未验证 提交 fb07475f 编写于 作者: J Jason 提交者: GitHub

Merge pull request #451 from SunAhong1993/paddle-2.0

modify the tf static
......@@ -132,18 +132,9 @@ def tf2paddle(model_path,
graph_opt = GraphOptimizer(source_frame="tf", paddle_type=paddle_type)
graph_opt.optimize(mapper.paddle_graph)
else:
from x2paddle.optimizer.tensorflow.bias import BiasOpt
from x2paddle.optimizer.tensorflow.transpose import TransposeOpt
from x2paddle.optimizer.tensorflow.batch_norm import BatchNormOpt
from x2paddle.optimizer.tensorflow.prelu import PReLUOpt
bias_opt = BiasOpt()
transpose_opt = TransposeOpt()
batch_norm_opt = BatchNormOpt()
prelu_opt = PReLUOpt()
bias_opt.run(mapper.paddle_graph)
batch_norm_opt.run(mapper.paddle_graph)
prelu_opt.run(mapper.paddle_graph)
transpose_opt.run(mapper.paddle_graph)
from x2paddle.optimizer.optimizer import GraphOptimizer
graph_opt = GraphOptimizer(source_frame="tf", paddle_type=paddle_type)
graph_opt.optimize(mapper.paddle_graph)
mapper.paddle_graph.gen_model(save_dir)
......
......@@ -284,7 +284,6 @@ class TFOpMapper(OpMapper):
inputs["shape"] = dims.name
layer_attrs["dtype"] = string(input_value.dtype)
layer_attrs["fill_value"] = input_value.value
self.paddle_graph.add_layer(
"paddle.full",
......@@ -578,6 +577,9 @@ class TFOpMapper(OpMapper):
inputs={"x": node.name},
outputs=[node.name],
perm=[0, 2, 3, 1])
def FusedBatchNormV3(self, node):
self.FusedBatchNorm(node)
def Mean(self, node):
input = self.graph.get_input_node(node, 0)
......@@ -930,6 +932,23 @@ class TFOpMapper(OpMapper):
outputs=[node.name],
axis=axis)
def Concat(self, node):
inputs_list = list()
for i in range(1, len(node.inputs)):
inputs_list.append(self.graph.get_input_node(node, i))
axis = self.graph.get_input_node(node, 0)
assert axis.layer_type == "Const", "axis for ConcatV2 must be type Const"
axis = axis.value
if axis < 0:
axis += len(inputs_list[0].out_shapes[0])
input_names = [i.name for i in inputs_list]
self.paddle_graph.add_layer(
kernel="paddle.concat",
inputs={"x": input_names},
outputs=[node.name],
axis=axis)
def AddN(self, node):
inputs_list = list()
for i in range(len(node.inputs) - 1):
......@@ -1400,6 +1419,7 @@ class TFOpMapper(OpMapper):
inputs = {"x": x.name, "y": y.name}
x_shape = x.out_shapes[0]
y_shape = y.out_shapes[0]
# TODO(syf)
layer_id = self.paddle_graph.add_layer(
"fluid.layers.elementwise_sub", inputs=inputs, outputs=[node.name])
self.paddle_graph.layers[layer_id].input_shapes = {"x": x_shape, "y": y_shape}
......
......@@ -457,7 +457,6 @@ class CaffeOpMapper(OpMapper):
def ReLU(self, node):
"""
:param node:
:return:
"""
......@@ -973,5 +972,4 @@ class CaffeOpMapper(OpMapper):
self.paddle_graph.add_layer(
kernel=op_info,
inputs={"x": self.get_input_name(input)},
outputs=[node.layer_name])
outputs=[node.layer_name])
\ No newline at end of file
# 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 .transpose_elimination import StaticTransposeElimination
from .transpose_eliminate_pass import StaticTransposeEliminatePass
\ No newline at end of file
# 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 x2paddle.optimizer.pass_ import Pass
from x2paddle.optimizer.elimination.static import StaticTransposeElimination
from x2paddle.optimizer.pass_manager import pass_register
@pass_register
class StaticTransposeEliminatePass(Pass):
name = "static_transpose_eliminate_pass"
def __init__(self):
Pass.__init__(self)
def apply(self, graph):
fuser = StaticTransposeElimination()
fuser.operate(graph)
# 用于注册
static_transpose_eliminate_pass = StaticTransposeEliminatePass()
\ No newline at end of file
# 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 copy
import sys
import numpy as np
from x2paddle.optimizer.pattern_matcher import FuseBase
from x2paddle.core.program import PaddleGraph, PaddleLayer
from x2paddle.core.util import *
class TransposeOpt:
class StaticTransposeElimination(FuseBase):
def __init__(self):
self.image_layers = [
'fluid.layers.conv2d', 'fluid.layers.batch_norm',
'fluid.layers.conv2d_transpose', 'fluid.layers.resize_nearest',
'fluid.layers.resize_bilinear', 'fluid.layers.pool2d',
'fluid.layers.pad2d'
]
super(StaticTransposeElimination, self).__init__(graph_type="static")
self.direct_layers = [
'fluid.layers.relu', 'fluid.layers.relu6', 'fluid.layers.abs',
'fluid.layers.sigmoid', 'fluid.layers.exp', 'fluid.layers.rsqrt',
'fluid.layers.swish_f32', 'fluid.layers.tanh',
'fluid.layers.softplus', 'fluid.layers.leaky_relu',
'fluid.layers.floor', 'fluid.layers.erf', 'fluid.layers.swish'
'paddle.nn.functional.relu', 'paddle.nn.functional.relu6', 'paddle.abs',
'paddle.nn.functional.sigmoid', 'paddle.exp', 'paddle.rsqrt',
'paddle.nn.functional.swish', 'paddle.tanh',
'paddle.nn.functional.softplus', 'paddle.nn.functional.leaky_relu',
'paddle.floor', 'paddle.erf', 'paddle.square'
]
self.elementwise_layers = [
'fluid.layers.elementwise_add', 'fluid.layers.elementwise_sub',
'fluid.layers.elementwise_mul', 'fluid.layers.elementwise_div'
'paddle.add', 'fluid.layers.elementwise_sub',
'paddle.multiply', 'paddle.divide'
]
self.reduce_layers = [
'fluid.layers.reduce_mean', 'fluid.layers.reduce_all',
'fluid.layers.reduce_max', 'fluid.layers.reduce_any',
'fluid.layers.reduce_sum', 'fluid.layers.reduce_prod'
'paddle.mean', 'paddle.all',
'paddle.max', 'paddle.any',
'paddle.sum', 'paddle.prod'
]
def get_transpose_num(self, graph):
count = 0
for layer_id, layer in graph.layers.items():
if layer.kernel == "fluid.layers.transpose":
if layer.kernel == "paddle.transpose":
count += 1
return count
def run(self, graph):
print("Optimize: TransposeOpt...")
