Skip to content

X
X2Paddle

PaddlePaddle / X2Paddle
大约 1 年 前同步成功

通知 328
Star 698
Fork 167
X
X2Paddle
未验证 提交 2228423e 编写于 作者: J Jason 提交者： GitHub
浏览文件
操作

Merge pull request #16 from MacroBull/master 

Rename onnx2paddle to onnx2fluid
上级 a4796334 f0dede1f
展开全部 隐藏空白更改
内联 并排
Showing with 4323 addition and 1582 deletion
onnx2paddle/.gitignore → onnx2fluid/.gitignore
浏览文件 @ 2228423e
......@@ -57,3 +57,4 @@ coverage.xml
/examples/*.aria2
/examples/*.onnx
/examples/*.np?
**/.*
onnx2paddle/README.md → onnx2fluid/README.md
浏览文件 @ 2228423e
Onnx2paddle
Onnx2Fluid
===
Inference model conversion from ONNX/PyTorch to Paddle
Inference model conversion from ONNX/PyTorch to Paddle fluid
快速开始
---
......
onnx2paddle/examples/convert_data_npz_0.py → onnx2fluid/examples/convert_data_npz_0.py
浏览文件 @ 2228423e
......@@ -22,4 +22,4 @@ output_data = data['outputs']
inputs = Dict(zip(input_names, [input_data]))
outputs = Dict(zip(output_name, [output_data]))
np.savez(fn, inputs=inputs, outputs=outputs) # overwrite
np.savez(fn, inputs=inputs, outputs=outputs) # overwrite
onnx2paddle/examples/gen_some_samples.py → onnx2fluid/examples/gen_some_samples.py
浏览文件 @ 2228423e
......@@ -6,7 +6,7 @@ Created on Fri Mar 22 11:19:45 2019
@author: Macrobull
Not all ops in this file are supported by both Pytorch and ONNX
This only demostrates the conversion/validation workflow from Pytorch to ONNX to Paddle
This only demostrates the conversion/validation workflow from Pytorch to ONNX to Paddle fluid
"""
......@@ -16,12 +16,10 @@ import torch
import torch.nn as nn
import torch.nn.functional as F
from onnx2paddle.torch_export_helper import export_onnx_with_validation
from onnx2fluid.torch_export_helper import export_onnx_with_validation
idx = 0
######### example: RNN ########
#
#class Model(nn.Module):
......@@ -44,7 +42,6 @@ idx = 0
# ['x'], ['y'],
# verbose=True, training=False)
######### example: random ########
#
#class Model(nn.Module):
......@@ -66,9 +63,9 @@ idx = 0
# ['x'], ['y'],
# verbose=True, training=False)
######## example: fc ########
class Model(nn.Module):
def __init__(self):
super(Model, self).__init__()
......@@ -85,13 +82,12 @@ xb = torch.rand((2, 3))
yp = model(xb)
idx += 1
print('index: ', idx)
export_onnx_with_validation(model, (xb, ), 't' + str(idx),
['x'], ['y'],
verbose=True, training=False)
export_onnx_with_validation(
model, (xb, ), 't' + str(idx), ['x'], ['y'], verbose=True, training=False)
######## example: compare ########
class Model(nn.Module):
def __init__(self):
super(Model, self).__init__()
......@@ -110,12 +106,15 @@ xb1 = torch.rand((2, 3))
ya, yb, yc = model(xb0, xb1)
idx += 1
print('index: ', idx)
export_onnx_with_validation(model, (xb0, xb1), 't' + str(idx),
['x0', 'x1'], ['ya', 'yb', 'yc'],
verbose=True, training=False)
export_onnx_with_validation(
model, (xb0, xb1),
't' + str(idx), ['x0', 'x1'], ['ya', 'yb', 'yc'],
verbose=True,
training=False)
######## example: affine_grid ########
class Model(nn.Module):
def __init__(self):
super(Model, self).__init__()
......@@ -130,13 +129,15 @@ theta = torch.rand((2, 2, 3))
grid = model(theta)
idx += 1
print('index: ', idx)
export_onnx_with_validation(model, (theta, ), 't' + str(idx),
['theta'], ['grid'],
verbose=True, training=False)
export_onnx_with_validation(
model, (theta, ),
't' + str(idx), ['theta'], ['grid'],
verbose=True,
training=False)
######## example: conv2d_transpose ########
class Model(nn.Module):
def __init__(self):
super(Model, self).__init__()
......@@ -155,12 +156,12 @@ xb = torch.rand((2, 3, 4, 5))
yp = model(xb)
idx += 1
print('index: ', idx)
export_onnx_with_validation(model, (xb, ), 't' + str(idx),
['x'], ['y'],
verbose=True, training=False)
export_onnx_with_validation(
model, (xb, ), 't' + str(idx), ['x'], ['y'], verbose=True, training=False)
######## example: conv2d ########
class Model(nn.Module):
def __init__(self):
super(Model, self).__init__()
......@@ -181,10 +182,8 @@ xb = torch.rand((2, 3, 4, 5))
yp = model(xb)
idx += 1
print('index: ', idx)
export_onnx_with_validation(model, (xb, ), 't' + str(idx),
['x'], ['y'],
verbose=True, training=False)
export_onnx_with_validation(
model, (xb, ), 't' + str(idx), ['x'], ['y'], verbose=True, training=False)
######### example: conv1d ########
#
......@@ -210,6 +209,7 @@ export_onnx_with_validation(model, (xb, ), 't' + str(idx),
######## example: empty ########
class Model(nn.Module):
def __init__(self):
super(Model, self).__init__()
......@@ -223,6 +223,5 @@ xb = torch.rand((2, 3))
yp = model(xb)
idx += 1
print('index: ', idx)
export_onnx_with_validation(model, (xb, ), 't' + str(idx),
['y'], ['y'],
verbose=True, training=False)
export_onnx_with_validation(
model, (xb, ), 't' + str(idx), ['y'], ['y'], verbose=True, training=False)
onnx2paddle/examples/onnx_model_zoo.sh → onnx2fluid/examples/onnx_model_zoo.sh
浏览文件 @ 2228423e
#! /usr/bin/env sh
get_url="proxychains4 aria2c -c -s8 -x8"
get_url="aria2c -c -s8 -x8"
base_url="https://s3.amazonaws.com/download.onnx/models/opset_9/"
flags="-de -o /tmp/export/"
flags="-e -o /tmp/export/"
bvlc_alexnet()
{
......@@ -18,13 +18,13 @@ bvlc_alexnet()
do
echo "converting $npz ..."
python convert_data_npz_0.py "$npz" "data_0" "prob_1"
python -m onnx2paddle $flags "$fn_model" -t $npz
python -m onnx2fluid $flags "$fn_model" -t $npz
done
for pb_dir in $bn_tar/*/
do
echo "converting $pb_dir ..."
python convert_data_pb_0.py "$pb_dir" "data_0" "prob_1"
python -m onnx2paddle $flags "$fn_model" -t echo $(dirname "$pb_dir/x").npz
python -m onnx2fluid $flags "$fn_model" -t $(dirname "$pb_dir/x").npz
done
}
......@@ -42,7 +42,7 @@ bvlc_googlenet()
do
echo "converting $pb_dir"
python convert_data_pb_0.py "$pb_dir" "data_0" "prob_1"
python -m onnx2paddle $flags "$fn_model" -t $(dirname "$pb_dir/x").npz
python -m onnx2fluid $flags "$fn_model" -t $(dirname "$pb_dir/x").npz
done
}
......@@ -60,7 +60,7 @@ bvlc_reference_caffenet()
do
echo "converting $pb_dir"
python convert_data_pb_0.py "$pb_dir" "data_0" "prob_1"
python -m onnx2paddle $flags "$fn_model" -t $(dirname "$pb_dir/x").npz
python -m onnx2fluid $flags "$fn_model" -t $(dirname "$pb_dir/x").npz
done
}
......@@ -69,7 +69,7 @@ bvlc_reference_rcnn_ilsvrc13()
bn_tar="bvlc_reference_rcnn_ilsvrc13"
fn_tar="$bn_tar.tar.gz"
fn_model="$bn_tar/model.onnx"
$get_url "$base_url$fn_tar"
echo "extracting ..."
tar xf "$fn_tar"
......@@ -77,8 +77,8 @@ bvlc_reference_rcnn_ilsvrc13()
for pb_dir in $bn_tar/*/
do
echo "converting $pb_dir"
python convert_data_pb_0.py "$pb_dir" "data_0" "softmaxout_1"
python -m onnx2paddle $flags "$fn_model" -t $(dirname "$pb_dir/x").npz
python convert_data_pb_0.py "$pb_dir" "data_0" "fc_rcnn_1"
python -m onnx2fluid $flags "$fn_model" -t $(dirname "$pb_dir/x").npz
done
}
......@@ -87,7 +87,7 @@ inception_v1()
bn_tar="inception_v1"
fn_tar="$bn_tar.tar.gz"
fn_model="$bn_tar/model.onnx"
$get_url "$base_url$fn_tar"
echo "extracting ..."
