rebase from develop and update readme

onnx2fluid/README.md

@@ -2,59 +2,82 @@
 
 [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
 
-onnx2fluid支持将onnx模型转换为PaddlePaddle模型，并用于预测，用户也可以通过将PyTorch模型导出为ONNX格式模型，再使用onnx2fluid将模型转为PaddlePaddle模型。
+onnx2fluid支持将ONNX模型转换为PaddlePaddle模型，并用于预测，用户也可以通过将PyTorch模型导出为ONNX模型，再使用onnx2fluid将模型转为PaddlePaddle模型。
 
-## 环境安装
+## 特色
 
-工具开发过程中，我们在如下环境配置中测试模型转换:
+* 导出Python代码和fluid ProgramDesc模型
+* 权重可嵌入支持的算子中
+* 转换验证打包三合一
+* 转换过程不依赖PaddlePaddle
+* 可自由扩展算子
+
+## 环境配置
+
+在如下环境配置中测试成功:
 
 * python 3.5+
 * onnx == 1.4.0
-* paddlepaddle == 1.3.0
-
-建议使用[anaconda](https://docs.anaconda.com/anaconda/install):
+* paddlepaddle == 1.3.0 (可选，仅用于验证)
 
+使用[Anaconda](https://docs.anaconda.com/anaconda/install):
 ``` shell
-# 安装onnx
-# 也可参考https://github.com/onnx/onnx
 conda install -c conda-forge onnx
+pip install paddlepaddle==1.3.0
 ```
 
-## 使用说明
-```shell
-# 安装
-git clone https://github.com/PaddlePaddle/X2Paddle.git
-cd X2Paddle/onnx2fluid
+## 动手玩
+
+测试ONNX官方预训练模型，包含alexnet, googlenet, caffenet, rcnn
+inception_v1, inception_v2, resnet50, shufflenet, squeezenet,
+vgg19, zfnet512等:
+
+``` shell
 python setup.py install
+cd examples
+sh onnx_model_zoo.sh
+```
 
-# 模型转换
-python -m onnx2fluid -o /path/to/export_dir/ /path/of/onnx/model.onnx
+使用PyTorch搭建模型，导出ONNX，转换并验证:
+
+``` shell
+python setup.py install
+cd examples
+python gen_some_samples.py
+onnx2fluid sample_1.onnx -t sample_1.npz
 ```
-**示例：VGG19模型**
+
+## 使用说明
+
+onnx2fluid:
+
 ```shell
-wget https://s3.amazonaws.com/download.onnx/models/opset_9/vgg19.tar.gz
-tar xzvf vgg19.tar.gz
+onnx2fluid [-dexy] [-o /path/to/export_dir/] [-z archive.zip] [-t test_data.npz] /path/to/onnx/model.onnx
 