def operate(self, graph):
total_layer_num = len(graph.layers)
scanned_layers = set()
optimized_transpose_layers = list()
optimized_reduce_layers = list()
optimized_concat_layers = list()
optimized_elementwise_layers = list()
def get_index(layer):
if layer.kernel.startswith("paddle.nn") and "functional" not in layer.kernel:
return 1
else:
return 0
def strip_transpose(_graph):
layers = copy.deepcopy(_graph.layers)
......@@ -53,7 +71,7 @@ class TransposeOpt:
sys.stderr.write("\rOptimize Transpose Layers...{}%".format(
percent))
if layer.kernel != "fluid.layers.transpose":
if layer.kernel != "paddle.transpose":
continue
if layer.attrs["perm"] != [0, 2, 3, 1]:
continue
......@@ -65,7 +83,7 @@ class TransposeOpt:
elementwise_layers = list()
can_be_optimized = True
for out in _graph.edges_out.get(layer_id, []):
if _graph.layers[out].kernel == "fluid.layers.transpose":
if _graph.layers[out].kernel == "paddle.transpose":
if _graph.layers[out].attrs["perm"] != [0, 3, 1, 2]:
can_be_optimized = False
break
......@@ -73,21 +91,24 @@ class TransposeOpt:
elif _graph.layers[out].kernel in self.elementwise_layers:
propagate_layers.append(out)
elif _graph.layers[out].kernel in self.direct_layers:
if _graph.layers[out].outputs[0] in _graph.outputs:
ouput_index = get_index(_graph.layers[out])
if _graph.layers[out].outputs[ouput_index] in _graph.outputs:
can_be_optimized = False
break
propagate_layers.append(out)
elif _graph.layers[out].kernel in self.reduce_layers:
if _graph.layers[out].outputs[0] in _graph.outputs:
ouput_index = get_index(_graph.layers[out])
if _graph.layers[out].outputs[ouput_index] in _graph.outputs:
can_be_optimized = False
break
if not _graph.layers[out].attrs.get('keep_dim', False):
if not _graph.layers[out].attrs.get('keepdim', False):
can_be_optimized = False
break
propagate_layers.append(out)
reduce_layers.append(out)
elif _graph.layers[out].kernel == "fluid.layers.concat":
if _graph.layers[out].outputs[0] in _graph.outputs:
elif _graph.layers[out].kernel == "paddle.concat":
ouput_index = get_index(_graph.layers[out])
if _graph.layers[out].outputs[ouput_index] in _graph.outputs:
can_be_optimized = False
break
propagate_layers.append(out)
......@@ -102,37 +123,41 @@ class TransposeOpt:
visited_layers.add(current_id)
for out in _graph.edges_out.get(current_id, []):
if _graph.layers[
out].kernel == "fluid.layers.transpose":
out].kernel == "paddle.transpose":
if _graph.layers[out].attrs["perm"] != [0, 3, 1, 2]:
can_be_optimized = False
break
transpose_layers.append(out)
elif _graph.layers[
out].kernel in self.elementwise_layers:
if _graph.layers[out].outputs[0] in _graph.outputs:
output_index = get_index(_graph.layers[out])
if _graph.layers[out].outputs[output_index] in _graph.outputs:
can_be_optimized = False
break
if out not in visited_layers:
propagate_layers.append(out)
elif _graph.layers[out].kernel in self.direct_layers:
if _graph.layers[out].outputs[0] in _graph.outputs:
output_index = get_index(_graph.layers[out])
if _graph.layers[out].outputs[output_index] in _graph.outputs:
can_be_optimized = False
break
if out not in visited_layers:
propagate_layers.append(out)
elif _graph.layers[out].kernel in self.reduce_layers:
if _graph.layers[out].outputs[0] in _graph.outputs:
output_index = get_index(_graph.layers[out])
if _graph.layers[out].outputs[output_index] in _graph.outputs:
can_be_optimized = False
break
if not _graph.layers[out].attrs.get('keep_dim',
if not _graph.layers[out].attrs.get('keepdim',
False):
can_be_optimized = False
break
if out not in visited_layers:
propagate_layers.append(out)
reduce_layers.append(out)
elif _graph.layers[out].kernel == "fluid.layers.concat":
if _graph.layers[out].outputs[0] in _graph.outputs:
elif _graph.layers[out].kernel == "paddle.concat":
output_index = get_index(_graph.layers[out])
if _graph.layers[out].outputs[output_index] in _graph.outputs:
can_be_optimized = False
break
if out not in visited_layers:
......@@ -149,14 +174,15 @@ class TransposeOpt:
current_id].input_shapes['x']
y_shape = _graph.layers[
current_id].input_shapes['y']
output_index = get_index(_graph.layers[ipt])
if _graph.layers[ipt].outputs[
0] == _graph.layers[current_id].inputs[
output_index] == _graph.layers[current_id].inputs[
'x']:
if len(x_shape) <= 1:
elementwise_layers.append(current_id)
continue
elif _graph.layers[ipt].outputs[
0] == _graph.layers[current_id].inputs[
output_index] == _graph.layers[current_id].inputs[
'y']:
if len(y_shape) <= 1:
elementwise_layers.append(current_id)
......@@ -168,8 +194,9 @@ class TransposeOpt:
except Exception as e:
can_be_optimized = False
break
output_index = get_index(_graph.layers[ipt])
if _graph.layers[
ipt].kernel == "fluid.layers.transpose":
ipt].kernel == "paddle.transpose":
if _graph.layers[ipt].attrs["perm"] != [0, 2, 3, 1]:
can_be_optimized = False
break
......@@ -177,30 +204,30 @@ class TransposeOpt:
transpose_layers.append(ipt)
elif _graph.layers[
ipt].kernel in self.elementwise_layers:
if _graph.layers[ipt].outputs[0] in _graph.outputs:
if _graph.layers[ipt].outputs[output_index] in _graph.outputs:
can_be_optimized = False