tar xf "$fn_tar"
......@@ -96,14 +96,14 @@ inception_v1()
do
echo "converting $npz ..."
python convert_data_npz_0.py "$npz" "data_0" "prob_1"
python -m onnx2paddle $flags "$fn_model" -t $npz
python -m onnx2fluid $flags "$fn_model" -t $npz
done
for pb_dir in $bn_tar/*/
do
echo "converting $pb_dir"
python convert_data_pb_0.py "$pb_dir" "data_0" "prob_1"
python -m onnx2paddle $flags "$fn_model" -t $(dirname "$pb_dir/x").npz
python -m onnx2fluid $flags "$fn_model" -t $(dirname "$pb_dir/x").npz
done
}
......@@ -112,7 +112,7 @@ inception_v2()
bn_tar="inception_v2"
fn_tar="$bn_tar.tar.gz"
fn_model="$bn_tar/model.onnx"
$get_url "$base_url$fn_tar"
echo "extracting ..."
tar xf "$fn_tar"
......@@ -121,14 +121,14 @@ inception_v2()
do
echo "converting $npz ..."
python convert_data_npz_0.py "$npz" "data_0" "prob_1"
python -m onnx2paddle $flags "$fn_model" -t $npz
python -m onnx2fluid $flags "$fn_model" -t $npz
done
for pb_dir in $bn_tar/*/
do
echo "converting $pb_dir"
python convert_data_pb_0.py "$pb_dir" "data_0" "prob_1"
python -m onnx2paddle $flags "$fn_model" -t $(dirname "$pb_dir/x").npz
python -m onnx2fluid $flags "$fn_model" -t $(dirname "$pb_dir/x").npz
done
}
......@@ -137,7 +137,7 @@ resnet50()
bn_tar="resnet50"
fn_tar="$bn_tar.tar.gz"
fn_model="$bn_tar/model.onnx"
$get_url "$base_url$fn_tar"
echo "extracting ..."
tar xf "$fn_tar"
......@@ -146,14 +146,14 @@ resnet50()
do
echo "converting $npz ..."
python convert_data_npz_0.py "$npz" "gpu_0/data_0" "gpu_0/softmaxout_1"
python -m onnx2paddle $flags "$fn_model" -t $npz
python -m onnx2fluid $flags "$fn_model" -t $npz
done
for pb_dir in $bn_tar/*/
do
echo "converting $pb_dir"
python convert_data_pb_0.py "$pb_dir" "gpu_0/data_0" "gpu_0/softmaxout_1"
python -m onnx2paddle $flags "$fn_model" -t $(dirname "$pb_dir/x").npz
python -m onnx2fluid $flags "$fn_model" -t $(dirname "$pb_dir/x").npz
done
}
......@@ -162,7 +162,7 @@ shufflenet()
bn_tar="shufflenet"
fn_tar="$bn_tar.tar.gz"
fn_model="$bn_tar/model.onnx"
$get_url "$base_url$fn_tar"
echo "extracting ..."
tar xf "$fn_tar"
......@@ -171,7 +171,7 @@ shufflenet()
do
echo "converting $pb_dir"
python convert_data_pb_0.py "$pb_dir" "gpu_0/data_0" "gpu_0/softmaxout_1"
python -m onnx2paddle $flags "$fn_model" -t $(dirname "$pb_dir/x").npz
python -m onnx2fluid $flags "$fn_model" -t $(dirname "$pb_dir/x").npz
done
}
......@@ -180,7 +180,7 @@ squeezenet()
bn_tar="squeezenet"
fn_tar="$bn_tar.tar.gz"
fn_model="$bn_tar/model.onnx"
$get_url "$base_url$fn_tar"
echo "extracting ..."
tar xf "$fn_tar"
......@@ -189,7 +189,7 @@ squeezenet()
do
echo "converting $pb_dir"
python convert_data_pb_0.py "$pb_dir" "data_0" "softmaxout_1"
python -m onnx2paddle $flags "$fn_model" -t $(dirname "$pb_dir/x").npz
python -m onnx2fluid $flags "$fn_model" -t $(dirname "$pb_dir/x").npz
done
}
......@@ -198,7 +198,7 @@ tiny_yolov2()
bn_tar="tiny_yolov2"
fn_tar="$bn_tar.tar.gz"
fn_model="$bn_tar/model.onnx"
$get_url "https://onnxzoo.blob.core.windows.net/models/opset_8/tiny_yolov2/$fn_tar"
echo "extracting ..."
tar xf "$fn_tar"
......@@ -207,7 +207,7 @@ tiny_yolov2()
do
echo "converting $pb_dir"
python convert_data_pb_0.py "$pb_dir" "image" "grid"
python -m onnx2paddle $flags "$fn_model" -t $(dirname "$pb_dir/x").npz -x
python -m onnx2fluid $flags "$fn_model" -t $(dirname "$pb_dir/x").npz -x
done
}
......@@ -216,7 +216,7 @@ vgg19()
bn_tar="vgg19"
fn_tar="$bn_tar.tar.gz"
fn_model="$bn_tar/model.onnx"
$get_url "$base_url$fn_tar"
echo "extracting ..."
tar xf "$fn_tar"
......@@ -225,7 +225,7 @@ vgg19()
do
echo "converting $pb_dir"
python convert_data_pb_0.py "$pb_dir" "data_0" "prob_1"
python -m onnx2paddle $flags "$fn_model" -t $(dirname "$pb_dir/x").npz
python -m onnx2fluid $flags "$fn_model" -t $(dirname "$pb_dir/x").npz
done
}
......@@ -234,7 +234,7 @@ zfnet512()
bn_tar="zfnet512"
fn_tar="$bn_tar.tar.gz"
fn_model="$bn_tar/model.onnx"
$get_url "$base_url$fn_tar"
echo "extracting ..."
tar xf "$fn_tar"
......@@ -243,20 +243,20 @@ zfnet512()
do
echo "converting $pb_dir"
python convert_data_pb_0.py "$pb_dir" "gpu_0/data_0" "gpu_0/softmax_1"
python -m onnx2paddle $flags "$fn_model" -t $(dirname "$pb_dir/x").npz
python -m onnx2fluid $flags "$fn_model" -t $(dirname "$pb_dir/x").npz
done
}
bvlc_alexnet # data error
bvlc_googlenet # desc error
bvlc_alexnet
bvlc_googlenet
bvlc_reference_caffenet
bvlc_reference_rcnn_ilsvrc13
inception_v1 ###
inception_v2 ###
resnet50 # data error
shufflenet ###
inception_v1
inception_v2
resnet50
shufflenet
squeezenet
tiny_yolov2 # not supported
vgg19
zfnet512 # data error
zfnet512
onnx2paddle/onnx2paddle/__main__.py → onnx2fluid/onnx2fluid/__main__.py
浏览文件 @ 2228423e
......@@ -5,7 +5,7 @@
#
################################################################################
"""
本文件允许模块包以python -m onnx2paddle方式直接执行。
本文件允许模块包以python -m onnx2fluid方式直接执行。
Authors: Macrobull
Date: 2019/02/22 10:25:46
......@@ -21,43 +21,67 @@ import argparse
import logging
import sys
parser = argparse.ArgumentParser(description='onnx2paddle',
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
parser.add_argument('model', nargs=1,
help='path to model.onnx',
)
parser.add_argument('--debug', '-d', action='store_true',
help='enable debug logging and checking',
)
parser.add_argument('--output-dir', '-o', type=str, default='',
help='output directory',
)
parser.add_argument('--test_data', '-t', type=str, default='',
help='I/O golden data for validation, e.g. test.npy, test.npz',
)
parser.add_argument('--embed_params', '-e', action='store_true',
help='try to embed parameters for trainable Paddle layers',
)
parser.add_argument('--pedantic', action='store_true', default=True,
help='accept and convert only standard ONNX opset',
)
parser.add_argument('--no-pedantic', '-x', action='store_false',
dest='pedantic',
help='process non-standard ONNX ops, this may lead to fails',
)
parser.add_argument('--precision', '-p', type=int, default=4,
help='assertion decimal for validation',
)
parser = argparse.ArgumentParser(
description='onnx2fluid',
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
parser.add_argument(
'model',
nargs=1,
help='path to model.onnx',
)
parser.add_argument(
'--debug',
'-d',
action='store_true',
help='enable debug logging and checking',
)
parser.add_argument(
'--output_dir',
'-o',
type=str,
default='',
help='output directory',
)
parser.add_argument(
'--test_data',
'-t',
type=str,
default='',
help='I/O golden data for validation, e.g. test.npy, test.npz',
)
parser.add_argument(
'--embed_params',
'-e',
action='store_true',
help='try to embed parameters for trainable Paddle fluid layers',
)
parser.add_argument(
'--pedantic',
action='store_true',
default=True,
help='accept and convert only standard ONNX opset',
)
parser.add_argument(
'--no-pedantic',
'-x',
action='store_false',
dest='pedantic',
help='process non-standard ONNX ops, this may lead to fails',
)
parser.add_argument(
'--precision',
'-p',
type=int,
default=4,
help='assertion decimal for validation',
)
args = parser.parse_args()
logging_format = '[%(levelname)8s]%(name)s::%(funcName)s:%(lineno)04d: %(message)s'
logging_level = logging.DEBUG if args.debug else logging.INFO
logging.basicConfig(format=logging_format, level=logging_level)
try:
from . import cmdline
except ImportError:
......@@ -66,5 +90,4 @@ except ImportError:
# imports
main = cmdline.main
sys.exit(main(**args.__dict__))
onnx2paddle/onnx2paddle/cmdline.py → onnx2fluid/onnx2fluid/cmdline.py
浏览文件 @ 2228423e
......@@ -21,7 +21,6 @@ import logging
import shutil
import zipfile
__all__ = [
'main',
]
......@@ -42,7 +41,7 @@ def main(**kwargs):
# imports
convert = conversion.convert
logger = logging.getLogger('onnx2paddle')
logger = logging.getLogger('onnx2fluid')
debug = kwargs.get('debug', False)
# prepare arguments
......@@ -58,13 +57,15 @@ def main(**kwargs):
onnx_opset_pedantic = kwargs.get('pedantic', True)
# convert
convert(filename, save_dir,
model_basename=model_basename,
model_func_name=model_func_name,
embed_params=embed_params,
onnx_opset_version=onnx_opset_version,
onnx_opset_pedantic=onnx_opset_pedantic,
debug=debug)
convert(
filename,
save_dir,
model_basename=model_basename,
model_func_name=model_func_name,
embed_params=embed_params,
onnx_opset_version=onnx_opset_version,
onnx_opset_pedantic=onnx_opset_pedantic,
debug=debug)
# validate
passed = True
......@@ -80,21 +81,23 @@ def main(**kwargs):
# in fact fluid can not fully clear the context
# continuous validation may be inaccurate
precision = 10 ** -kwargs.get('precision', 4)
precision = 10**-kwargs.get('precision', 4)
logger.info('starting validation on desc ...')