-python -m onnx2fluid -o paddle_model vgg19/model.onnx
+optional arguments:
+  --debug, -d           启用调试
+  --embed_params, -e    尝试权重内嵌
+  --no-pedantic, -x     转换扩展的ONNX算子
+  --skip-version-conversion, -y
+                        跳过ONNX算子版本转换
+  --output_dir, -o      指定输出目录
+  --archive [ARCHIVE], -z [ARCHIVE]
+                        如果验证通过，打包到指定的ZIP文件
 ```
-转换后的PaddlePaddle模型加载可参考文档[加载预测模型](http://www.paddlepaddle.org/documentation/docs/zh/1.3/api_guides/low_level/inference.html#id4)
 
-## 模型测试
-目录[examples](examples)中集成了部分ONNX预训练模型的转换测试
+转换工具onnx2fluid.conversion:
+
 ```shell
-cd examples
-# 测试和验证各onnx模型的转换
-sh onnx_model_zoo.sh
+onnx2fluid.conversion [-dexy] [-o /path/to/export_dir/] /path/to/onnx/model.onnx
 ```
-目前测试脚本中已包含的测试模型如下，  
-1. [bvlc_alexnet](https://s3.amazonaws.com/download.onnx/models/opset_9/bvlc_alexnet.tar.gz)  
-2. [bvlc_googlenet](https://s3.amazonaws.com/download.onnx/models/opset_9/bvlc_googlenet.tar.gz)  
-3. [bvlc_reference_caffenet](https://s3.amazonaws.com/download.onnx/models/opset_9/bvlc_reference_caffenet.tar.gz)  
-4. [bvlc_reference_rcnn_ilsvrc13](https://s3.amazonaws.com/download.onnx/models/opset_9/bvlc_reference_rcnn_ilsvrc13.tar.gz)  
-5. [inception_v1](https://s3.amazonaws.com/download.onnx/models/opset_9/inception_v1.tar.gz)  
-6. [inception_v2](https://s3.amazonaws.com/download.onnx/models/opset_9/inception_v2.tar.gz)  
-7. [resnet50](https://s3.amazonaws.com/download.onnx/models/opset_9/resnet50.tar.gz)  
-8. [shufflenet](https://s3.amazonaws.com/download.onnx/models/opset_9/shufflenet.tar.gz)  
-9. [squeezenet](https://s3.amazonaws.com/download.onnx/models/opset_9/squeezenet.tar.gz)  
-10. [vgg19](https://s3.amazonaws.com/download.onnx/models/opset_9/vgg19.tar.gz)  
-11. [zfnet512](https://s3.amazonaws.com/download.onnx/models/opset_9/zfnet512.tar.gz)
+
+验证工具onnx2fluid.validate:
+
+```shell
+onnx2fluid.validate [-d] [-t test_data.npz] [-p 1e-3] /path/to/onnx/model.onnx
+```
+
+## 参考
+
+* PaddlePaddle [算子](http://www.paddlepaddle.org/documentation/docs/zh/1.4/api_cn/layers_cn.html)
+* PaddlePaddle [加载预测模型](http://www.paddlepaddle.org/documentation/docs/zh/1.4/api_guides/low_level/inference.html#id4)

onnx2fluid/README_en.md

@@ -2,13 +2,25 @@
 
 [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
 
-onnx2fluid supports converting ONNX model to PaddlePaddle Model for prediction.
+onnx2fluid supports converting ONNX model to PaddlePaddle fluid model for prediction.
 
-## Running Environment
+PyTorch to Paddlepaddle model conversion can be easily achieved with PyTorch ONNX export functions.
 
-* python 3.5+ (python 2 working in progress)
+## Features
+
+* Python code + ProgramDesc proto generation, flexible and compatible
+* fluid layer weight embedding support
+* conversion, validation, archiving all in one
+* convert without PaddlePaddle dependency
+* export and validation helper functions for PyTorch to PaddlePaddle conversion
+* extra ONNX operator optimization for inference
+* easily extensible for user-defined operators
+
+## Environment and dependency
+
+* python 3.5+ (python 2 not fully supported yet)
 * onnx == 1.4.0
-* paddlepaddle == 1.3.0
+* paddlepaddle == 1.3.0 (optional for validation)
 
 ## Get started
 
@@ -20,26 +32,50 @@ cd examples
 sh onnx_model_zoo.sh
 ```
 
-Try exporting from PyTorch to Paddle fluid:
+Try exporting and validating from PyTorch to PaddlePaddle fluid:
 
 ``` shell
 python setup.py install
 cd examples
+
 python gen_some_samples.py
 onnx2fluid sample_1.onnx -t sample_1.npz
+
+python gen_unet.py
+onnx2fluid sample_unet.onnx -t sample_unet.npz
 ```
 