break
if ipt not in visited_layers:
propagate_layers.append(ipt)
elif _graph.layers[ipt].kernel in self.direct_layers:
if _graph.layers[ipt].outputs[0] in _graph.outputs:
if _graph.layers[ipt].outputs[output_index] in _graph.outputs:
can_be_optimized = False
break
if ipt not in visited_layers:
propagate_layers.append(ipt)
elif _graph.layers[ipt].kernel in self.reduce_layers:
if _graph.layers[ipt].outputs[0] in _graph.outputs:
if _graph.layers[ipt].outputs[output_index] in _graph.outputs:
can_be_optimized = False
break
if not _graph.layers[ipt].attrs.get('keep_dim',
if not _graph.layers[ipt].attrs.get('keepdim',
False):
can_be_optimized = False
break
if ipt not in visited_layers:
propagate_layers.append(ipt)
reduce_layers.append(ipt)
elif _graph.layers[ipt].kernel == "fluid.layers.concat":
if _graph.layers[ipt].outputs[0] in _graph.outputs:
elif _graph.layers[ipt].kernel == "paddle.concat":
if _graph.layers[ipt].outputs[output_index] in _graph.outputs:
can_be_optimized = False
break
if ipt not in visited_layers:
......@@ -217,7 +244,8 @@ class TransposeOpt:
transpose_layers.append(layer_id)
transpose_layers = list(set(transpose_layers))
for l in transpose_layers:
if graph.layers[l].outputs[0] in graph.outputs:
output_index = get_index(graph.layers[l])
if graph.layers[l].outputs[output_index] in graph.outputs:
can_be_optimized = False
break
if not can_be_optimized:
......@@ -243,17 +271,19 @@ class TransposeOpt:
for layer_id in list(set(optimized_transpose_layers)):
graph.del_layer(layer_id)
for layer_id in list(set(optimized_reduce_layers)):
dim = graph.layers[layer_id].attrs.get('dim', None)
dim = graph.layers[layer_id].attrs.get('axis', None)
if dim is not None:
for i in range(len(dim)):
dim[i] = [0, 2, 3, 1][dim[i]]
graph.layers[layer_id].attrs['dim'] = dim
graph.layers[layer_id].attrs['axis'] = dim
for layer_id in list(set(optimized_concat_layers)):
axis = graph.layers[layer_id].attrs.get('axis', 0)
graph.layers[layer_id].attrs['axis'] = [0, 2, 3, 1][axis]
for layer_id in list(set(optimized_elementwise_layers)):
axis = graph.layers[layer_id].attrs.get('axis', -1)
graph.layers[layer_id].attrs['axis'] = [0, 2, 3, 1][axis]
if graph.layers[layer_id].kernel == "paddle.add":
graph.layers[layer_id].kernel = "fluid.layers.elementwise_add"
current_transpose_num = self.get_transpose_num(graph)
print(
......
......@@ -105,10 +105,6 @@ class DygraphConv2DAddFuser(FuseBase):
if layer.kernel == "paddle.nn.Conv2D":
conv_id = layer_id
for layer_id, layer in matches.items():
if layer.kernel == "paddle.nn.functional.conv2d_transpose":
layer.bias = bias_name
if not is_transpose:
layer.outputs[0] = output_name
if layer.kernel == "paddle.nn.Conv2D":
layer.attrs["bias_attr"] = bias_name
if not is_transpose:
......
......@@ -13,4 +13,11 @@
# limitations under the License.
from .bn_scale_fuser import Static_BNScaleFuser
from .bn_scale_fuse_pass import Static_BNScaleFusePass
\ No newline at end of file
from .bn_scale_fuse_pass import Static_BNScaleFusePass
from .conv2d_add_fuser import StaticConv2DAddFuser
from .conv2d_add_fuse_pass import StaticConv2DAddFusePass
from .prelu_fuser import StaticPReLUFuser
from .prelu_fuse_pass import StaticPReLUFusePass
from .tf_batchnorm_fuser import StaticTFBatchNormFuser
from .tf_batchnorm_fuse_pass import StaticTFBatchNormFusePass
......@@ -79,7 +79,6 @@ class Static_BNScaleFuser(FuseBase):
graph.layers[new_layer_id] = new_layer
matches.pop(new_layer_id)
def gen_new_layer(self, parameters, matches):
layers_id = list(matches.keys())
layer = matches[layers_id[0]]
......
# 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 x2paddle.optimizer.pass_ import Pass
from x2paddle.optimizer.fusion.static import StaticConv2DAddFuser
from x2paddle.optimizer.pass_manager import pass_register
@pass_register
class StaticConv2DAddFusePass(Pass):
name = "static_conv2d_add_fuse_pass"
def __init__(self):
Pass.__init__(self)
def apply(self, graph):
fuser = StaticConv2DAddFuser()
fuser.operate(graph, match_kind="edge")
# 用于注册
static_conv2d_add_fuse_pass = StaticConv2DAddFusePass()
\ No newline at end of file
# 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 copy
import numpy as np
from x2paddle.optimizer.pattern_matcher import FuseBase
from x2paddle.core.program import PaddleGraph, PaddleLayer
from x2paddle.core.util import *
class StaticConv2DAddFuser(FuseBase):
def __init__(self):
super(StaticConv2DAddFuser, self).__init__(graph_type="static")
self.patterns = list()
def build_pattern(self):
""" 描述需要替换的conv2d+add图结构。
conv2d+add层模式python实现代码示例:
模式一:
MobilenetV1_Logits_Conv2d_1c_1x1_biases = paddle.static.create_parameter(dtype='float32', shape=[1001], name='MobilenetV1_Logits_Conv2d_1c_1x1_biases', default_initializer=paddle.nn.initializer.Constant(value=0.0))
conv2d_transpose_14 = paddle.transpose(x=MobilenetV1_Logits_AvgPool_1a_AvgPool, perm=[0, 3, 1, 2])
MobilenetV1_Logits_Conv2d_1c_1x1_Conv2D = paddle.nn.functional.conv2d(x=conv2d_transpose_14, weight=MobilenetV1_Logits_Conv2d_1c_1x1_weights, bias=None, stride=[1, 1], dilation=[1, 1], padding='SAME')