passed &= validate(shutil.os.path.join(save_dir, '__model__'),
golden_data_filename,
precision=precision,
)
passed &= validate(
shutil.os.path.join(save_dir, '__model__'),
golden_data_filename,
precision=precision,
)
logger.info('starting validation on code ...')
passed &= validate(shutil.os.path.join(save_dir, model_basename),
golden_data_filename,
model_func_name=model_func_name,
precision=precision,
save_inference_model=debug, # this overwrite desc file for test
)
passed &= validate(
shutil.os.path.join(save_dir, model_basename),
golden_data_filename,
model_func_name=model_func_name,
precision=precision,
save_inference_model=debug, # this overwrite desc file for test
)
if not passed:
logger.error('validation failed, exit')
......@@ -112,20 +115,22 @@ def main(**kwargs):
if __name__ == '__main__':
logging.basicConfig(
format='[%(levelname)8s]%(name)s::%(funcName)s:%(lineno)04d: %(message)s',
level=logging.DEBUG,
)
# main(model=['../examples/t5.onnx'],
# output_dir='/tmp/export/',
# embed_params=False,
# pedantic=False,
# test_data='../examples/t5.npz',
# debug=True)
main(model=['../examples/shufflenet/model.onnx'],
output_dir='/tmp/export/',
embed_params=True,
pedantic=False,
test_data='../examples/shufflenet/test_data_set_0.npz',
debug=True)
format=
'[%(levelname)8s]%(name)s::%(funcName)s:%(lineno)04d: %(message)s',
level=logging.DEBUG,
)
# main(model=['../examples/t5.onnx'],
# output_dir='/tmp/export/',
# embed_params=False,
# pedantic=False,
# test_data='../examples/t5.npz',
# debug=True)
main(
model=['../examples/inception_v2/model.onnx'],
output_dir='/tmp/export/',
embed_params=True,
pedantic=False,
test_data='../examples/inception_v2/test_data_set_2.npz',
debug=True)
onnx2paddle/onnx2paddle/conversion.py → onnx2fluid/onnx2fluid/conversion.py
浏览文件 @ 2228423e
......@@ -12,19 +12,21 @@ from __future__ import division
import logging
import shutil
__all__ = [
'convert',
]
def convert(onnx_model_filename, save_dir,
model_basename='model.py', model_func_name='inference',
def convert(onnx_model_filename,
save_dir,
model_basename='model.py',
model_func_name='inference',
embed_params=False,
onnx_opset_version=9, onnx_opset_pedantic=True,
onnx_opset_version=9,
onnx_opset_pedantic=True,
debug=False):
"""
convert an ONNX model to Paddle Python code and desc pb
convert an ONNX model to Paddle fluid Python code and desc pb
"""
import onnx
......@@ -59,10 +61,11 @@ def convert(onnx_model_filename, save_dir,
logger.info('checking model ...')
check_model(onnx_model)
logger.debug('using opset version: %d', onnx_opset_version)
if onnx_opset_pedantic: # WORKAROUND: RuntimeError: No Adapter For OP
if onnx_opset_pedantic: # WORKAROUND: RuntimeError: No Adapter For OP
onnx_model = convert_version(onnx_model, onnx_opset_version)
else: # TODO: add new argument for this option
logger.warning('opset conversion skipped for onnx_opset_pedantic is OFF')
else: # TODO: add new argument for this option
logger.warning(
'opset conversion skipped for onnx_opset_pedantic is OFF')
onnx_model = polish_model(onnx_model)
except ValidationError as e:
if onnx_opset_pedantic:
......@@ -90,13 +93,13 @@ def convert(onnx_model_filename, save_dir,
onnx.save(model, debug_model_filename + '.optimized_and_inffered.onnx')
# onnx.save(model, '/tmp/export/optimized_and_inffered.onnx')
# I/O instances
# I/O instances
onnx_graph = onnx_model.graph
paddle_program = Program()
paddle_writer = Writer()
fluid_program = Program()
fluid_writer = Writer()
# model components
# graph_name = onnx_graph.name
# graph_name = onnx_graph.name
graph_inputs = [value.name for value in onnx_graph.input]
graph_outputs = [value.name for value in onnx_graph.output]
graph_params = []
......@@ -107,29 +110,37 @@ def convert(onnx_model_filename, save_dir,
for name, weight in graph_weights(onnx_graph):
value_info = graph_value_infos[name]
value_info['embeded_as'] = []
value_info['get_weight'] = lambda: weight.tolist() # lazy getter
value_info['get_weight'] = (lambda w: lambda: w.tolist())(
weight) # lazy getter
logger.info('conversion started')
# op set conversion
# topo = 'backward' if embed_params else 'forward'
# topo = 'backward' if embed_params else 'forward'
topo = 'forward'
for name, domain, op_type, inputs, outputs, attrs in graph_ops(onnx_graph, topo=topo):
for name, domain, op_type, inputs, outputs, attrs in graph_ops(
onnx_graph, topo=topo):
logger.debug('translating op %s %s::%s ...', name, domain, op_type)
if domain == DEFAULT_OP_DOMAIN:
domain = ''
try:
paddle_writer.emit_op(paddle_program, name, domain, op_type,
inputs, outputs, attrs,
graph_value_infos,
embed_params=embed_params,
)
fluid_writer.emit_op(
fluid_program,
name,
domain,
op_type,
inputs,
outputs,
attrs,
graph_value_infos,
embed_params=embed_params,
)
except BaseException as e:
logger.fatal('conversion failed for:\n\t%s -> %s::%s -> %s',
inputs, domain, op_type, outputs)
logger.fatal('conversion failed for:\n\t%s -> %s::%s -> %s', inputs,
domain, op_type, outputs)
raise e
op_codes = paddle_program.codes
paddle_program.codes = []
logger.info('%d ops converted', len(paddle_program.op_descs))
op_codes = fluid_program.codes
fluid_program.codes = []
logger.info('%d ops converted', len(fluid_program.op_descs))
# weight writer
for name, weight in graph_weights(onnx_graph):
......@@ -138,18 +149,24 @@ def convert(onnx_model_filename, save_dir,
var_names = value_info.get('embeded_as', [])
if var_names:
if len(var_names) > 1:
logger.info('weight %s is shared between ops, more disk space will be consumed', name)
logger.debug('saving weight %s with size of %d, in %d bytes, as %s ...',
name, weight.size, weight.nbytes, var_names)
for var_name in var_names: # multiple references
paddle_writer.write_weight(weight, shutil.os.path.join(save_dir, var_name))
logger.info(
'weight %s is shared between ops, more disk space will be consumed',
name)
logger.debug(
'saving weight %s with size of %d, in %d bytes, as %s ...',
name, weight.size, weight.nbytes, var_names)
for var_name in var_names: # multiple references
fluid_writer.write_weight(
weight, shutil.os.path.join(save_dir, var_name))
else:
logger.debug('saving weight %s with size of %d, in %d bytes, to %s ...',
name, weight.size, weight.nbytes, make_var_name(name))
paddle_writer.write_weight(weight, shutil.os.path.join(save_dir, make_var_name(name)))
paddle_writer.emit_param(paddle_program, name, value_info)
param_codes = paddle_program.codes
paddle_program.codes = []
logger.debug(
'saving weight %s with size of %d, in %d bytes, to %s ...',
name, weight.size, weight.nbytes, make_var_name(name))
fluid_writer.write_weight(
weight, shutil.os.path.join(save_dir, make_var_name(name)))
fluid_writer.emit_param(fluid_program, name, value_info)
param_codes = fluid_program.codes
fluid_program.codes = []
logger.info('%d weights converted', len(graph_params))
# input writer
......@@ -159,9 +176,11 @@ def convert(onnx_model_filename, save_dir,
value_info = graph_value_infos[name]
assert value_info['external']
external_inputs.append(name)
paddle_writer.emit_inputs(paddle_program, external_inputs, graph_value_infos, remove_batch=False) # TODO:
input_codes = paddle_program.codes
paddle_program.codes = []
fluid_writer.emit_inputs(
fluid_program, external_inputs, graph_value_infos,
remove_batch=False) # TODO:
input_codes = fluid_program.codes
fluid_program.codes = []
logger.info('%d inputs converted', len(external_inputs))