 ## Usage
 
+onnx2fluid (all in one):
+
 ```shell
-onnx2fluid [-dexy] -o /path/to/export_dir/ /path/of/onnx/model.onnx
+onnx2fluid [-dexy] [-o /path/to/export_dir/] [-z archive.zip] [-t test_data.npz] /path/to/onnx/model.onnx
 
 optional arguments:
-  --embed_params, -e    try to embed parameters for trainable Paddle fluid layers
+  --debug, -d           enable debug logging and checking
+  --embed_params, -e    try to embed parameters for trainable PaddlePaddle fluid layers
   --no-pedantic, -x     process non-standard ONNX ops
   --skip-version-conversion, -y
-                        skip ONNX op version conversion, workaround for
-                        RumtimeErrors
+                        skip ONNX op version conversion, workaround for RumtimeErrors
+  --output_dir, -o      output directory
   --archive [ARCHIVE], -z [ARCHIVE]
                         compress outputs to ZIP file if conversion successed
 ```
+
+onnx2fluid.conversion:
+
+```shell
+onnx2fluid.conversion [-dexy] [-o /path/to/export_dir/] /path/to/onnx/model.onnx
+```
+
+onnx2fluid.validate:
+
+```shell
+onnx2fluid.validate [-d] [-t test_data.npz] [-p 1e-3] /path/to/onnx/model.onnx
+```
+
+## Reference
+
+* [PaddlePaddle fluid operators](http://www.paddlepaddle.org/documentation/docs/en/1.4/api/layers.html)
+* load converted model via [load_inference_model](http://www.paddlepaddle.org/documentation/docs/en/1.4/api/io.html#permalink-1-load_inference_model)

onnx2fluid/examples/convert_data_pb_0.py

@@ -6,9 +6,7 @@ Created on Wed Mar 27 11:50:03 2019
 @author: Macrobull
 """
 
-# import os, sys
-import os
-import sys
+import os, sys
 import numpy as np
 import onnx
 import onnx.numpy_helper as numpy_helper

onnx2fluid/examples/gen_some_samples.py

@@ -7,7 +7,6 @@ Created on Fri Mar 22 11:19:45 2019
 
 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 fluid
-
 """
 
 from __future__ import print_function

onnx2fluid/examples/gen_unet.py

@@ -4,7 +4,6 @@
 Created on Fri Mar 22 11:19:45 2019
 
 @author: Macrobull
-
 """
 
 from __future__ import print_function

onnx2fluid/examples/gen_yolov2.py

@@ -4,7 +4,6 @@
 Created on Fri Mar 22 11:19:45 2019
 
 @author: Macrobull
-
 """
 
 from __future__ import print_function

onnx2fluid/examples/onnx_model_zoo.sh

@@ -97,8 +97,8 @@ bvlc_reference_rcnn_ilsvrc13()
     do
         echo "converting $pb_dir"
         python convert_data_pb_0.py "$pb_dir" data_0 fc-rcnn_1
-        python -m onnx2fluid.validation $validate_flags1 -t $(dirname "$pb_dir/x").npz -p 0
-        python -m onnx2fluid.validation $validate_flags2 -t $(dirname "$pb_dir/x").npz -p 0
+        python -m onnx2fluid.validation $validate_flags1 -t $(dirname "$pb_dir/x").npz
+        python -m onnx2fluid.validation $validate_flags2 -t $(dirname "$pb_dir/x").npz
     done
 }
 

onnx2fluid/onnx2fluid/__main__.py

@@ -16,10 +16,7 @@ from __future__ import division
 from __future__ import print_function
 from __future__ import unicode_literals
 
-# import argparse, logging, sys
-import argparse
-import logging
-import sys
+import argparse, logging, sys
 
 parser = argparse.ArgumentParser(
     description='onnx2fluid',
@@ -86,11 +83,17 @@ parser.add_argument(
     help='compress outputs to ZIP file if conversion successed',
 )
 parser.add_argument(
-    '--precision',
+    '--atol',
     '-p',
     type=float,
-    default=3.,
-    help='assertion decimal for validation',
+    default=1e-3,
+    help='assertion absolute tolerance for validation',
+)
+parser.add_argument(
+    '--rtol',
+    type=float,
+    default=1e-4,
+    help='assertion relative tolerance for validation',
 )
 args = parser.parse_args()
 
@@ -98,12 +101,6 @@ logging_format = '[%(levelname)8s]%(name)s::%(funcName)s:%(lineno)04d: %(message
 logging_level = logging.DEBUG if args.debug else logging.INFO
 logging.basicConfig(format=logging_format, level=logging_level)
 