MobilenetV1_Logits_Conv2d_1c_1x1_Conv2D = paddle.transpose(x=MobilenetV1_Logits_Conv2d_1c_1x1_Conv2D, perm=[0, 2, 3, 1])
MobilenetV1_Logits_Conv2d_1c_1x1_BiasAdd = paddle.add(x=MobilenetV1_Logits_Conv2d_1c_1x1_Conv2D, y=MobilenetV1_Logits_Conv2d_1c_1x1_biases)
模式二:
MobilenetV1_Logits_Conv2d_1c_1x1_biases = paddle.static.create_parameter(dtype='float32', shape=[1001], name='MobilenetV1_Logits_Conv2d_1c_1x1_biases', default_initializer=paddle.nn.initializer.Constant(value=0.0))
MobilenetV1_Logits_Conv2d_1c_1x1_Conv2D = paddle.nn.functional.conv2d(x=conv2d_transpose_14, weight=MobilenetV1_Logits_Conv2d_1c_1x1_weights, bias=None, stride=[1, 1], dilation=[1, 1], padding='SAME')
MobilenetV1_Logits_Conv2d_1c_1x1_BiasAdd = paddle.add(x=MobilenetV1_Logits_Conv2d_1c_1x1_Conv2D, y=MobilenetV1_Logits_Conv2d_1c_1x1_biases)
"""
def gen_name(id):
return "x" + str(id)
pattern = PaddleGraph(graph_type="dygraph")
pattern.add_layer(
"paddle.static.create_parameter",
inputs={},
outputs=[gen_name(0)])
pattern.add_layer(
kernel="paddle.transpose",
inputs={"x": "conv-input-0"},
outputs=[gen_name(1)],
perm=[0, 3, 1, 2])
pattern.add_layer(
kernel="paddle.nn.functional.conv2d",
inputs={"input": gen_name(1),
"weight": "conv-input-1"},
outputs=[gen_name(2)])
pattern.add_layer(
kernel="paddle.transpose",
inputs={"x": gen_name(2)},
outputs=[gen_name(2)],
perm=[0, 2, 3, 1])
pattern.add_layer(
kernel="paddle.add",
inputs={"x": gen_name(2),
"y": gen_name(0)},
outputs=[gen_name(3)])
pattern.build(inputs={"input-0": "conv-input-0",
"input-1": "conv-input-1"})
self.patterns.append(pattern)
pattern = PaddleGraph(graph_type="dygraph")
pattern.add_layer(
"paddle.static.create_parameter",
inputs={},
outputs=[gen_name(0)])
pattern.add_layer(
kernel="paddle.nn.functional.conv2d",
inputs={"input": "conv-input-0",
"weight": "conv-input-1"},
outputs=[gen_name(1)])
pattern.add_layer(
kernel="paddle.add",
inputs={"x": gen_name(1),
"y": gen_name(0)},
outputs=[gen_name(2)])
pattern.build(inputs={"input-0": "conv-input-0",
"input-1": "conv-input-1"})
self.patterns.append(pattern)
def insert_new_layer(self, graph, parameters, matches):
self.gen_new_layer(matches, graph)
matches_copy = copy.deepcopy(matches)
for layer_id, layer in matches_copy.items():
if layer.kernel not in ["paddle.add"]:
matches.pop(layer_id)
def gen_new_layer(self, matches, graph):
is_transpose = False
for layer_id, layer in matches.items():
if layer.kernel == "paddle.static.create_parameter":
bias_name = layer.attrs["name"][1: -1]
if layer.kernel == "paddle.transpose":
is_transpose = True
if layer.kernel == "paddle.add":
output_name = layer.outputs[0]
if layer.kernel == "paddle.nn.functional.conv2d":
conv_id = layer_id
for layer_id, layer in matches.items():
if layer.kernel == "paddle.nn.functional.conv2d":
layer.inputs["bias"] = bias_name
layer.attrs.pop("bias")
if not is_transpose:
layer.outputs[0] = output_name
if layer.kernel == "paddle.transpose":
if conv_id in graph.edges_in[layer_id]:
layer.outputs[0] = output_name
# 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 x2paddle.optimizer.pass_ import Pass
from x2paddle.optimizer.fusion.static import StaticPReLUFuser
from x2paddle.optimizer.pass_manager import pass_register
@pass_register
class StaticPReLUFusePass(Pass):
name = "static_prelu_fuse_pass"
def __init__(self):
Pass.__init__(self)
def apply(self, graph):
fuser = StaticPReLUFuser()
fuser.operate(graph, match_kind="edge")
# 用于注册
static_prelu_fuse_pass = StaticPReLUFusePass()
\ No newline at end of file
# 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 copy
import numpy as np
from collections import OrderedDict
from x2paddle.optimizer.pattern_matcher import FuseBase
from x2paddle.core.program import PaddleGraph, PaddleLayer
from x2paddle.core.util import *
class StaticPReLUFuser(FuseBase):
def __init__(self):
super(StaticPReLUFuser, self).__init__(graph_type="static")
def build_pattern(self):
""" 描述需要替换的prelu图结构。
prelu层模式python实现代码示例:
conv4_alphas = paddle.static.create_parameter(dtype='float32', shape=[128], name='conv4_alphas', default_initializer=paddle.nn.initializer.Constant(value=0.0))
conv4_mul_1_y = paddle.full(dtype='float32', shape=[1], fill_value=0.5)
conv4_Relu = paddle.nn.functional.relu(x=conv4_BiasAdd)
conv4_Abs = paddle.abs(x=conv4_BiasAdd)
conv4_sub = fluid.layers.elementwise_sub(x=conv4_BiasAdd, y=conv4_Abs)
conv4_mul = paddle.multiply(x=conv4_alphas, y=conv4_sub)
conv4_mul_1 = paddle.multiply(x=conv4_mul, y=conv4_mul_1_y)
conv4_add = paddle.add(x=conv4_Relu, y=conv4_mul_1)
"""
def gen_name(id):
return "x" + str(id)
self.pattern.add_layer(
"paddle.static.create_parameter",
inputs={},
outputs=[gen_name(0)])
self.pattern.add_layer(
"paddle.full",
inputs={},
outputs=[gen_name(1)],
shape=[1],
fill_value=0.5)
self.pattern.add_layer(
"paddle.nn.functional.relu",
inputs={"x": "prelu-input-0"},
outputs=[gen_name(2)])
self.pattern.add_layer(
"paddle.abs",
inputs={"x": "prelu-input-0"},
outputs=[gen_name(3)])
self.pattern.add_layer(
"fluid.layers.elementwise_sub",
inputs={"x": "prelu-input-0",
"y": gen_name(3)},
outputs=[gen_name(4)])
self.pattern.add_layer(
"paddle.multiply",
