self.layers_with_act:
index = i
break
input.fluid_code.layers[index].param_attr['act'] = string(
self.activation_ops[node.layer_type])
input.fluid_code.layers[-1].output = node.fluid_code.layers[
0].output
self.graph.remove_node(act_node_name)
def merge_bias(self):
for node_name in self.graph.topo_sort:
node = self.graph.get_node(node_name)
if node is None:
continue
if node.layer_type == "BiasAdd":
input = self.graph.get_node(node.inputs[0])
if input.layer_type not in self.layers_with_bias:
continue
if len(input.outputs) != 1:
continue
if len(input.fluid_code.layers) == 0:
continue
bias_with_act = False
if 'act' in node.fluid_code.layers[-1].param_attr:
bias_with_act = True
layer_with_act = False
index = -1
for i in range(len(input.fluid_code.layers)):
if input.fluid_code.layers[i].op in self.layers_with_bias:
index = i
break
if 'act' in input.fluid_code.layers[
index].param_attr and input.fluid_code.layers[
index].param_attr['act'] is not None:
layer_with_act = True
if bias_with_act and layer_with_act:
continue
if not input.fluid_code.layers[index].param_attr['bias_attr']:
bias_name = node.inputs[1]
input.fluid_code.layers[index].param_attr[
'bias_attr'] = string(bias_name)
input.fluid_code.layers[-1].output = node.fluid_code.layers[
0].output
if bias_with_act:
input.fluid_code.layers[index].param_attr[
'act'] = node.fluid_code.layers[-1].param_attr[
'act']
node.fluid_code.clear()
self.graph.remove_node(node.layer_name)
self.graph.identity_map[node.layer_name] = input.layer_name
def remove_transpose(self):
graph_copy = cp.deepcopy(self.graph)
elementwise_ops = [
'Sub', 'Add', 'RealDiv', 'Maximum', 'Mul', 'FloorDiv',
'GreateerEqual'
]
can_be_optimized_ops = [
'Conv2D', 'MaxPool', 'FusedBatchNorm', 'DepthwiseConv2dNative',
'AvgPool', 'Pad', 'Conv2DBackpropInput', 'ResizeNearestNeighbor',
'Placeholder', 'Relu', 'Relu6', 'Abs', 'Sigmoid', 'Exp', 'Rsqrt',
'swish_f32', 'LeakyRelu', 'Cast', 'Tanh'
]
# These ops may have one more Variable input
can_be_optimized_special_ops = ['ResizeBilinear']
for node_name in self.graph.topo_sort:
node = graph_copy.get_node(node_name)
if node is None:
continue
if node.layer_type in can_be_optimized_ops:
if node.fluid_code.layers[
-1].op != "transpose" or node.fluid_code.layers[
-1].param_attr["perm"] != [0, 2, 3, 1]:
continue
can_be_removed = True
output_names = node.outputs
for out_name in output_names:
out_node = graph_copy.get_node(out_name)
if hasattr(out_node, "can_be_removed"):
if not out_node.can_be_removed:
can_be_removed = False
break
elif out_node.fluid_code.layers[
0].op != "transpose" or out_node.fluid_code.layers[
0].param_attr["perm"] != [0, 3, 1, 2]:
can_be_removed = False
break
elif out_node.layer_type in elementwise_ops or out_node.layer_type in can_be_optimized_special_ops:
can_be_removed = False
break
if can_be_removed and len(node.fluid_code.layers) > 1:
true_node = self.graph.get_node(node_name)
if true_node.layer_type == "Placeholder":
index = self.graph.input_nodes.index(
true_node.fluid_code.layers[-2].output)
if isinstance(true_node.fluid_code.layers[-1].output,
str):
self.graph.input_nodes[
index] = true_node.fluid_code.layers[-1].output
else:
self.graph.input_nodes[
index] = true_node.fluid_code.layers[
-1].output.layer_name
true_node.fluid_code.layers[
-2].output = true_node.fluid_code.layers[-1].output
node.removed = True
del true_node.fluid_code.layers[-1]
for out_name in output_names:
out_node = self.graph.get_node(out_name)
out_node.fluid_code.layers[