# output writer
......@@ -171,49 +190,93 @@ def convert(onnx_model_filename, save_dir,
value_info = graph_value_infos[name]
assert value_info['external']
external_outputs.append(name)
paddle_writer.emit_outputs(paddle_program, external_outputs)
output_codes = [''] + paddle_program.codes # add an empty line
paddle_program.codes = []
fluid_writer.emit_outputs(fluid_program, external_outputs)
output_codes = [''] + fluid_program.codes # add an empty line
fluid_program.codes = []
logger.info('%d outputs converted', len(external_outputs))
# code generation
header_codes = fluid_writer.header_code(
model_func_name, 'From: {}'.format(onnx_model_filename))
code_filename = shutil.os.path.join(save_dir, model_basename)
paddle_writer.write_code_file(code_filename, paddle_writer.header_code(model_func_name),
input_codes, param_codes, op_codes, output_codes)
logger.info('code saved to %s, factory function: %s', code_filename, model_func_name)
fluid_writer.write_code_file(code_filename, header_codes, input_codes,
param_codes, op_codes, output_codes)
logger.info('code saved to %s, factory function: %s', code_filename,
model_func_name)
# desc generation
desc_filename = shutil.os.path.join(save_dir, '__model__')
paddle_writer.write_desc_file(desc_filename,
op_descs=paddle_program.op_descs,
var_descs=paddle_program.var_descs,
)
fluid_writer.write_desc_file(
desc_filename,
op_descs=fluid_program.op_descs,
var_descs=fluid_program.var_descs,
)
logger.info('program saved to %s', desc_filename)
logger.info('conversion finished')
# globals().update(locals())
# globals().update(locals())
if __name__ == '__main__':
logging.basicConfig(
format='[%(levelname)8s]%(name)s::%(funcName)s:%(lineno)04d: %(message)s',
level=logging.DEBUG,
)
import argparse
parser = argparse.ArgumentParser(
description='onnx2fluid.convert',
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
parser.add_argument(
'model',
nargs=1,
help='path to model.onnx',
)
parser.add_argument(
'--debug',
'-d',
action='store_true',
help='enable debug logging and checking',
)
parser.add_argument(
'--output_dir',
'-o',
type=str,
default='',
help='output directory',
)
parser.add_argument(
'--embed_params',
'-e',
action='store_true',
help='try to embed parameters for trainable Paddle fluid layers',
)
parser.add_argument(
'--pedantic',
action='store_true',
default=True,
help='accept and convert only standard ONNX opset',
)
parser.add_argument(
'--no-pedantic',
'-x',
action='store_false',
dest='pedantic',
help='process non-standard ONNX ops, this may lead to fails',
)
args = parser.parse_args()
logging_format = '[%(levelname)8s]%(name)s::%(funcName)s:%(lineno)04d: %(message)s'
logging_level = logging.DEBUG if args.debug else logging.INFO
logging.basicConfig(format=logging_format, level=logging_level)
debug = args.debug
model_filename = args.model[0]
save_dir = args.output_dir
embed_params = args.embed_params
pedantic = args.pedantic
model_list = [
'../examples/t1.onnx',
'../examples/t2.onnx',
'../examples/t3.onnx',
'../examples/t4.onnx',
'../examples/t5.onnx',
'../examples/t6.onnx',
# '../examples/t7.onnx',
# '../examples/t8.onnx',
]
for model in model_list:
pathname, _ = shutil.os.path.splitext(model)
convert(model, pathname,
onnx_opset_pedantic=False, debug=True)
convert(model, pathname + '.embeded',
embed_params=True, onnx_opset_pedantic=False, debug=True)
convert(
model_filename,
save_dir,
embed_params=embed_params,
onnx_opset_pedantic=pedantic,
debug=debug)
onnx2fluid/onnx2fluid/framework_pb2.py 0 → 100644
浏览文件 @ 2228423e
此差异已折叠。
onnx2paddle/onnx2paddle/onnx_utils.py → onnx2fluid/onnx2fluid/onnx_utils.py
浏览文件 @ 2228423e
......@@ -12,34 +12,36 @@ import logging
import numpy as np
import onnx
from collections import OrderedDict as Dict # as default dict
from collections import OrderedDict as Dict # as default dict
from onnx.helper import get_attribute_value, make_attribute
from onnx.mapping import TENSOR_TYPE_TO_NP_TYPE
from onnx.numpy_helper import to_array
from onnx.shape_inference import infer_shapes
logger = logging.getLogger(__name__)
__all__ = [
'print_pb_structure',
'build_value_refs',
'node_attrs', 'node_topo', 'node_iter',
'node_attrs',
'node_topo',
'node_iter',
'tensor_shape',
'graph_ops', 'graph_weights',
'graph_ops',
'graph_weights',
'inferred_model_value_info',
'optimize_model_skip_op_for_inference',
'optimize_model_strip_initializer',
'optimize_model_cast', 'optimize_model_slice',
'optimize_model_cast',
'optimize_model_slice',
]
ONNX_INT_MAX = 2 ** 63 - 1
ONNX_INT_MAX = 2**63 - 1
DEFAULT_OP_DOMAIN = 'ai.onnx'
def print_pb_structure(message,
loop_iterative=False, depth=0):
def print_pb_structure(message, loop_iterative=False, depth=0):
"""
print pb fields in its structure
"""
......@@ -47,14 +49,17 @@ def print_pb_structure(message,
if hasattr(message, 'DESCRIPTOR') and hasattr(message.DESCRIPTOR, 'fields'):
for field in message.DESCRIPTOR.fields:
print('\t' * depth + '-', field.name)
print_pb_structure(getattr(message, field.name),
loop_iterative=loop_iterative, depth=(depth + 1))
print_pb_structure(
getattr(message, field.name),
loop_iterative=loop_iterative,
depth=(depth + 1))
if loop_iterative and hasattr(message, 'MergeFrom') and hasattr(message, '__len__'):
if loop_iterative and hasattr(message, 'MergeFrom') and hasattr(
message, '__len__'):
for idx, item in enumerate(message):
print('\t' * depth + '-', idx)
print_pb_structure(item,
loop_iterative=loop_iterative, depth=(depth + 1))
print_pb_structure(
item, loop_iterative=loop_iterative, depth=(depth + 1))
def build_value_refs(nodes):
......@@ -80,7 +85,8 @@ def get_attribute_value2(attr):
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))
value = np.frombuffer(
data, dtype=dtype, count=(len(data) // dtype.itemsize))
else:
value = get_attribute_value(attr)
return value
......@@ -91,7 +97,8 @@ def node_attrs(node):
convert ONNX node attributes to dict
"""
return {attr.name: get_attribute_value2(attr) for attr in node.attribute} # dict
return {attr.name: get_attribute_value2(attr)
for attr in node.attribute} # dict
def tensor_shape(tensor):
......@@ -137,7 +144,7 @@ def node_topo(nodes, topo='default'):
for next_idx in input_refs[val_name]:
node_in_degrees[next_idx] -= 1
if node_in_degrees[next_idx] == 0:
queue.insert(0, next_idx) # make it lazy
queue.insert(0, next_idx) # make it lazy
return node_topo
if topo == 'backward':
......@@ -162,14 +169,13 @@ def node_topo(nodes, topo='default'):
for next_idx in output_refs[val_name]:
node_out_degrees[next_idx] -= 1
if node_out_degrees[next_idx] == 0:
queue.insert(0, next_idx) # make it lazy
queue.insert(0, next_idx) # make it lazy
return node_topo
raise ValueError('unkown given topo: {}'.format(topo))
def node_iter(nodes,
indices=None):
def node_iter(nodes, indices=None):
"""
generator for ONNX node graph with given indices
"""
......@@ -194,8 +200,7 @@ def node_iter(nodes,
yield name, domain, op_type, inputs, outputs, attrs
def graph_ops(graph,
topo='default'):
def graph_ops(graph, topo='default'):
"""
generator for ONNX node graph with given topology
"""
......@@ -232,24 +237,24 @@ def inferred_model_value_info(model):
value_info = Dict()
for item in graph.value_info:
value_info[item.name] = dict(
dtype=TENSOR_TYPE_TO_NP_TYPE[item.type.tensor_type.elem_type],
shape=tensor_shape(item),
external=False,
)
dtype=TENSOR_TYPE_TO_NP_TYPE[item.type.tensor_type.elem_type],
shape=tensor_shape(item),