-try:
-    from . import cmdline
-except ImportError:
-    import cmdline
-
-# imports
-main = cmdline.main
+from .cmdline import main
 
 sys.exit(main(**args.__dict__))

onnx2fluid/onnx2fluid/cmdline.py

@@ -17,9 +17,6 @@ from __future__ import print_function
 from __future__ import unicode_literals
 
 import logging, shutil, zipfile
-#import logging
-#import shutil
-#import zipfile
 
 __all__ = [
     'main',
@@ -33,31 +30,23 @@ DEFAULT_MODEL_FUNC = 'inference'
 def main(**kwargs):
     """主程序入口"""
 
-    try:
-        from . import conversion
-    except ImportError:
-        import conversion
-
-    # imports
-    convert = conversion.convert
+    from .conversion import convert
 
     logger = logging.getLogger('onnx2fluid')
-    debug = kwargs.get('debug', False)
+    #    debug = kwargs.get('debug', False)
 
     # prepare arguments
-    filename = kwargs['model'][0]
+    filename = kwargs.pop('model')[0]
     basepath, _ = shutil.os.path.splitext(filename)
-    save_dir = kwargs.get('output_dir', '')
+    save_dir = kwargs.pop('output_dir', '')
     # model.onnx -> model/
     save_dir = (save_dir.rstrip(shutil.os.sep)
                 if save_dir else basepath) + shutil.os.sep
     model_basename = DEFAULT_MODEL_MODULE + '.py'
     model_func_name = DEFAULT_MODEL_FUNC
-    embed_params = kwargs.get('embed_params', False)
     onnx_opset_version = DEFAULT_ONNX_OPSET_VERSION
-    onnx_opset_pedantic = kwargs.get('pedantic', True)
-    onnx_skip_version_conversion = kwargs.get('skip_version_conversion', False)
-    archive = kwargs.get('archive', None)
+    onnx_opset_pedantic = kwargs.pop('pedantic', True)
+    onnx_skip_version_conversion = kwargs.pop('skip_version_conversion', False)
 
     # convert
     convert(
@@ -65,49 +54,35 @@ def main(**kwargs):
         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,
         onnx_skip_version_conversion=onnx_skip_version_conversion,
-        debug=debug)
+        **kwargs)
 
     # validate
     passed = True
-    golden_data_filename = kwargs.get('test_data', '')
+    golden_data_filename = kwargs.pop('test_data', '')
     if golden_data_filename:
-        try:
-            from . import validation
-        except ImportError:
-            import validation
-
-        # imports
-        validate = validation.validate
-
-        # in fact fluid can not fully clear the context
-        # continuous validation may be inaccurate
-        decimal = kwargs.get('precision', 3.)
+        from .validation import validate
 
         logger.info('starting validation on desc ...')
         passed &= validate(
-            shutil.os.path.join(save_dir, '__model__'),
-            golden_data_filename,
-            decimal=decimal,
-        )
+            shutil.os.path.join(save_dir, '__model__'), golden_data_filename,
+            **kwargs)
 
         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,
-            decimal=decimal,
-            save_inference_model=debug,  # this overwrite desc file for test
-        )
+            **kwargs)
 
     if not passed:
         logger.error('validation failed, exit')
         return
 
     # create zip file
+    archive = kwargs.pop('archive', None)
     if archive is not None:
         if archive == '':
             archive = save_dir.rstrip(shutil.os.sep) + '.zip'
@@ -132,6 +107,10 @@ if __name__ == '__main__':
         level=logging.DEBUG,
     )
 
+    del main
+
+    from onnx2fluid.cmdline import main
+
     main(
         model=['../examples/t1.onnx'],
         output_dir='/tmp/export/',

onnx2fluid/onnx2fluid/conversion.py

@@ -9,8 +9,6 @@ Created on Mon Feb 25 09:50:35 2019
 from __future__ import division
 
 import logging, shutil
-#import logging
-#import shutil
 
 __all__ = [
     'convert',
@@ -25,7 +23,8 @@ def convert(onnx_model_filename,
             onnx_opset_version=9,
             onnx_opset_pedantic=True,
             onnx_skip_version_conversion=False,
-            debug=False):
+            debug=False,
+            **kwargs):
     """
     convert an ONNX model to Paddle fluid Python code and desc pb
     """
@@ -37,21 +36,14 @@ def convert(onnx_model_filename,
     from onnx.utils import polish_model
     from onnx.version_converter import convert_version
 