1].inputs = out_node.fluid_code.layers[0].inputs
del out_node.fluid_code.layers[0]
for node_name in self.graph.topo_sort:
node = graph_copy.get_node(node_name)
if node is None:
continue
if node.layer_type in elementwise_ops:
can_be_removed = True
if node.fluid_code.layers[
-1].op != "transpose" or node.fluid_code.layers[
-1].param_attr["perm"] != [0, 2, 3, 1]:
continue
can_be_removed = True
output_names = node.outputs
for out_name in output_names:
out_node = graph_copy.get_node(out_name)
if len(out_node.fluid_code.layers) < 3:
can_be_removed = False
break
if hasattr(out_node, "can_be_removed"):
if not out_node.can_be_removed:
can_be_removed = False
break
if out_node.layer_type in can_be_optimized_ops:
if out_node.fluid_code.layers[
0].op != "transpose" or out_node.fluid_code.layers[
0].param_attr["perm"] != [0, 3, 1, 2]:
can_be_removed = False
break
elif out_node.layer_type in elementwise_ops:
if out_node.fluid_code.layers[
0].op != "transpose" and out_node.fluid_code.layers[
1].op != "transpose":
can_be_removed = False
break
if out_node.fluid_code.layers[0].op == "transpose":
if out_node.fluid_code.layers[0].param_attr[
"perm"] != [0, 3, 1, 2]:
can_be_removed = False
break
if out_node.fluid_code.layers[1].op == "transpose":
if out_node.fluid_code.layers[1].param_attr[
"perm"] != [0, 3, 1, 2]:
can_be_removed = False
break
if can_be_removed and len(node.fluid_code.layers) > 1:
true_node = self.graph.get_node(node_name)
true_node.fluid_code.layers[
-2].output = true_node.fluid_code.layers[-1].output
del true_node.fluid_code.layers[-1]
for out_name in output_names:
out_node = self.graph.get_node(out_name)
if out_node.layer_type in can_be_optimized_ops:
out_node.fluid_code.layers[
1].inputs = out_node.fluid_code.layers[0].inputs
del out_node.fluid_code.layers[0]
elif out_node.layer_type in elementwise_ops:
if out_node.inputs[0] in node.layer_name:
if out_node.fluid_code.layers[
1].op == 'transpose':
out_node.fluid_code.layers[2].inputs[
'x'] = out_node.fluid_code.layers[
0].inputs
del out_node.fluid_code.layers[0]
else:
out_node.fluid_code.layers[1].inputs[
'x'] = out_node.fluid_code.layers[
0].inputs
del out_node.fluid_code.layers[0]
elif out_node.inputs[1] in node.layer_name:
if out_node.fluid_code.layers[
1].op == 'transpose':
out_node.fluid_code.layers[2].inputs[
'y'] = out_node.fluid_code.layers[
1].inputs
del out_node.fluid_code.layers[1]
else:
out_node.fluid_code.layers[1].inputs[
'y'] = out_node.fluid_code.layers[
0].inputs
del out_node.fluid_code.layers[0]
graph_copy = cp.deepcopy(self.graph)
for node_name in self.graph.topo_sort:
node = graph_copy.get_node(node_name)
if node is None or len(node.fluid_code.layers) < 2:
continue
if node.layer_type in can_be_optimized_ops and node.layer_type != "Placeholder":
if node.fluid_code.layers[
-1].op != "transpose" or node.fluid_code.layers[
-1].param_attr["perm"] != [0, 2, 3, 1]:
continue
can_be_removed = True
output_names = node.outputs
for out_name in output_names:
out_node = graph_copy.get_node(out_name)
if hasattr(out_node, "can_be_removed"):
if not out_node.can_be_removed:
can_be_removed = False
break
if len(out_node.fluid_code.layers) < 2:
can_be_removed = False
break
if out_node.layer_type in can_be_optimized_ops:
if out_node.fluid_code.layers[
0].op != "transpose" or out_node.fluid_code.layers[
0].param_attr["perm"] != [0, 3, 1, 2]:
can_be_removed = False
break
elif out_node.layer_type in elementwise_ops:
if out_node.fluid_code.layers[