external=False,
)
for item in graph.input:
assert item.name not in value_info
value_info[item.name] = dict(
dtype=TENSOR_TYPE_TO_NP_TYPE[item.type.tensor_type.elem_type],
shape=tensor_shape(item),
external=True,
)
dtype=TENSOR_TYPE_TO_NP_TYPE[item.type.tensor_type.elem_type],
shape=tensor_shape(item),
external=True,
)
for item in graph.output:
# assert item.name not in value_info, 'bypass-model not supported'
# assert item.name not in value_info, 'bypass-model not supported'
value_info[item.name] = dict(
dtype=TENSOR_TYPE_TO_NP_TYPE[item.type.tensor_type.elem_type],
shape=tensor_shape(item),
external=True,
)
dtype=TENSOR_TYPE_TO_NP_TYPE[item.type.tensor_type.elem_type],
shape=tensor_shape(item),
external=True,
)
return value_info
......@@ -283,9 +288,7 @@ def skip_node_backward(nodes, src_input_name, dst_output_name, output_refs):
return processed
def optimize_model_skip_op_for_inference(
model,
op_list=None):
def optimize_model_skip_op_for_inference(model, op_list=None):
"""
skip ops can be bypassed for inference
"""
......@@ -297,38 +300,42 @@ def optimize_model_skip_op_for_inference(
ret = type(model)()
ret.CopyFrom(model)
ret.graph.ClearField('value_info') # WORKAROUND: onnx do not drop old value_info
ret.graph.ClearField(
'value_info') # WORKAROUND: onnx do not drop old value_info
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 == ''):
if not (node.domain == DEFAULT_OP_DOMAIN or node.domain == ''):
continue
op_type = node.op_type
if not(op_type in op_list):
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):
logger.warning('currently only 1-input-1-output op supported, skip required %d: %s',
node_idx, node.op_type)
elif not (len(node.input) == 1 and len(node.output) == 1):
logger.warning(
'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 = skip_node_forward(ret_nodes, output_name, input_name, input_refs)
processed = skip_node_forward(ret_nodes, output_name, input_name,
input_refs)
elif input_name in output_refs:
processed = skip_node_backward(ret_nodes, input_name, output_name, output_refs)
processed = skip_node_backward(ret_nodes, input_name, output_name,
output_refs)
else:
processed = -1
if processed > 0:
nodes_to_remove.append(node_idx)
logger.debug('skip op %d: %s -> %s -> %s',
node_idx, input_name, node.op_type, output_name)
logger.debug('skip op %d: %s -> %s -> %s', node_idx, input_name,
node.op_type, output_name)
elif processed == 0:
logger.warning('weird, no node processed')
else:
......@@ -342,8 +349,7 @@ def optimize_model_skip_op_for_inference(
return ret
def optimize_model_strip_initializer(model,
keep_input_only=True):
def optimize_model_strip_initializer(model, keep_input_only=True):
"""
strip weights for inference
"""
......@@ -354,7 +360,8 @@ def optimize_model_strip_initializer(model,
ret = type(model)()
ret.CopyFrom(model)
ret.graph.ClearField('value_info') # WORKAROUND: onnx do not drop old value_info
ret.graph.ClearField(
'value_info') # WORKAROUND: onnx do not drop old value_info
# strip initializers
ret.graph.ClearField('initializer')
......@@ -366,8 +373,7 @@ def optimize_model_strip_initializer(model,
elif not keep_input_only and name in output_refs:
ret_initializers.add().CopyFrom(initializer)
else:
logger.debug('initializer %s(%s[%d]) stripped',
name,
logger.debug('initializer %s(%s[%d]) stripped', name,
TENSOR_TYPE_TO_NP_TYPE[initializer.data_type],
len(initializer.raw_data))
......@@ -379,10 +385,10 @@ def optimize_model_strip_initializer(model,
if name in input_refs or name in out_names:
ret_inputs.add().CopyFrom(item)
else:
logger.debug('input %s(%s%s) stripped',
name,
TENSOR_TYPE_TO_NP_TYPE[item.type.tensor_type.elem_type],
tensor_shape(item))
logger.debug(
'input %s(%s%s) stripped', name,
TENSOR_TYPE_TO_NP_TYPE[item.type.tensor_type.elem_type],
tensor_shape(item))
return ret
......@@ -397,18 +403,19 @@ def optimize_model_cast(model):
ret = type(model)()
ret.CopyFrom(model)
ret.graph.ClearField('value_info') # WORKAROUND: onnx do not drop old value_info
ret.graph.ClearField(
'value_info') # WORKAROUND: onnx do not drop old value_info
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 == ''):
if not (node.domain == DEFAULT_OP_DOMAIN or node.domain == ''):
continue
if not(node.op_type == 'Cast'):
if not (node.op_type == 'Cast'):
continue
attrs = node_attrs(node)
output_dtype = TENSOR_TYPE_TO_NP_TYPE[attrs['to']]
input_name = node.input[0]
info = value_info.get('input_name', None) # relax for un-inferrable
info = value_info.get('input_name', None) # relax for un-inferrable
if info is None:
continue
input_dtype = info.get('dtype', None)
......@@ -417,21 +424,23 @@ def optimize_model_cast(model):
output_name = node.output[0]
if output_name in input_refs:
processed = skip_node_forward(ret_nodes, output_name, input_name, input_refs)
processed = skip_node_forward(ret_nodes, output_name, input_name,
input_refs)
elif input_name in output_refs:
processed = skip_node_backward(ret_nodes, input_name, output_name, output_refs)
processed = skip_node_backward(ret_nodes, input_name, output_name,
output_refs)
else:
processed = -1
if processed > 0:
nodes_to_remove.append(node_idx)
logger.debug('skip %s: %s -> %s Cast op',
node.name, input_dtype, output_dtype)
logger.debug('skip %s: %s -> %s Cast op', node.name, input_dtype,
output_dtype)
elif processed == 0:
logger.warning('weird, no node processed')
else:
logger.debug('keep standalone %s: %s -> %s Cast op',
node.name, input_dtype, output_dtype)
logger.debug('keep standalone %s: %s -> %s Cast op', node.name,
input_dtype, output_dtype)
nodes_to_remove.sort(reverse=True)
for node_idx in nodes_to_remove:
......@@ -452,13 +461,14 @@ def optimize_model_slice(model):
chain = []
while True:
node = nodes[node_idx]
if not(node.domain == DEFAULT_OP_DOMAIN or node.domain == ''):
if not (node.domain == DEFAULT_OP_DOMAIN or node.domain == ''):
return chain
if not node.op_type == 'Slice':
return chain
chain.append(node_idx)
output_name = node.output[0]
if output_name not in input_refs or len(input_refs[output_name]) != 1:
if output_name not in input_refs or len(
input_refs[output_name]) != 1:
return chain
node_idx = list(input_refs[output_name])[0]
......@@ -468,7 +478,8 @@ def optimize_model_slice(model):
for slice_node_idx in slice_chain:
node = nodes[slice_node_idx]
attrs = node_attrs(node)
for axis, start, end in zip(attrs['axes'], attrs['starts'], attrs['ends']):
for axis, start, end in zip(attrs['axes'], attrs['starts'],
attrs['ends']):
if start == 0 and end == ONNX_INT_MAX:
continue
if axis in merged_slice:
......@@ -480,7 +491,8 @@ def optimize_model_slice(model):
ret = type(model)()
ret.CopyFrom(model)
ret.graph.ClearField('value_info') # WORKAROUND: onnx do not drop old value_info
ret.graph.ClearField(
'value_info') # WORKAROUND: onnx do not drop old value_info
ret_nodes = ret.graph.node
nodes_to_remove = []
for node_idx in range(len(nodes)):
......@@ -488,7 +500,7 @@ def optimize_model_slice(model):
if len(slice_chain) == 0:
continue
merged_slice = _merge_slice(slice_chain)
if len(merged_slice) > 0 and len(slice_chain) == 1: # no need to merge
if len(merged_slice) > 0 and len(slice_chain) == 1: # no need to merge
continue
attrs = dict(axes=[], starts=[], ends=[])
......@@ -501,42 +513,50 @@ def optimize_model_slice(model):
input_name = first_node.input[0]
output_name = last_node.output[0]
processed = -1
if output_name in input_refs: # 0, [1...]