inputs={"x": gen_name(0),
"y": gen_name(4)},
outputs=[gen_name(5)])
self.pattern.add_layer(
"paddle.multiply",
inputs={"x": gen_name(5),
"y": gen_name(1)},
outputs=[gen_name(6)])
self.pattern.add_layer(
"paddle.add",
inputs={"x": gen_name(2),
"y": gen_name(6)},
outputs=[gen_name(7)])
self.pattern.build(inputs={"input-0": "prelu-input-0", })
def insert_new_layer(self, graph, parameters, matches):
new_layers, last_layer_id = self.gen_new_layer(matches, parameters, graph)
matches_copy = copy.deepcopy(matches)
for layer_id, layer in matches_copy.items():
for i in range(4):
if layer_id == new_layers[i].id:
matches.pop(new_layers[i].id)
prefix_layers = OrderedDict()
mid_layers = OrderedDict()
suffix_layers = OrderedDict()
is_need_id = False
for layer_id, layer in graph.layers.items():
if is_need_id:
suffix_layers[layer_id] = layer
else:
if layer_id == last_layer_id:
for i in range(4):
mid_layers[new_layers[i].id] = new_layers[i]
is_need_id = True
prefix_layers[layer_id] = layer
prefix_layers.update(mid_layers)
prefix_layers.update(suffix_layers)
graph.layers = prefix_layers
def gen_new_layer(self, matches, parameters, graph):
layer_id_list = list(matches.keys())
layer_id_list.sort(key = int)
for layer_id, layer in matches.items():
if layer.kernel == "paddle.nn.functional.relu":
input_name = layer.inputs["x"]
if layer.kernel == "paddle.static.create_parameter":
param_layer = layer
param_name = layer.outputs[0]
if layer.kernel == "paddle.add":
output_name = layer.outputs[0]
transpose0 = PaddleLayer(
id=layer_id_list[-1] + "_1",
kernel="paddle.transpose",
inputs={"x": input_name},
outputs=["{}_transpose_for_prelu".format(input_name)],
perm=[0, 3, 1, 2])
param = parameters[param_name]
c = param.shape[0]
prelu = PaddleLayer(id=layer_id_list[-1] + "_2",
kernel="paddle.nn.functional.prelu",
inputs={"x": "{}_transpose_for_prelu".format(input_name),
"weight": param_name},
outputs=["{}_prelu".format(input_name)])
transpose1 = PaddleLayer(
id=layer_id_list[-1] + "_3",
kernel="paddle.transpose",
inputs={"x": "{}_prelu".format(input_name)},
outputs=[output_name],
perm=[0, 2, 3, 1])
return [param_layer, transpose0, prelu, transpose1], layer_id_list[-1]
# 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 x2paddle.optimizer.pass_ import Pass
from x2paddle.optimizer.fusion.static import StaticTFBatchNormFuser
from x2paddle.optimizer.pass_manager import pass_register
@pass_register
class StaticTFBatchNormFusePass(Pass):
name = "static_tf_batchnorm_fuse_pass"
def __init__(self):
Pass.__init__(self)
def apply(self, graph):
fuser = StaticTFBatchNormFuser()
fuser.operate(graph, match_kind="edge")
# 用于注册
static_tf_batchnorm_fuse_pass = StaticTFBatchNormFusePass()
\ No newline at end of file
# 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 copy
import numpy as np
from collections import OrderedDict
from x2paddle.optimizer.pattern_matcher import FuseBase
from x2paddle.core.program import PaddleGraph, PaddleLayer
from x2paddle.core.util import *
class StaticTFBatchNormFuser(FuseBase):
def __init__(self):
super(StaticTFBatchNormFuser, self).__init__(graph_type="static")
self.patterns = list()
def build_pattern(self):
""" 描述需要替换的batchnorm图结构。
batchnorm层模式python实现代码示例:
"""
def gen_name(id):
return "x" + str(id)
pattern = PaddleGraph(graph_type="dygraph")
pattern.add_layer(
"paddle.static.create_parameter",
inputs={},
outputs=[gen_name(0)])
pattern.add_layer(
"paddle.full",
inputs={},
outputs=[gen_name(1)],
shape=[1])
pattern.add_layer(
"paddle.add",
inputs={"x": gen_name(0), "y": gen_name(1)},
outputs=[gen_name(2)])
pattern.add_layer(
"paddle.rsqrt",
inputs={"x": gen_name(2)},
outputs=[gen_name(3)])
pattern.add_layer(
"paddle.static.create_parameter",
inputs={},
outputs=[gen_name(4)])
pattern.add_layer(
"paddle.multiply",
inputs={"x": gen_name(3), "y": gen_name(4)},
outputs=[gen_name(5)])
pattern.add_layer(
"paddle.static.create_parameter",
inputs={},
outputs=[gen_name(6)])
pattern.add_layer(
"paddle.multiply",
inputs={"x": gen_name(6), "y": gen_name(5)},
outputs=[gen_name(7)])
pattern.add_layer(
"paddle.static.create_parameter",
inputs={},
outputs=[gen_name(8)])
pattern.add_layer(
"fluid.layers.elementwise_sub",
inputs={"x": gen_name(8), "y": gen_name(7)},
outputs=[gen_name(9)])
pattern.add_layer(
"paddle.multiply",
inputs={"x": "bn-input-0", "y": gen_name(5)},
outputs=[gen_name(10)])
pattern.add_layer(
"paddle.add",
inputs={"x": gen_name(10), "y": gen_name(9)},
outputs=[gen_name(11)])
pattern.build(inputs={"input-0": "bn-input-0", })
self.patterns.append(pattern)
pattern = PaddleGraph(graph_type="dygraph")
pattern.add_layer(
"paddle.static.create_parameter",
inputs={},
outputs=[gen_name(0)])
pattern.add_layer(
"paddle.full",
inputs={},
outputs=[gen_name(1)],
shape=[1])
pattern.add_layer(
"paddle.add",
inputs={"x": gen_name(0), "y": gen_name(1)},
outputs=[gen_name(2)])
pattern.add_layer(
"paddle.rsqrt",
inputs={"x": gen_name(2)},
outputs=[gen_name(3)])
pattern.add_layer(
"paddle.static.create_parameter",
inputs={},
outputs=[gen_name(4)])
pattern.add_layer(
"paddle.multiply",
inputs={"x": gen_name(3), "y": gen_name(4)},
outputs=[gen_name(5)])