0].op != "transpose" and out_node.fluid_code.layers[
1].op != "transpose":
can_be_removed = False
break
if out_node.fluid_code.layers[
0].op == "expand" or out_node.fluid_code.layers[
1].op == "expand":
can_be_removed = False
break
if out_node.fluid_code.layers[0].op == "transpose":
if out_node.fluid_code.layers[0].param_attr[
"perm"] != [0, 3, 1, 2]:
can_be_removed = False
break
if out_node.fluid_code.layers[1].op == "transpose":
if out_node.fluid_code.layers[1].param_attr[
"perm"] != [0, 3, 1, 2]:
can_be_removed = False
break
elif out_node.layer_type not in elementwise_ops and out_node.layer_type not in can_be_optimized_ops:
can_be_removed = False
break
if can_be_removed:
true_node = self.graph.get_node(node_name)
if len(true_node.fluid_code.layers) < 2:
continue
true_node.fluid_code.layers[
-2].output = true_node.fluid_code.layers[-1].output
del true_node.fluid_code.layers[-1]
for out_name in output_names:
out_node = self.graph.get_node(out_name)
if out_node.layer_type in can_be_optimized_ops:
out_node.fluid_code.layers[
1].inputs = out_node.fluid_code.layers[0].inputs
del out_node.fluid_code.layers[0]
elif out_node.layer_type in elementwise_ops:
if out_node.inputs[0] in node.layer_name:
if out_node.fluid_code.layers[
1].op == 'transpose':
if out_node.fluid_code.layers[
2].op == 'transpose':
out_node.fluid_code.layers[3].inputs[
'x'] = out_node.fluid_code.layers[
0].inputs
else:
out_node.fluid_code.layers[2].inputs[
'x'] = out_node.fluid_code.layers[
0].inputs
del out_node.fluid_code.layers[0]
else:
out_node.fluid_code.layers[1].inputs[
'x'] = out_node.fluid_code.layers[
0].inputs
del out_node.fluid_code.layers[0]
elif out_node.inputs[1] in node.layer_name:
if out_node.fluid_code.layers[
1].op == 'transpose':
out_node.fluid_code.layers[2].inputs[
'y'] = out_node.fluid_code.layers[
1].inputs
del out_node.fluid_code.layers[1]
else:
out_node.fluid_code.layers[1].inputs[
'y'] = out_node.fluid_code.layers[
0].inputs
del out_node.fluid_code.layers[0]
graph_copy = cp.deepcopy(self.graph)
for node_name in self.graph.topo_sort:
node = graph_copy.get_node(node_name)
if node is None:
continue
if node.layer_type in elementwise_ops:
can_be_removed = True
if len(node.fluid_code.layers) < 3:
continue
numTranspose = 0
numNotTranspose = 0
for i in range(len(node.fluid_code.layers)):
if node.fluid_code.layers[i].op == 'transpose':
numTranspose += 1
elif node.fluid_code.layers[i].op != 'expand':
numNotTranspose += 1
if numTranspose > numNotTranspose:
if node.fluid_code.layers[0].op == 'expand':
if node.fluid_code.layers[
1].op != 'transpose' or node.fluid_code.layers[
2].op != 'transpose':
continue
else:
true_node = self.graph.get_node(node_name)
true_node.fluid_code.layers[3].inputs[
'x'] = true_node.fluid_code.layers[1].inputs
true_node.fluid_code.layers[3].inputs[
'y'] = true_node.fluid_code.layers[2].inputs
l = Layer()
l.op = 'transpose'
l.inputs = true_node.fluid_code.layers[3].output
l.param_attr = {'perm': [0, 3, 1, 2]}
if isinstance(l.inputs, six.string_types):
l.output = l.inputs
else:
l.output = l.inputs.layer_name
true_node.fluid_code.layers.append(l)
del true_node.fluid_code.layers[1]
del true_node.fluid_code.layers[1]
else:
if node.fluid_code.layers[
0].op != 'transpose' or node.fluid_code.layers[
1].op != 'transpose':
continue
else:
true_node = self.graph.get_node(node_name)
true_node.fluid_code.layers[2].inputs[
'x'] = true_node.fluid_code.layers[0].inputs
true_node.fluid_code.layers[2].inputs[
'y'] = true_node.fluid_code.layers[1].inputs