-    try:
-        from . import onnx_utils, writer
-    except ImportError:
-        import onnx_utils, writer
-
-    # imports
-    DEFAULT_OP_DOMAIN = onnx_utils.DEFAULT_OP_DOMAIN
-    graph_ops, graph_weights = onnx_utils.graph_ops, onnx_utils.graph_weights
-    inferred_model_value_info = onnx_utils.inferred_model_value_info
-    optimize_model_skip_op_for_inference = onnx_utils.optimize_model_skip_op_for_inference
-    optimize_model_strip_initializer = onnx_utils.optimize_model_strip_initializer
-    optimize_model_cast = onnx_utils.optimize_model_cast
-    optimize_model_slice = onnx_utils.optimize_model_slice
-    Program, Writer = writer.Program, writer.Writer
-    make_var_name = writer.make_var_name
+    from .onnx_utils import DEFAULT_OP_DOMAIN
+    from .onnx_utils import graph_ops, graph_weights
+    from .onnx_utils import inferred_model_value_info
+    from .onnx_utils import optimize_model_skip_op_for_inference
+    from .onnx_utils import optimize_model_strip_initializer
+    from .onnx_utils import optimize_model_cast, optimize_model_slice
+    from .writer import Program, Writer
+    from .writer import make_var_name
 
     logger = logging.getLogger('convert')
 
@@ -94,6 +86,8 @@ def convert(onnx_model_filename,
         model = onnx.shape_inference.infer_shapes(onnx_model)
         debug_model_filename, _ = shutil.os.path.splitext(onnx_model_filename)
         onnx.save(model, debug_model_filename + '.optimized_and_inffered.onnx')
+
+
 #        onnx.save(model, '/tmp/export/optimized_and_inffered.onnx')
 
 # I/O instances
@@ -218,12 +212,13 @@ def convert(onnx_model_filename,
 
     logger.info('conversion finished')
 
-
-#    globals().update(locals())
-
 if __name__ == '__main__':
+    del convert
+
     import argparse
 
+    from onnx2fluid.conversion import convert
+
     parser = argparse.ArgumentParser(
         description='onnx2fluid.convert',
         formatter_class=argparse.ArgumentDefaultsHelpFormatter,
@@ -280,7 +275,10 @@ if __name__ == '__main__':
 
     debug = args.debug
     model_filename = args.model[0]
+    basepath, _ = shutil.os.path.splitext(model_filename)
     save_dir = args.output_dir
+    save_dir = (save_dir.rstrip(shutil.os.sep)
+                if save_dir else basepath) + shutil.os.sep
     embed_params = args.embed_params
     pedantic = args.pedantic
     skip_version_conversion = args.skip_version_conversion

onnx2fluid/onnx2fluid/symbolic.py

@@ -13,7 +13,7 @@ Created on Mon Feb 25 09:33:43 2019
 from __future__ import division
 
 import logging as _logging
-import numpy as np
+import numpy as _np
 
 from collections import OrderedDict as _dict
 from onnx.mapping import TENSOR_TYPE_TO_NP_TYPE
@@ -164,7 +164,7 @@ def _make_var_name(name):
 
 
 def _dtype(value_infos, val_name):
-    return np.dtype(value_infos[val_name]['dtype'])
+    return _np.dtype(value_infos[val_name]['dtype'])
 
 
 def _dtype_or_none(value_infos, val_name):
@@ -173,7 +173,7 @@ def _dtype_or_none(value_infos, val_name):
     value_info = value_infos[val_name]
     if 'dtype' not in value_info:
         return None
-    return np.dtype(value_info['dtype'])
+    return _np.dtype(value_info['dtype'])
 
 
 def _shape(value_infos, val_name):
@@ -268,14 +268,13 @@ def _default(prog, op_type, inputs, outputs, attrs, *args, name='', **kwargs):
                 (var_outs, *fluid_output_args), fluid_attrs)
 