new_input_name = first_node.output[0] if len(merged_slice) > 0 else input_name
processed = skip_node_forward(ret_nodes, output_name, new_input_name, input_refs)
if output_name in input_refs: # 0, [1...]
new_input_name = first_node.output[0] if len(
merged_slice) > 0 else input_name
processed = skip_node_forward(ret_nodes, output_name,
new_input_name, input_refs)
if processed > 0:
if len(merged_slice) > 0:
remain_idx = slice_chain[0]
remove_chain = slice_chain[1:]
slice_node = ret_nodes[remain_idx]
for attr in slice_node.attribute:
attr.CopyFrom(make_attribute(attr.name, attrs[attr.name]))
attr.CopyFrom(
make_attribute(attr.name, attrs[attr.name]))
logger.debug('merged slice chain %s -> %s%s -> %s',
input_name, remain_idx, remove_chain, output_name)
input_name, remain_idx, remove_chain,
output_name)
else:
remove_chain = slice_chain
if processed < 0 and input_name in output_refs:
new_output_name = last_node.input[0] if len(merged_slice) > 0 else output_name
processed = skip_node_backward(ret_nodes, input_name, new_output_name, output_refs)
new_output_name = last_node.input[0] if len(
merged_slice) > 0 else output_name
processed = skip_node_backward(ret_nodes, input_name,
new_output_name, output_refs)
if processed > 0:
if len(merged_slice) > 0:
remain_idx = slice_chain[-1]
remove_chain = slice_chain[:-1]
slice_node = ret_nodes[remain_idx]
for attr in slice_node.attribute:
attr.CopyFrom(make_attribute(attr.name, attrs[attr.name]))
attr.CopyFrom(
make_attribute(attr.name, attrs[attr.name]))
logger.debug('merged slice chain %s -> %s%s -> %s',
input_name, remove_chain, remain_idx, output_name)
input_name, remove_chain, remain_idx,
output_name)
else:
remove_chain = slice_chain
if processed > 0:
nodes_to_remove.extend(remove_chain)
if len(merged_slice) == 0:
logger.debug('skip slice chain %s -> %s -> %s',
input_name, slice_chain, output_name)
elif processed < 0: # NEVERFIX: not merge standalone slice chain
logger.debug('skip slice chain %s -> %s -> %s', input_name,
slice_chain, output_name)
elif processed < 0: # NEVERFIX: not merge standalone slice chain
logger.debug('keep standalone slice chain %s -> %s -> %s',
input_name, slice_chain, output_name)
......@@ -549,9 +569,10 @@ def optimize_model_slice(model):
if __name__ == '__main__':
logging.basicConfig(
format='[%(levelname)8s]%(name)s::%(funcName)s:%(lineno)04d: %(message)s',
level=logging.DEBUG,
)
format=
'[%(levelname)8s]%(name)s::%(funcName)s:%(lineno)04d: %(message)s',
level=logging.DEBUG,
)
from onnx.checker import check_model
from onnx.utils import polish_model
......
onnx2paddle/onnx2paddle/torch_export_helper.py → onnx2fluid/onnx2fluid/torch_export_helper.py
浏览文件 @ 2228423e
......@@ -24,8 +24,7 @@ def _ensure_tuple(obj):
return (obj, )
def _flatten_list(obj,
out=None):
def _flatten_list(obj, out=None):
assert isinstance(obj, list)
if out is None:
out = type(obj)()
......@@ -37,8 +36,7 @@ def _flatten_list(obj,
return out
def export_data(state_dict,
prefix=''):
def export_data(state_dict, prefix=''):
"""
export binary data with meta text for raw C++ inference engines
"""
......@@ -65,10 +63,14 @@ def export_data(state_dict,
fp.close()
def export_onnx_with_validation(model, inputs, export_basepath,
input_names=None, output_names=None,
def export_onnx_with_validation(model,
inputs,
export_basepath,
input_names=None,
output_names=None,
use_npz=True,
*args, **kwargs):
*args,
**kwargs):
"""
export PyTorch model to ONNX model and export sample inputs and outputs in a Numpy file
"""
......@@ -95,12 +97,16 @@ def export_onnx_with_validation(model, inputs, export_basepath,
ret[key] = value
return ret
torch_inputs = _ensure_tuple(inputs) # WORKAROUND: for torch.onnx
outputs = torch.onnx.export(model, torch_inputs, export_basepath + '.onnx',
input_names=_flatten_list(input_names),
output_names=_flatten_list(output_names),
*args, **kwargs)
if outputs is None: # WORKAROUND: for torch.onnx
torch_inputs = _ensure_tuple(inputs) # WORKAROUND: for torch.onnx
outputs = torch.onnx.export(
model,
torch_inputs,
export_basepath + '.onnx',
input_names=_flatten_list(input_names),
output_names=_flatten_list(output_names),
*args,
**kwargs)
if outputs is None: # WORKAROUND: for torch.onnx
outputs = model(*inputs)
torch_outputs = _ensure_tuple(outputs)
......
onnx2paddle/onnx2paddle/validation.py → onnx2fluid/onnx2fluid/validation.py
浏览文件 @ 2228423e
......@@ -13,8 +13,7 @@ import os
import sys
def _flatten_dict(obj,
out=None):
def _flatten_dict(obj, out=None):
assert isinstance(obj, dict)
if out is None:
out = type(obj)()
......@@ -34,12 +33,13 @@ def _ensure_list(obj):
return [obj]
def validate(paddle_model_filename, golden_data_filename,
def validate(fluid_model_filename,
golden_data_filename,
model_func_name='inference',
precision=1e-4,
save_inference_model=False):
"""
inferece the converted Paddle model, validate with given golden data
inferece the converted Paddle fluid model, validate with given golden data
"""
import numpy as np
......@@ -52,17 +52,17 @@ def validate(paddle_model_filename, golden_data_filename,
exe.run(fluid.default_startup_program())
# load model
paddle_model_dir, basename = os.path.split(paddle_model_filename)
if basename == '__model__': # is desc model
fluid_model_dir, basename = os.path.split(fluid_model_filename)
if basename == '__model__': # is desc model
logger.debug('using desc file %s', basename)
prog, in_names, var_outs = fluid.io.load_inference_model(paddle_model_dir, exe)
out_names = var_outs # HINT: pass var if fetch ops already created
prog, _, var_outs = fluid.io.load_inference_model(fluid_model_dir, exe)
out_names = var_outs # HINT: pass var if fetch ops already created
logger.info('model load passed')
elif basename.endswith('.py'): # is python code
elif basename.endswith('.py'): # is python code
logger.debug('using python code file %s', basename)
module_name, _ = os.path.splitext(basename)
sys_path = sys.path.copy()
sys.path.append(paddle_model_dir)
sys.path.append(fluid_model_dir)
try:
module = importlib.import_module(module_name)
func = getattr(module, model_func_name)
......@@ -71,18 +71,21 @@ def validate(paddle_model_filename, golden_data_filename,
module = importlib.import_module(module_name)
func = getattr(module, model_func_name)
sys.path = sys_path
logger.debug('from %s imported %s: %s', module_name, model_func_name, func)
logger.debug('from %s imported %s: %s', module_name, model_func_name,
func)
var_outs = func()
var_outs = _ensure_list(var_outs)
out_names = [var.name for var in var_outs] # HINT: pass string to create fetch ops
out_names = [var.name for var in var_outs
] # HINT: pass string to create fetch ops
logger.info('import passed')
prog = fluid.default_main_program()
fluid.io.load_persistables(executor=exe, dirname=paddle_model_dir, main_program=prog)
fluid.io.load_persistables(
executor=exe, dirname=fluid_model_dir, main_program=prog)
logger.info('weight load passed')
else:
raise ValueError('unsupported Paddle model')
raise ValueError('unsupported Paddle fluid model')
# load data
logger.info('using golden data %s', golden_data_filename)
......@@ -100,10 +103,15 @@ def validate(paddle_model_filename, golden_data_filename,
# DEBUG: reload test for python code
if basename.endswith('.py') and save_inference_model:
fluid.io.save_inference_model(paddle_model_dir, input_data.keys(), var_outs, exe,
main_program=prog, export_for_deployment=True)
fluid.io.save_inference_model(
fluid_model_dir,
input_data.keys(),
var_outs,
exe,
main_program=prog,
export_for_deployment=True)
logger.info('model re-save passed')
fluid.io.load_inference_model(paddle_model_dir, exe)
fluid.io.load_inference_model(fluid_model_dir, exe)
logger.info('model re-load passed')
# execute
......@@ -124,49 +132,54 @@ def validate(paddle_model_filename, golden_data_filename,
else:
logger.info('accuracy not passed')
# globals().update(locals())
return passed
if __name__ == '__main__':