l = Layer()
l.op = 'transpose'
l.inputs = true_node.fluid_code.layers[2].output
l.param_attr = {'perm': [0, 3, 1, 2]}
l.output = l.inputs.layer_name
true_node.fluid_code.layers.append(l)
del true_node.fluid_code.layers[0]
del true_node.fluid_code.layers[0]
def make_nchw_input_output(self):
for i, name in enumerate(self.graph.input_nodes):
node = self.graph.get_node(name)
if len(node.out_shapes[0]) == 4 and node.tf_data_format == "NHWC":
shape = node.fluid_code.layers[0].param_attr["shape"]
shape = [shape[j] for j in [0, 3, 1, 2]]
node.fluid_code.layers[0].param_attr["shape"] = shape
node.fluid_code.layers[0].output = "nhwc_" + name
attr = {"perm": [0, 2, 3, 1]}
node.fluid_code.add_layer(
"transpose",
inputs="nhwc_" + name,
output=node,
param_attr=attr)
self.graph.input_nodes[i] = "nhwc_" + name
for i, name in enumerate(self.graph.output_nodes):
node = self.graph.get_node(name)
if node.layer_type != "transpose":
if node.fluid_code.layers[-1].op == "transpose":
node.fluid_code.layers[-2].output = name
del node.fluid_code.layers[-1]
def optimize_sub_graph(self):
self.merge_batch_norm()
self.merge_prelu()
self.merge_scale()
self.merge_affine_channel()
def merge_batch_norm(self):
for i, name in enumerate(self.graph.topo_sort):
node = self.graph.get_node(name)
if node is None:
continue
is_batch_norm = True
if node.layer_type == "Add":
in_nodes0 = [
self.graph.get_node(in_name) for in_name in node.inputs
]
if in_nodes0[0].layer_type != "Mul" or in_nodes0[
1].layer_type != "Sub":
is_batch_norm = False
continue
if exist_act(in_nodes0[0]) or exist_act(in_nodes0[1]):
is_batch_norm = False
continue
in_nodes1 = [
self.graph.get_node(in_name)
for in_name in in_nodes0[0].inputs
]
in_nodes2 = [
self.graph.get_node(in_name)
for in_name in in_nodes0[1].inputs
]
if len(in_nodes1[0].out_shapes[0]) != 4:
is_batch_norm = False
continue
if in_nodes1[1].layer_type != "Mul":
is_batch_norm = False
continue
if exist_act(in_nodes1[1]):
is_batch_norm = False
continue
if in_nodes2[0].layer_type != "Const" or in_nodes2[
1].layer_type != "Mul":
is_batch_norm = False
continue
if exist_act(in_nodes2[1]):
is_batch_norm = False
continue
in_nodes3 = [
self.graph.get_node(in_name)
for in_name in in_nodes1[1].inputs
]
if in_nodes3[0].layer_type != "Rsqrt" or in_nodes3[
1].layer_type != "Const":
is_batch_norm = False
continue
in_nodes4 = [
self.graph.get_node(in_name)
for in_name in in_nodes2[1].inputs
]
if in_nodes4[0].layer_type != "Const" or in_nodes4[
1].layer_name != in_nodes1[1].layer_name:
is_batch_norm = False
continue
in_nodes5 = self.graph.get_node(in_nodes3[0].inputs[0])
if in_nodes5.layer_type != "Add":
is_batch_norm = False
continue
if exist_act(in_nodes5):
is_batch_norm = False
continue
in_nodes6 = [
self.graph.get_node(in_name) for in_name in in_nodes5.inputs
]
if in_nodes6[0].layer_type != "Const" or in_nodes6[
1].layer_type != "Const":
is_batch_norm = False
continue
if len(in_nodes0[0].outputs) != 1:
is_batch_norm = False
continue
if len(in_nodes0[1].outputs) != 1:
is_batch_norm = False
continue
if len(in_nodes1[1].outputs) != 2:
is_batch_norm = False
continue
if len(in_nodes2[0].outputs) != 1:
is_batch_norm = False
continue
if len(in_nodes2[1].outputs) != 1:
is_batch_norm = False
continue
if len(in_nodes3[0].outputs) != 1:
is_batch_norm = False
continue
if len(in_nodes3[1].outputs) != 1:
is_batch_norm = False
continue
if len(in_nodes4[0].outputs) != 1:
is_batch_norm = False
continue
if len(in_nodes5.outputs) != 1:
is_batch_norm = False
continue
if len(in_nodes6[0].outputs) != 1:
is_batch_norm = False
continue
if len(in_nodes6[1].outputs) != 1:
is_batch_norm = False
continue
conv_shape = in_nodes1[0].out_shapes[0]
if conv_shape[3] < 0:
is_batch_norm = False
continue
# moving_variance
if in_nodes6[0].value.size != conv_shape[3]:
is_batch_norm = False
continue
# epsilon
if in_nodes6[1].value.size != 1:
is_batch_norm = False
continue
# gamma
if in_nodes3[1].value.size != conv_shape[3]:
is_batch_norm = False
continue
# moving_mean
if in_nodes4[0].value.size != conv_shape[3]:
is_batch_norm = False
continue
# beta
if in_nodes2[0].value.size != conv_shape[3]:
is_batch_norm = False
continue
if is_batch_norm:
index = in_nodes1[0].outputs.index(in_nodes0[0].layer_name)
in_nodes1[0].outputs[index] = node.layer_name
node.layer_type = "FusedBatchNorm"
node.inputs = [in_nodes1[0].layer_name]
act = node.fluid_code.layers[-1].param_attr.get("act", None)
node.fluid_code.clear()
attr = {
"epsilon": in_nodes6[1].value,
"param_attr": string(in_nodes3[1].layer_name),
"bias_attr": string(in_nodes2[0].layer_name),
"moving_mean_name": string(in_nodes4[0].layer_name),
"moving_variance_name": string(in_nodes6[0].layer_name),
"is_test": True,
"act": act
}
node.fluid_code.add_layer(
"batch_norm",
inputs=in_nodes1[0].fluid_code.layers[-1].output,
output=node,
param_attr=attr)
del self.graph.node_map[in_nodes0[0].layer_name]
del self.graph.node_map[in_nodes0[1].layer_name]
del self.graph.node_map[in_nodes1[1].layer_name]
del self.graph.node_map[in_nodes2[1].layer_name]
del self.graph.node_map[in_nodes3[0].layer_name]
del self.graph.node_map[in_nodes4[0].layer_name]
del self.graph.node_map[in_nodes5.layer_name]
def merge_prelu(self):
for i, name in enumerate(self.graph.topo_sort):
node = self.graph.get_node(name)
if node is None:
continue
is_prelu = True
if node.layer_type == "Add":
if exist_act(node):
is_prelu = False
continue
in_nodes0 = [
self.graph.get_node(in_name) for in_name in node.inputs
]
if in_nodes0[0].layer_type != "Relu" or in_nodes0[
1].layer_type != "Mul":
is_prelu = False
continue
if exist_act(in_nodes0[1]):
is_prelu = False
continue
if len(in_nodes0[0].outputs) != 1 or len(in_nodes0[1]
.outputs) != 1:
is_prelu = False
continue
in_nodes1 = self.graph.get_node(in_nodes0[0].inputs[0])
in_nodes2 = [
self.graph.get_node(in_name)
for in_name in in_nodes0[1].inputs
]
if in_nodes2[1].layer_type != "Const" or numpy.fabs(in_nodes2[
1].value - 0.5) > 1e-06:
is_prelu = False
continue
if in_nodes2[0].layer_type != "Mul":
is_prelu = False
continue
if exist_act(in_nodes2[0]):
is_prelu = False
continue
if len(in_nodes2[1].outputs) != 1 or len(in_nodes2[0]
.outputs) != 1:
is_prelu = False
continue
in_nodes3 = [
self.graph.get_node(in_name)
for in_name in in_nodes2[0].inputs
]
if in_nodes3[0].layer_type != "Const" or in_nodes3[
1].layer_type != "Sub":
is_prelu = False
continue
if exist_act(in_nodes3[1]):
is_prelu = False
continue
if len(in_nodes3[0].outputs) != 1 or len(in_nodes3[1]
.outputs) != 1:
is_prelu = False
continue
in_nodes4 = [
self.graph.get_node(in_name)
for in_name in in_nodes3[1].inputs
]
if in_nodes4[0].layer_name != in_nodes1.layer_name or in_nodes4[
1].layer_type != "Abs":
is_prelu = False
continue
if len(in_nodes4[1].outputs) != 1:
is_prelu = False
continue
in_nodes5 = self.graph.get_node(in_nodes4[1].inputs[0])
if in_nodes5.layer_name != in_nodes1.layer_name:
is_prelu = False
continue
if len(in_nodes0[0].outputs) != 1:
is_prelu = false
continue