 
-def _assign(prog, attrs):
-    mapping = attrs['mapping']  # additional
+def _assign(prog, mapping):
     fluid_op = 'assign'
 
     for val_dst, val_src in mapping.items():
         var_dst = _make_var_name(val_dst)
         var_src = _make_var_name(val_src)
-        prog.Code('{} = {}'.format(var_dst, var_src))
+        prog.Code('{} = {} # assign'.format(var_dst, var_src))
         #        prog.Code('{} = layers.{}({})'
         #                  .format(var_dst,
         #                          fluid_op,
@@ -774,7 +773,7 @@ def Cast(prog, inputs, outputs, attrs, value_infos, *args, **kwargs):
 
     # interpretation
     dtype = attrs['to']  # required
-    if not isinstance(dtype, np.dtype):  # additional: possible np.dtype
+    if not isinstance(dtype, _np.dtype):  # additional: possible np.dtype
         dtype = TENSOR_TYPE_TO_NP_TYPE[dtype]
     output_dtype = _dtype_or_none(value_infos, val_output)
     if output_dtype:
@@ -851,13 +850,13 @@ def Constant(prog, inputs, outputs, attrs, value_infos, *args, **kwargs):
 
     # interpretation
     value = attrs['value']  # required
-    dtype = np.dtype(value.dtype)
+    dtype = _np.dtype(value.dtype)
     output_dtype = _dtype_or_none(value_infos, val_output)
     if output_dtype:
         assert dtype == output_dtype, 'tensor dtype unmatches storage dtype'
 
 
-#    dtype = np.dtype('float32') # HINT: force to float32
+#    dtype = _np.dtype('float32') # HINT: force to float32
     shape = attrs.get('shape', None)  #
     if shape is None:
         shape = _shape_or_none(value_infos, val_output)
@@ -1160,7 +1159,7 @@ def ConvTranspose(prog,
 #    val_output, = outputs[:1]
 #
 #    _assign(prog,
-#            dict(mapping=dict([(val_output, val_data)])),
+#            dict([(val_output, val_data)]),
 #            value_infos,
 #            )
 
@@ -1215,13 +1214,13 @@ def Gemm(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs):
             if beta.is_integer():
                 vm_dtype = _dtype_or_none(value_infos, val_c)
                 if vm_dtype is None:
-                    vm_dtype = np.dtype('float32')
+                    vm_dtype = _np.dtype('float32')
                     _logger.warning(
                         'in %s(%s -> Gemm -> %s): '
                         'attribute "beta" seems to be an interger, '
                         'however dtype can not be inferred, '
                         'still use float32', name, inputs, outputs)
-                beta = np.dtype(vm_dtype).type(beta)
+                beta = _np.dtype(vm_dtype).type(beta)
             prog.Op(
                 '',
                 'Constant',
@@ -1387,7 +1386,7 @@ def Pad(prog, inputs, outputs, attrs, value_infos, name='', *args, **kwargs):
             mode)
         fluid_op = 'pad'
         pad2d_attr = ''
-    paddings = np.array(pads).reshape(
+    paddings = _np.array(pads).reshape(
         (-1, 2)).transpose().flatten().tolist()  # SSEE -> SESE
     od_attrs['paddings'] = paddings
     name_attr = ', name={}'.format(repr(name)) if name else ''
@@ -1525,7 +1524,7 @@ def Reshape(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs):
             'Cast',
             [val_shape],
             [val_shape_int32],  # var
-            dict(to=np.dtype('int32')),  # use np.dtype
+            dict(to=_np.dtype('int32')),  # use np.dtype
             value_infos=value_infos,
             name=(name + '_cast'),
         )
@@ -1840,7 +1839,9 @@ if __name__ == '__main__':
     )
     logger = _logging.getLogger('symbolic_test')
 