pattern.add_layer(
"paddle.multiply",
inputs={"x": "bn-input-0", "y": gen_name(5)},
outputs=[gen_name(10)])
pattern.add_layer(
"paddle.static.create_parameter",
inputs={},
outputs=[gen_name(6)])
pattern.add_layer(
"paddle.multiply",
inputs={"x": gen_name(6), "y": gen_name(5)},
outputs=[gen_name(7)])
pattern.add_layer(
"paddle.static.create_parameter",
inputs={},
outputs=[gen_name(8)])
pattern.add_layer(
"fluid.layers.elementwise_sub",
inputs={"x": gen_name(8), "y": gen_name(7)},
outputs=[gen_name(9)])
pattern.add_layer(
"paddle.add",
inputs={"x": gen_name(10), "y": gen_name(9)},
outputs=[gen_name(11)])
pattern.build(inputs={"input-0": "bn-input-0", })
self.patterns.append(pattern)
def insert_new_layer(self, graph, parameters, matches):
new_layers, last_layer_id = self.gen_new_layer(matches, parameters, graph)
matches_copy = copy.deepcopy(matches)
for layer_id, layer in matches_copy.items():
for i in range(7):
if layer_id == new_layers[i].id:
matches.pop(new_layers[i].id)
prefix_layers = OrderedDict()
mid_layers = OrderedDict()
suffix_layers = OrderedDict()
is_need_id = False
for layer_id, layer in graph.layers.items():
if is_need_id:
suffix_layers[layer_id] = layer
else:
if layer_id == last_layer_id:
for i in range(7):
mid_layers[new_layers[i].id] = new_layers[i]
is_need_id = True
prefix_layers[layer_id] = layer
prefix_layers.update(mid_layers)
prefix_layers.update(suffix_layers)
graph.layers = prefix_layers
def gen_new_layer(self, matches, parameters, graph):
layer_id_list = list(matches.keys())
layer_id_list.sort(key = int)
for layer_id, layer in matches.items():
if layer.kernel == "paddle.full":
full_layer = layer
out_layer_id = graph.edges_out[layer_id][0]
if matches[out_layer_id].kernel == "paddle.add":
var_layer_id = graph.edges_in[out_layer_id][0]
var_layer = matches[var_layer_id]
if layer.kernel == "paddle.rsqrt":
out_layer_id = graph.edges_out[layer_id][0]
if matches[out_layer_id].kernel == "paddle.multiply":
gamma_layer_id = graph.edges_in[out_layer_id][1]
gamma_layer = matches[gamma_layer_id]
if layer.kernel == "fluid.layers.elementwise_sub":
in_layer_id = graph.edges_in[layer_id][0]
beta_layer = matches[in_layer_id]
in_layer_id = graph.edges_in[layer_id][1]
in_layer_id = graph.edges_in[in_layer_id][0]
mean_layer = matches[in_layer_id]
out_layer_id = graph.edges_out[layer_id][0]
add_layer = matches[out_layer_id]
if layer.kernel == "paddle.multiply":
in_layer_id = graph.edges_in[layer_id][1]
mul_layer = matches[in_layer_id]
if mul_layer.kernel == "paddle.multiply":
in_layer_id = graph.edges_in[layer_id][0]
if in_layer_id not in matches:
input_name = layer.inputs["x"]
transpose0 = PaddleLayer(
id=layer_id_list[-1] + "_1",
kernel="paddle.transpose",
inputs={"x": input_name},
outputs=["{}_transpose_for_bn".format(input_name)],
perm=[0, 3, 1, 2])
params = parameters[gamma_layer.outputs[0]]
c = params.shape[0]
bn = PaddleLayer(
id=layer_id_list[-1] + "_2",
kernel="paddle.nn.functional.batch_norm",
inputs={"x": "{}_transpose_for_bn".format(input_name),
"running_mean": mean_layer.outputs[0],
"running_var": var_layer.outputs[0],
"weight": gamma_layer.outputs[0],
"bias": beta_layer.outputs[0]},
outputs=["{}_bn".format(input_name)],
epsilon=full_layer.attrs["fill_value"])
transpose1 = PaddleLayer(
id=layer_id_list[-1] + "_3",
kernel="paddle.transpose",
inputs={"x": "{}_bn".format(input_name)},
outputs=add_layer.outputs,
perm=[0, 2, 3, 1])
mean_layer.id = layer_id_list[-1] + "_01"
var_layer.id = layer_id_list[-1] + "_02"
gamma_layer.id = layer_id_list[-1] + "_03"
beta_layer.id = layer_id_list[-1] + "_04"
return [mean_layer, var_layer, gamma_layer, beta_layer,
transpose0, bn, transpose1], layer_id_list[-1]
......@@ -16,13 +16,13 @@ from x2paddle.optimizer.pass_manager import PassManager
from x2paddle.optimizer.fusion.dygraph import *
from x2paddle.optimizer.fusion.static import *
from x2paddle.optimizer.elimination.dygraph import *
from x2paddle.optimizer.elimination.static import *
class GraphOptimizer(object):
def __init__(self, source_frame, paddle_type="dygraph", jit_type="trace"):
if source_frame == "pytorch":
if jit_type == "trace":
self.passes = ["dygraph_constant_fuse_pass",
"trace_fc_fuse_pass"]
self.passes = ["trace_fc_fuse_pass"]
else:
self.passes = [
"dygraph_constant_fuse_pass",
......@@ -39,12 +39,20 @@ class GraphOptimizer(object):
else:
self.passes = ["static_bn_scale_fuse_pass"]
elif source_frame == "tf":
self.passes = [
"dygraph_conv2d_add_fuse_pass",
"dygraph_tf_batchnorm_fuse_pass",
"dygraph_prelu_fuse_pass",
"transpose_eliminate_pass"
]
if paddle_type == "dygraph":
self.passes = [
"dygraph_conv2d_add_fuse_pass",
"dygraph_tf_batchnorm_fuse_pass",
"dygraph_prelu_fuse_pass",
"transpose_eliminate_pass"
]
else:
self.passes = [
"static_conv2d_add_fuse_pass",
"static_tf_batchnorm_fuse_pass",
"static_prelu_fuse_pass",
"static_transpose_eliminate_pass"
]
else:
self.passes = []
......
import copy
from collections import OrderedDict
from x2paddle.core.program import PaddleLayer
class BatchNormOpt:
def __init__(self):
pass
def run(self, graph):
print("Optimize: BatchNormOpt...")