if len(in_nodes0[1].outputs) != 1:
is_prelu = False
continue
if len(in_nodes1.outputs) < 3:
is_prelu = False
continue
if len(in_nodes2[0].outputs) != 1:
is_prelu = false
continue
if len(in_nodes2[1].outputs) != 1:
is_prelu = False
continue
if len(in_nodes3[0].outputs) != 1:
is_prelu = False
continue
if len(in_nodes3[1].outputs) != 1:
is_prelu = false
continue
if len(in_nodes4[1].outputs) != 1:
is_prelu = False
continue
mode = None
in_shape = in_nodes1.out_shapes[0]
if in_shape == list(in_nodes3[0].value.shape):
mode = "element"
elif len(in_nodes3[0].value.shape) == 0:
mode = "all"
elif len(in_nodes3[0].value.shape) == 1 and in_nodes3[
0].value.shape[0] == 1:
mode = "all"
elif len(in_shape) == 4 and len(in_nodes3[
0].value.shape) == 1 and in_nodes3[0].value.shape[
0] == in_shape[-1]:
mode = "channel"
weight = self.op_mapper.weights[in_nodes3[0].layer_name]
weight = numpy.expand_dims(weight, 0)
weight = numpy.expand_dims(weight, 2)
weight = numpy.expand_dims(weight, 3)
self.op_mapper.weights[in_nodes3[0].layer_name] = weight
# fix bug in Paddle1.8.3 and may change in next version.
# self.op_mapper.weights[in_nodes3[0].layer_name +
# '_1'] = weight.reshape(1, -1)
in_nodes3[0].fluid_code.layers[0].param_attr["shape"] = [
1, in_shape[-1], 1, 1
]
else:
is_prelu = False
continue
if is_prelu:
index = in_nodes1.outputs.index(in_nodes0[0].layer_name)
del in_nodes1.outputs[index]
index = in_nodes1.outputs.index(in_nodes3[1].layer_name)
del in_nodes1.outputs[index]
index = in_nodes1.outputs.index(in_nodes4[1].layer_name)
del in_nodes1.outputs[index]
in_nodes1.outputs.append(node.layer_name)
node.layer_type = "Prelu"
node.inputs = [in_nodes1.layer_name]
act = node.fluid_code.layers[-1].param_attr.get("act", None)
node.fluid_code.clear()
attr = {
"mode": string(mode),
"param_attr": string(in_nodes3[0].layer_name)
}
node.fluid_code.add_layer(
"prelu",
inputs=in_nodes1.fluid_code.layers[-1].output,
output=node,
param_attr=attr)
del self.graph.node_map[in_nodes0[0].layer_name]
del self.graph.node_map[in_nodes0[1].layer_name]
del self.graph.node_map[in_nodes2[0].layer_name]
del self.graph.node_map[in_nodes2[1].layer_name]
del self.graph.node_map[in_nodes3[1].layer_name]
del self.graph.node_map[in_nodes4[1].layer_name]
def merge_scale(self):
for i, name in enumerate(self.graph.topo_sort):
node = self.graph.get_node(name)
if node is None:
continue
is_scale = True
if node.layer_type == "Sub":
in_nodes0 = [
self.graph.get_node(in_name) for in_name in node.inputs
]
if in_nodes0[0].layer_type != "Mul" or in_nodes0[
1].layer_type != "Const" or in_nodes0[
1].value.size != 1:
is_scale = False
continue
if exist_act(in_nodes0[0]):
is_scale = False
continue
if len(in_nodes0[0].outputs) != 1 or len(in_nodes0[1]
.outputs) != 1:
is_scale = False
continue
in_nodes1 = [
self.graph.get_node(in_name)
for in_name in in_nodes0[0].inputs
]
if in_nodes1[0].layer_type != "Const" or in_nodes1[
1].layer_type != "RealDiv" or in_nodes1[
0].value.size != 1:
is_scale = False
continue
if exist_act(in_nodes1[1]):
is_scale = False
continue
if len(in_nodes1[0].outputs) != 1 or len(in_nodes1[1]
.outputs) != 1:
is_scale = False
continue
in_nodes2 = [
self.graph.get_node(in_name)
for in_name in in_nodes1[1].inputs
]
if in_nodes2[1].layer_type != "Const" or in_nodes2[
1].value.size != 1:
is_scale = False
continue
if is_scale:
in_node = self.graph.get_node(in_nodes1[1].inputs[0])
index = in_node.outputs.index(in_nodes1[1].layer_name)
in_node.outputs[index] = node.layer_name