logging.basicConfig(
format='[%(levelname)8s]%(name)s::%(funcName)s:%(lineno)04d: %(message)s',
level=logging.DEBUG,
)
logger = logging.getLogger('validation_test')
model_rc_list = [
'../examples/t{}/model.py',
'../examples/t{}/__model__',
'../examples/t{}.embeded/model.py',
'../examples/t{}.embeded/__model__',
]
import numpy as np
idx_model = np.random.randint(1, 7)
model = np.random.choice(model_rc_list).format(idx_model)
precision = 10 ** (np.random.rand() * -4 - 2)
debug = False
model = '/tmp/export/model.py'
# model = '../examples/t1/__model__'
# model = '../examples/t1.embeded/model.py'
# model = '../examples/t1.embeded/__model__'
debug = True
logger.info('args: %s %.6f', model, precision)
data_dir, dir_name = os.path.split(os.path.split(model)[0])
data_pathname = os.path.splitext(dir_name)[0]
# proto debug test
from framework_pb2 import ProgramDesc
pd = ProgramDesc()
pd.ParseFromString(open(os.path.join(data_dir, dir_name, '__model__'), 'rb').read())
# validate
# validate(model, os.path.join(data_dir, data_pathname + '.npz'),
# precision=precision, save_inference_model=debug)
validate(model, '../examples/bvlc_alexnet/test_data_0.npz',
precision=precision, save_inference_model=debug)
import argparse
parser = argparse.ArgumentParser(
description='onnx2fluid.validate',
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
parser.add_argument(
'model',
nargs=1,
help='path to model.py or __model__',
)
parser.add_argument(
'--debug',
'-d',
action='store_true',
help='enable debug logging and checking',
)
parser.add_argument(
'--test_data',
'-t',
type=str,
help='I/O golden data for validation, e.g. test.npy, test.npz',
)
parser.add_argument(
'--precision',
'-p',
type=int,
default=4,
help='assertion decimal for validation',
)
args = parser.parse_args()
logging_format = '[%(levelname)8s]%(name)s::%(funcName)s:%(lineno)04d: %(message)s'
logging_level = logging.DEBUG if args.debug else logging.INFO
logging.basicConfig(format=logging_format, level=logging_level)
debug = args.debug
fluid_model_filename = args.model[0]
golden_data_filename = args.test_data
precision = args.precision
validate(
fluid_model_filename,
golden_data_filename,
precision=precision,
save_inference_model=debug)
onnx2paddle/onnx2paddle/writer.py → onnx2fluid/onnx2fluid/writer.py
浏览文件 @ 2228423e
......@@ -34,15 +34,13 @@ except ImportError:
logger.warning('importing paddle.fluid.proto.framework_pb2d failed,'
'using fallback framework_pb2')
__all__ = [
'Program',
'Writer',
'Program',
'Writer',
]
def _irepr(obj,
to='_'):
def _irepr(obj, to='_'):
"""inline repr"""
s = repr(obj)
......@@ -53,8 +51,7 @@ def _irepr(obj,
return s
def _flatten_list(obj,
out=None):
def _flatten_list(obj, out=None):
if out is None:
out = type(obj)()
for item in obj:
......@@ -72,7 +69,7 @@ def make_attr_name(name):
if name == '':
raise ValueError('name should not be empty')
for s in ' *?\/-:': #
for s in ' *?\\/-:': #
name = name.replace(s, '_')
if not name.startswith('_'):
name = '_' + name
......@@ -85,15 +82,15 @@ class Program(object):
"""
DTYPE_TO_FRAMEWORK_DTYPE = {
'bool': framework_pb2.VarType.BOOL,
'int8': framework_pb2.VarType.INT8,
'uint8': framework_pb2.VarType.UINT8,
'int16': framework_pb2.VarType.INT16,
'int32': framework_pb2.VarType.INT32,
'int64': framework_pb2.VarType.INT64,
'float16': framework_pb2.VarType.FP16,
'float32': framework_pb2.VarType.FP32,
'float64': framework_pb2.VarType.FP64
'bool': framework_pb2.VarType.BOOL,
'int8': framework_pb2.VarType.INT8,
'uint8': framework_pb2.VarType.UINT8,
'int16': framework_pb2.VarType.INT16,
'int32': framework_pb2.VarType.INT32,
'int64': framework_pb2.VarType.INT64,
'float16': framework_pb2.VarType.FP16,
'float32': framework_pb2.VarType.FP32,
'float64': framework_pb2.VarType.FP64
}
@staticmethod
......@@ -116,7 +113,7 @@ class Program(object):
od_var = framework_pb2.OpDesc.Var()
od_var.parameter = key
if idx < len(vals):
od_var.arguments.append(vals[idx]) #
od_var.arguments.append(vals[idx]) #
od_vars.append(od_var)
return od_vars
......@@ -130,10 +127,10 @@ class Program(object):
for key, value in attrs.items():
od_attr = framework_pb2.OpDesc.Attr()
od_attr.name = key
if isinstance(value, bool): # bool.mro() = [bool, int, object]
if isinstance(value, bool): # bool.mro() = [bool, int, object]
od_attr.type = framework_pb2.BOOLEAN
od_attr.b = value
elif isinstance(value, int): # only cast to int32
elif isinstance(value, int): # only cast to int32
od_attr.type = framework_pb2.INT
od_attr.i = value
elif isinstance(value, float):
......@@ -143,10 +140,10 @@ class Program(object):
od_attr.type = framework_pb2.STRING
od_attr.s = value
elif isinstance(value, list) and len(value) > 0:
if isinstance(value, bool): # bool.mro() = [bool, int, object]
if isinstance(value, bool): # bool.mro() = [bool, int, object]
od_attr.type = framework_pb2.BOOLEANS
od_attr.bools.extend(value)
elif isinstance(value[0], int): # only cast to int32 list
elif isinstance(value[0], int): # only cast to int32 list
od_attr.type = framework_pb2.INTS
od_attr.ints.extend(value)
elif isinstance(value[0], float):
......@@ -168,11 +165,8 @@ class Program(object):
return ('Program(code mutable: {}) with:\n'
'codes: {}\n'
'op_descs: {}\n'
'var_descs: {}\n').format(
self.code_mutable,
self.codes,
self.op_descs,
self.var_descs)
'var_descs: {}\n').format(self.code_mutable, self.codes,
self.op_descs, self.var_descs)
def __repr__(self):
return self.__str__()
......@@ -185,8 +179,11 @@ class Program(object):
if self.code_mutable:
self.codes.append(code)
def OpDesc(self, name,
input_val_keys=None, output_val_keys=None, attrs=None):
def OpDesc(self,
name,
input_val_keys=None,
output_val_keys=None,
attrs=None):
"""
add OpDesc
"""
......@@ -202,10 +199,15 @@ class Program(object):
self.op_descs.append(desc)
return desc
def VarDesc(self, name,
persistable=False, value_info=None, remove_batch=None):
def VarDesc(self,
name,
persistable=False,
value_info=None,
remove_batch=None,
dummy_dtype='float32'):
"""
add VarDesc
add VarDesc,
dummy_dtype: WORKAROUND for Netron viewer
"""
var_desc = framework_pb2.VarDesc()
......@@ -213,14 +215,19 @@ class Program(object):
var_desc.persistable = persistable
var_desc.type.type = framework_pb2.VarType.LOD_TENSOR
# REMOVEIT: WORKAROUND: Netron: null.tensor error
tensor_desc = var_desc.type.lod_tensor.tensor
tensor_desc.data_type = self.Dtype(dummy_dtype) # required
if value_info and 'dtype' in value_info:
tensor_desc = var_desc.type.lod_tensor.tensor
tensor_desc.data_type = self.Dtype(value_info['dtype']) # required
tensor_desc.data_type = self.Dtype(value_info['dtype']) # required
if 'shape' in value_info:
tensor_desc.dims.extend(value_info['shape'])
if len(value_info['shape']) > 0: # skip scalars
if len(value_info['shape']) > 0: # skip scalars
if remove_batch is None:
remove_batch = value_info.get('remove_batch', not persistable)
remove_batch = value_info.get('remove_batch',
not persistable)
if remove_batch:
tensor_desc.dims[0] = -1
......@@ -231,7 +238,7 @@ class Program(object):
convert an ONNX op and add it to program
"""
if domain != '': # TODO: symbolic file routing by domain
if domain != '': # TODO: symbolic file routing by domain
raise ValueError('only default domain supported')
if op_type in symbolic.DEFAULT_OP_MAPPING:
......@@ -240,8 +247,8 @@ class Program(object):
fn = getattr(symbolic, op_type)
fn(self, *args, **kwargs)
else:
raise ValueError('conversion for {}::{} not supported'
.format(domain, op_type))
raise ValueError('conversion for {}::{} not supported'.format(
domain, op_type))
def IntermediateOp(self, domain, op_type, *args, **kwargs):
"""
......@@ -267,14 +274,15 @@ class Writer(object):
CODE_INDENT = ' ' * 4
@staticmethod
def header_code(func_name):
def header_code(func_name, info=''):
"""
Python header codes
"""
codes = list()
codes.append('"""')
codes.append('This code is generated by onnx2paddle.')