layers = copy.deepcopy(graph.layers)
for layer_id, layer in layers.items():
if layer.kernel != "fluid.layers.elementwise_add":
continue
axis = layer.attrs.get('axis', -1)
if axis != -1 and axis != 3:
continue
input_ids0 = graph.edges_in[layer_id]
mul_layer0 = graph.layers[input_ids0[0]]
sub_layer0 = graph.layers[input_ids0[1]]
if mul_layer0.kernel != "fluid.layers.elementwise_mul":
continue
if sub_layer0.kernel != "fluid.layers.elementwise_sub":
continue
axis = mul_layer0.attrs.get('axis', -1)
if axis != -1 and axis != 3:
continue
axis = sub_layer0.attrs.get('axis', -1)
if axis != -1 and axis != 0:
continue
if len(graph.edges_out.get(input_ids0[0], [])) != 1:
continue
if len(graph.edges_out.get(input_ids0[1], [])) != 1:
continue
input_ids1 = graph.edges_in[input_ids0[0]]
nhwc_input = graph.layers[input_ids1[0]]
mul_layer1 = graph.layers[input_ids1[1]]
if mul_layer1.kernel != "fluid.layers.elementwise_mul":
continue
axis = mul_layer1.attrs.get('axis', -1)
if axis != -1 and axis != 0:
continue
if len(graph.edges_out.get(input_ids1[1], [])) != 2:
continue
input_ids2 = graph.edges_in[input_ids0[1]]
beta = graph.layers[input_ids2[0]]
mul_layer2 = graph.layers[input_ids2[1]]
if beta.kernel != "fluid.layers.create_parameter":
continue
axis = mul_layer2.attrs.get('axis', -1)
if axis != -1 and axis != 0:
continue
if len(graph.edges_out.get(input_ids2[0], [])) != 1:
continue
if len(graph.edges_out.get(input_ids2[1], [])) != 1:
continue
if beta.outputs[0] not in graph.parameters:
continue
beta_shape = graph.parameters[beta.outputs[0]].shape
if len(beta_shape) != 1:
continue
input_ids3 = graph.edges_in[input_ids2[1]]
mean = graph.layers[input_ids3[0]]
mul_layer3 = graph.layers[input_ids3[1]]
if mean.kernel != "fluid.layers.create_parameter":
continue
axis = mul_layer3.attrs.get('axis', -1)
if axis != -1 and axis != 0:
continue
if len(graph.edges_out.get(input_ids3[0], [])) != 1:
continue
if len(graph.edges_out.get(input_ids3[1], [])) != 2:
continue
if mul_layer3.id != mul_layer1.id:
continue
if mean.outputs[0] not in graph.parameters:
continue
mean_shape = graph.parameters[mean.outputs[0]].shape
if mean_shape != beta_shape:
continue
input_ids4 = graph.edges_in[input_ids3[1]]
rsqrt_layer = graph.layers[input_ids4[0]]
gamma = graph.layers[input_ids4[1]]
if rsqrt_layer.kernel != "fluid.layers.rsqrt":
continue
if gamma.kernel != "fluid.layers.create_parameter":
continue
if len(graph.edges_out.get(input_ids4[0], [])) != 1:
continue
if len(graph.edges_out.get(input_ids4[1], [])) != 1:
continue
if gamma.outputs[0] not in graph.parameters:
continue
gamma_shape = graph.parameters[gamma.outputs[0]].shape
if gamma_shape != beta_shape:
continue
input_ids5 = graph.edges_in[input_ids4[0]]
add_layer = graph.layers[input_ids5[0]]
if add_layer.kernel != "fluid.layers.elementwise_add":
continue
axis = add_layer.attrs.get('axis', -1)
if axis != -1 and axis != 0:
continue
if len(graph.edges_out.get(input_ids5[0], [])) != 1:
continue
input_ids6 = graph.edges_in[input_ids5[0]]
variance = graph.layers[input_ids6[0]]
other = graph.layers[input_ids6[1]]
if variance.kernel != "fluid.layers.create_parameter":
continue
if other.kernel != "fluid.layers.fill_constant":
continue
if len(graph.edges_out.get(input_ids6[0], [])) != 1:
continue
if len(graph.edges_out.get(input_ids6[1], [])) != 1:
continue
if variance.outputs[0] not in graph.parameters:
continue
variance_shape = graph.parameters[variance.outputs[0]].shape
if variance_shape != beta_shape:
continue
ids = set([
layer_id, mul_layer0.id, sub_layer0.id, mul_layer1.id, beta.id,
mul_layer2.id, mean.id, mul_layer2.id, rsqrt_layer.id, gamma.id,
add_layer.id, variance.id, other.id
])
for id in ids:
del graph.layers[id]
if id in graph.edges_in:
del graph.edges_in[id]
if id in graph.edges_out:
del graph.edges_out[id]
copy_layers = copy.deepcopy(graph.layers)
graph.layers = OrderedDict()
for k, v in copy_layers.items():
if k != nhwc_input.id:
graph.layers[k] = v
continue
graph.layers[k] = v
transpose0 = PaddleLayer(
id='{}_1'.format(k),
kernel="fluid.layers.transpose",
inputs={"x": v.outputs[0]},
outputs=["transpose_for_bn"],
perm=[0, 3, 1, 2])
bn = PaddleLayer(
id='{}_2'.format(k),
kernel="fluid.layers.batch_norm",
inputs={"input": "transpose_for_bn"},
outputs=layer.outputs,
epsilon=other.attrs["value"],
param_attr="'{}'".format(gamma.outputs[0]),
bias_attr="'{}'".format(beta.outputs[0]),
moving_mean_name="'{}'".format(mean.outputs[0]),
moving_variance_name="'{}'".format(variance.outputs[0]))
transpose1 = PaddleLayer(
id=layer_id,
kernel="fluid.layers.transpose",
inputs={"x": layer.outputs[0]},
outputs=layer.outputs,
perm=[0, 2, 3, 1])
graph.layers[transpose0.id] = transpose0
graph.layers[bn.id] = bn
graph.layers[transpose1.id] = transpose1
graph.build()
import copy
class BiasOpt:
def __init__(self):
self.conv_layers = [
'fluid.layers.conv2d', 'fluid.layers.conv2d_transpose'
]
def run(self, graph):
print("Optimize: BiasOpt...")