node.layer_type = "Scale"
node.inputs = [in_node.layer_name]
scale = 1.0 / in_nodes2[1].value * in_nodes1[0].value
act = None
if node.fluid_code.layers[0].param_attr is not None:
act = node.fluid_code.layers[0].param_attr.get("act",
None)
node.fluid_code.clear()
attr = {
"scale": scale,
"bias": in_nodes0[1].value,
"bias_after_scale": True,
"act": act
}
node.fluid_code.add_layer(
"scale", inputs=in_node, output=node, param_attr=attr)
del self.graph.node_map[in_nodes0[0].layer_name]
del self.graph.node_map[in_nodes0[1].layer_name]
del self.graph.node_map[in_nodes1[0].layer_name]
del self.graph.node_map[in_nodes1[1].layer_name]
del self.graph.node_map[in_nodes2[1].layer_name]
def merge_affine_channel(self):
for i, name in enumerate(self.graph.topo_sort):
node = self.graph.get_node(name)
if node is None:
continue
is_affine_channel = True
if node.layer_type == "RealDiv":
in_nodes0 = [
self.graph.get_node(in_name) for in_name in node.inputs
]
bias_add = True
if (in_nodes0[0].layer_type != "Sub" and in_nodes0[0].layer_type
!= "Add") or in_nodes0[1].layer_type != "Const" or len(
in_nodes0[1].value.shape) != 3:
is_affine_channel = False
continue
if in_nodes0[0].layer_type == "Sub":
bias_add = False
if exist_act(in_nodes0[0]):
is_affine_channel = False
continue
if len(in_nodes0[0].outputs) != 1 or len(in_nodes0[1]
.outputs) != 1:
is_affine_channel = False
continue
in_nodes1 = [
self.graph.get_node(in_name)
for in_name in in_nodes0[0].inputs
]
if len(in_nodes1[0].out_shapes[0]) != 4 or in_nodes1[
1].layer_type != "Const" or len(in_nodes1[1]
.value.shape) != 3:
is_affine_channel = False
continue
if len(in_nodes1[1].outputs) != 1:
is_affine_channel = False
continue
channel = in_nodes1[0].out_shapes[0][-1]
if channel < 0 or channel != in_nodes0[
1].value.size or channel != in_nodes1[1].value.size:
is_affine_channel = False
continue
if in_nodes0[1].out_shapes[0][-1] != in_nodes0[
1].value.size or in_nodes1[1].out_shapes[0][
-1] != in_nodes1[1].value.size:
is_affine_channel = False
continue
if is_affine_channel:
in_node = in_nodes1[0]
index = in_node.outputs.index(in_nodes0[0].layer_name)
in_node.outputs[index] = node.layer_name
node.layer_type = "AffineChannel"
node.inputs = [in_node.layer_name]
scale = 1.0 / in_nodes0[1].value.flatten()
bias = in_nodes1[1].value.flatten() / in_nodes0[
1].value.flatten()
if not bias_add:
bias *= -1.0
self.op_mapper.weights[node.layer_name + "_scale"] = scale
self.op_mapper.weights[node.layer_name + "_bias"] = bias
act = None
if node.fluid_code.layers[0].param_attr is not None:
act = node.fluid_code.layers[0].param_attr.get("act",
None)
node.fluid_code.clear()
attr = {
"dtype": string(scale.dtype),
"shape": [channel],
"name": string(node.layer_name + "_scale")
}
node.fluid_code.add_layer(
"create_parameter",
inputs=None,
output=node.layer_name + "_scale",
param_attr=attr)
attr = {
"dtype": string(scale.dtype),
"shape": [channel],
"name": string(node.layer_name + "_bias")
}
node.fluid_code.add_layer(
"create_parameter",
inputs=None,
output=node.layer_name + "_bias",
param_attr=attr)
inputs = {
"x": in_node,
"scale": node.layer_name + "_scale",
"bias": node.layer_name + "_bias"
}
attr = {"act": act}
node.fluid_code.add_layer(
"affine_channel",
inputs=inputs,
output=node,
param_attr=attr)
del self.graph.node_map[in_nodes0[0].layer_name]
del self.graph.node_map[in_nodes0[1].layer_name]
del self.graph.node_map[in_nodes1[1].layer_name]
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或