codes.append('This code is generated by onnx2fluid.')
codes.append('{}'.format(info))
codes.append('"""')
codes.append('')
codes.append('from __future__ import division')
......@@ -287,16 +295,25 @@ class Writer(object):
return codes
@staticmethod
def emit_op(prog, name, domain, op_type, inputs, outputs, attrs, value_infos, *args, **kwargs):
def emit_op(prog, name, domain, op_type, inputs, outputs, attrs,
value_infos, *args, **kwargs):
"""
emit an ONNX op into program
"""
prog.Code('# {}, {}::{}: {} -> {}, {}'
.format(name, domain, op_type, inputs, outputs, _irepr(attrs, to=', ')))
prog.Op(domain, op_type, inputs, outputs, attrs,
value_infos=value_infos, name=name,
*args, **kwargs)
prog.Code('# {}, {}::{}: {} -> {}, {}'.format(name, domain, op_type,
inputs, outputs,
_irepr(attrs, to=', ')))
prog.Op(
domain,
op_type,
inputs,
outputs,
attrs,
value_infos=value_infos,
name=name,
*args,
**kwargs)
@staticmethod
def emit_param(prog, name, value_info):
......@@ -313,18 +330,18 @@ class Writer(object):
var_name = make_var_name(name)
attr_name = make_attr_name(name)
prog.Code('# parameter: {}'.format(name))
prog.Code('{} = ParamAttr(name={})' # , trainable=True
prog.Code('{} = ParamAttr(name={})' # , trainable=True
.format(attr_name, repr(var_name)))
prog.Code('{} = layers.create_parameter(shape={}, dtype={}, name={}, attr={}'
', default_initializer=initializer.Constant(0))' #, is_bias={}
.format(var_name,
value_info['shape'], repr(value_info['dtype'].name),
repr(name), attr_name)) #, value_info.get('is_bias', False)))
prog.Code(
'{} = layers.create_parameter(shape={}, dtype={}, name={}, attr={}'
', default_initializer=initializer.Constant(0))' #, is_bias={}
.format(var_name, value_info['shape'],
repr(value_info['dtype'].name), repr(name),
attr_name)) #, value_info.get('is_bias', False)))
prog.VarDesc(var_name, persistable=True, value_info=value_info)
@staticmethod
def emit_inputs(prog, names, value_infos,
remove_batch=None):
def emit_inputs(prog, names, value_infos, remove_batch=None):
"""
emit ONNX inputs into program
"""
......@@ -334,27 +351,33 @@ class Writer(object):
value_info = value_infos[name]
shape = value_info['shape']
if remove_batch is None:
remove_batch = value_info.get('remove_batch', True) # HINT: True by default ?
remove_batch = value_info.get('remove_batch',
True) # HINT: True by default ?
if remove_batch:
shape = shape[1:]
prog.Code('# input: {}'.format(name))
prog.Code(('{} = layers.data(name={}, shape={}, dtype={}, '
'append_batch_size={})' # , stop_gradient=True
).format(var_name, repr(name),
shape,
repr(value_info['dtype'].name),
remove_batch,
))
prog.OpDesc('feed',
(['feed'], 'X'),
([var_name], 'Out'),
dict(col=idx),
)
prog.VarDesc(var_name, value_info=value_info, remove_batch=remove_batch)
prog.Code((
'{} = layers.data(name={}, shape={}, dtype={}, '
'append_batch_size={})' # , stop_gradient=True
).format(
var_name,
repr(name),
shape,
repr(value_info['dtype'].name),
remove_batch,
))
prog.OpDesc(
'feed',
(['feed'], 'X'),
([var_name], 'Out'),
dict(col=idx),
)
prog.VarDesc(
var_name, value_info=value_info, remove_batch=remove_batch)
@staticmethod
def emit_outputs(prog, names): #, value_infos
def emit_outputs(prog, names): #, value_infos
"""
emit ONNX outputs into program
"""
......@@ -364,11 +387,12 @@ class Writer(object):
var_name = make_var_name(name)
code += var_name + ', '
prog.OpDesc('fetch',
([var_name], 'X'),
(['fetch'], 'Out'),
dict(col=idx),
)
prog.OpDesc(
'fetch',
([var_name], 'X'),
(['fetch'], 'Out'),
dict(col=idx),
)
# var is emitted over ops
prog.Code(code)
......@@ -396,9 +420,9 @@ class Writer(object):
tensor_desc.dims.extend(weight.shape)
fp = open(filename, 'wb')
np.array([0], dtype=np.int32).tofile(fp) # version
np.array([0], dtype=np.int64).tofile(fp) # LOD level
np.array([0], dtype=np.int32).tofile(fp) # tensor version
np.array([0], dtype=np.int32).tofile(fp) # version
np.array([0], dtype=np.int64).tofile(fp) # LOD level
np.array([0], dtype=np.int32).tofile(fp) # tensor version
np.array([tensor_desc.ByteSize()], dtype=np.int32).tofile(fp)
fp.write(tensor_desc.SerializeToString())
weight.tofile(fp)
......@@ -463,4 +487,4 @@ class Writer(object):
fp = open(filename, 'wb')
fp.write(prog_desc.SerializeToString())
fp.close()
logger.debug('saved descs to %s', filename)
\ No newline at end of file
logger.debug('saved descs to %s', filename)
onnx2paddle/requirements.txt → onnx2fluid/requirements.txt
浏览文件 @ 2228423e
-e .
onnx>=1.4.0
paddlepaddle
\ No newline at end of file
paddlepaddle
onnx2paddle/setup.cfg → onnx2fluid/setup.cfg
浏览文件 @ 2228423e
......@@ -2,14 +2,14 @@
# https://setuptools.readthedocs.io/en/latest/setuptools.html#configuring-setup-using-setup-cfg-files
[metadata]
# 项目名称，发布、安装时以此作为包名
name = onnx2paddle
name = onnx2fluid
# 作者姓名和邮箱地址
author = Macrobull
# author_email = .Github@github.com
# 项目版本号，1.0以上版本才视为正式版
version = 0.1.0
# 项目概要描述信息，一句话让用户明白项目概要，不支持中文
description = Inference model conversion from ONNX/PyTorch to Paddle
description = Inference model conversion from ONNX/PyTorch to Paddle fluid
# 项目的详细描述内容和格式，包括readme和changelog等，通常使用md或rst等格式
long_description = file: README.md, CHANGELOG.md
long_description_content_type = text/markdown
......@@ -25,7 +25,7 @@ classifier =
Programming Language :: Python :: 3.5
# 关键字，用于检索，方便用户搜索到你的项目
keywords =
onnx paddle
onnx paddlepaddle
[options]
# 包名称，find:表示自动寻找，可在options.packages.find中进行详细配置
......@@ -44,21 +44,21 @@ install_requires =
# mock
# 单测代码目录
#test_suite = onnx2paddle.tests
#test_suite = onnx2fluid.tests
# 自动添加被版本控制的数据文件
include_package_data = True
# 项目是纯py项目，可以直接执行zip源码包
zip_safe = False
# 可以通过以下配置将指定的函数变成命令行工具，允许用户直接执行
#[options.entry_points]
#console_scripts =
# onnx2paddle = onnx2paddle.cmdline:main
[options.entry_points]
console_scripts =
onnx2fluid = onnx2fluid.cmdline:main
# 可以通过以下配置向包中添加conf或data等非py文件，安装时会一同安装到site-packages目录下
# 仅支持文件，不支持目录，但可以使用通配
#[options.package_data]
#onnx2paddle =
#onnx2fluid =
# conf/*
# data/*
......
onnx2paddle/setup.py → onnx2fluid/setup.py
浏览文件 @ 2228423e
......@@ -15,4 +15,3 @@ Date: 2019/02/22 10:25:46
import setuptools
setuptools.setup()
onnx2paddle/onnx2paddle/framework_pb2.py 已删除 100644 → 0
浏览文件 @ a4796334
此差异已折叠。
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