layers = copy.deepcopy(graph.layers)
for layer_id, layer in layers.items():
if layer.kernel in self.conv_layers or layer.kernel == "fluid.layers.transpose":
if len(graph.edges_out.get(layer_id, [])) > 1:
continue
if layer.outputs[0] in graph.outputs:
continue
out_layer_id = graph.edges_out[layer_id][0]
if graph.layers[
out_layer_id].kernel != "fluid.layers.elementwise_add":
continue
if graph.layers[out_layer_id].attrs.get('axis', -1) != -1:
continue
in_layer_id = graph.edges_in[out_layer_id]
bias_layer_id = in_layer_id[1 - in_layer_id.index(layer_id)]
if graph.layers[
bias_layer_id].kernel != "fluid.layers.create_parameter":
continue
bias_layer = graph.layers[bias_layer_id]
if len(bias_layer.attrs['shape']) != 1:
continue
if len(graph.edges_out[bias_layer_id]) != 1:
continue
if layer.kernel == "fluid.layers.transpose":
if layer.attrs['perm'] != [0, 2, 3, 1]:
continue
in_layer_id = graph.edges_in[layer_id][0]
if graph.layers[in_layer_id].kernel not in self.conv_layers:
continue
if graph.layers[in_layer_id].attrs['bias_attr'] != False:
continue
if graph.layers[in_layer_id].outputs[0] in graph.outputs:
continue
if len(graph.edges_out[in_layer_id]) != 1:
continue
graph.layers[in_layer_id].attrs[
'bias_attr'] = bias_layer.attrs['name']
else:
graph.layers[layer_id].attrs[
'bias_attr'] = bias_layer.attrs['name']
bias_add_outs = graph.edges_out.get(out_layer_id, [])
bias_add_output = graph.layers[out_layer_id].outputs[0]
graph.del_layer(bias_layer_id)
graph.del_layer(out_layer_id)
for out in bias_add_outs:
for k, v in graph.layers[out].inputs.items():
if v == layer.outputs[0]:
graph.layers[out].inputs[k] = bias_add_output
graph.layers[layer_id].outputs[0] = bias_add_output
if layer.kernel == "fluid.layers.transpose":
in_layer_id = graph.edges_in[layer_id][0]
graph.layers[in_layer_id].outputs[0] = bias_add_output
graph.layers[layer_id].inputs['x'] = bias_add_output
import copy
import numpy as np
from collections import OrderedDict
from x2paddle.core.program import PaddleLayer
from x2paddle.core.util import *
class PReLUOpt:
def __init__(self):
pass
def run(self, graph):
print("Optimize: PReLUOpt...")
layers = copy.deepcopy(graph.layers)
for layer_id, layer in layers.items():
if layer.kernel != "fluid.layers.elementwise_add":
continue
axis = layer.attrs.get('axis', -1)
if axis != -1 and axis != 3:
continue
input_ids0 = graph.edges_in[layer_id]
relu_layer0 = graph.layers[input_ids0[0]]
mul_layer0 = graph.layers[input_ids0[1]]
if relu_layer0.kernel != "fluid.layers.relu":
continue
if mul_layer0.kernel != "fluid.layers.elementwise_mul":
continue
axis = mul_layer0.attrs.get('axis', -1)
if axis != -1 and axis != 3:
continue
if len(graph.edges_out.get(input_ids0[0], [])) != 1:
continue
if len(graph.edges_out.get(input_ids0[1], [])) != 1:
continue
input_ids1_0 = graph.edges_in[input_ids0[0]]
input_ids1_1 = graph.edges_in[input_ids0[1]]
fill_layer = graph.layers[input_ids1_1[1]]
mul_layer1 = graph.layers[input_ids1_1[0]]
if fill_layer.kernel != "fluid.layers.fill_constant":
continue
if mul_layer1.kernel != "fluid.layers.elementwise_mul":
continue
axis = mul_layer1.attrs.get('axis', -1)
if axis != -1 and axis != 0:
continue
if len(graph.edges_out.get(input_ids1_1[1], [])) != 1:
continue
if len(graph.edges_out.get(input_ids1_0[0], [])) != 3:
continue
input_ids2 = graph.edges_in[input_ids1_1[0]]
alpha = graph.layers[input_ids2[0]]
sub_layer = graph.layers[input_ids2[1]]
if alpha.kernel != "fluid.layers.create_parameter":
continue
if sub_layer.kernel != "fluid.layers.elementwise_sub":
continue
axis = sub_layer.attrs.get('axis', -1)
if axis != -1 and axis != 3:
continue
if len(graph.edges_out.get(input_ids2[0], [])) != 1:
continue
if len(graph.edges_out.get(input_ids2[1], [])) != 1:
continue
if alpha.outputs[0] not in graph.parameters:
continue
input_ids3 = graph.edges_in[input_ids2[1]]
add_layer = graph.layers[input_ids3[0]]
abs_layer = graph.layers[input_ids3[1]]
if abs_layer.kernel != "fluid.layers.abs":
continue
if len(graph.edges_out.get(input_ids3[1], [])) != 1:
continue
ids = set([
layer.id, relu_layer0.id, mul_layer0.id, fill_layer.id, mul_layer1.id, alpha.id,
sub_layer.id, abs_layer.id])
for id in ids:
del graph.layers[id]
if id in graph.edges_in:
del graph.edges_in[id]
if id in graph.edges_out:
del graph.edges_out[id]
copy_layers = copy.deepcopy(graph.layers)
graph.layers = OrderedDict()
for k, v in copy_layers.items():
if k != add_layer.id:
graph.layers[k] = v
continue
graph.layers[k] = v
transpose0 = PaddleLayer(
id='{}_1'.format(k),
kernel="fluid.layers.transpose",
inputs={"x": v.outputs[0]},
outputs=["transpose_for_prelu"],
perm=[0, 3, 1, 2])
prelu = PaddleLayer(
id='{}_2'.format(k),
kernel="fluid.layers.prelu",
inputs={"x": "transpose_for_prelu"},
outputs=layer.outputs,
mode=string("channel"),
param_attr="'{}'".format(alpha.outputs[0]))
transpose1 = PaddleLayer(
id=layer_id,
kernel="fluid.layers.transpose",
inputs={"x": layer.outputs[0]},
outputs=layer.outputs,
perm=[0, 2, 3, 1])
graph.layers[transpose0.id] = transpose0
graph.layers[prelu.id] = prelu
graph.layers[transpose1.id] = transpose1
first_axis = graph.parameters[alpha.outputs[0]].shape[0]
graph.parameters[alpha.outputs[0]] = np.reshape(graph.parameters[alpha.outputs[0]], (1, first_axis, 1, 1))
graph.build()
\ No newline at end of file
Markdown is supported
0% .
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