-    from writer import Program
+    import numpy as np
+
+    from onnx2fluid.writer import Program
 
     prog = Program()
     AdaptiveAveragePool(

onnx2fluid/onnx2fluid/validation.py

@@ -8,11 +8,6 @@ Created on Fri Mar 22 12:17:19 2019
 
 import importlib, logging, os, sys
 
-#import importlib
-#import logging
-#import os
-#import sys
-
 
 def _flatten_dict(obj, out=None):
     assert isinstance(obj, dict)
@@ -37,8 +32,10 @@ def _ensure_list(obj):
 def validate(fluid_model_filename,
              golden_data_filename,
              model_func_name='inference',
-             decimal=3,
-             save_inference_model=False):
+             atol=1e-3,
+             rtol=1e-4,
+             save_inference_model=False,
+             **kwargs):
     """
     inference the converted Paddle fluid model, validate with given golden data
     """
@@ -125,7 +122,13 @@ def validate(fluid_model_filename,
     for (name, truth), output in zip(output_data.items(), outputs):
         logger.info('testing output {} ...'.format(name))
         try:
-            np.testing.assert_almost_equal(output, truth, decimal=decimal)
+            np.testing.assert_allclose(
+                output,
+                truth,
+                rtol=rtol,
+                atol=atol,
+                equal_nan=False,
+                verbose=True)
         except AssertionError as e:
             passed = False
             logger.error('failed: %s\n', e)
@@ -134,10 +137,9 @@ def validate(fluid_model_filename,
     else:
         logger.info('accuracy not passed')
 
-
-#    globals().update(locals())
     return passed
 
+
 if __name__ == '__main__':
     import argparse
 
@@ -163,11 +165,17 @@ if __name__ == '__main__':
         help='I/O golden data for validation, e.g. test.npy, test.npz',
     )
     parser.add_argument(
-        '--precision',
+        '--atol',
         '-p',
         type=float,
-        default=3.,
-        help='assertion decimal for validation',
+        default=1e-3,
+        help='assertion absolute tolerance for validation',
+    )
+    parser.add_argument(
+        '--rtol',
+        type=float,
+        default=1e-4,
+        help='assertion relative tolerance for validation',
     )
     args = parser.parse_args()
 
@@ -178,10 +186,11 @@ if __name__ == '__main__':
     debug = args.debug
     fluid_model_filename = args.model[0]
     golden_data_filename = args.test_data
-    decimal = args.precision
+    atol, rtol = args.atol, args.rtol
 
     validate(
         fluid_model_filename,
         golden_data_filename,
-        decimal=decimal,
+        atol=atol,
+        rtol=rtol,
         save_inference_model=debug)

onnx2fluid/onnx2fluid/writer.py

@@ -9,27 +9,17 @@ Created on Sun Feb 24 20:44:43 2019
 from __future__ import division
 
 import logging, os
-#import logging
-#import os
 import numpy as np
 
 logger = logging.getLogger(__name__)
 
-try:
-    from . import symbolic
-except ImportError:
-    import symbolic
-
-# imports
-make_var_name = symbolic._make_var_name
+from . import symbolic
+from .symbolic import _make_var_name as make_var_name
 
 try:
     import paddle.fluid.proto.framework_pb2 as framework_pb2
 except ImportError:
-    try:
-        from . import framework_pb2
-    except ImportError:
-        import framework_pb2
+    from . import framework_pb2
 
     logger.warning('importing paddle.fluid.proto.framework_pb2d failed,'
                    'using fallback framework_pb2')
@@ -215,11 +205,9 @@ class Program(object):
                 name,
                 persistable=False,
                 value_info=None,
-                remove_batch=None,
-                dummy_dtype='float32'):
+                remove_batch=None):
         """
         add VarDesc,
-        dummy_dtype: WORKAROUND for Netron viewer
         """
 
         var_desc = framework_pb2.VarDesc()
@@ -238,9 +226,6 @@ class Program(object):
                                                       not persistable)
                     if remove_batch:
                         tensor_desc.dims[0] = -1
-        else:  # REMOVEIT: WORKAROUND: Netron: null.tensor error
-            tensor_desc = var_desc.type.lod_tensor.tensor
-            tensor_desc.data_type = self.Dtype(dummy_dtype)  # required
 
         self.var_descs.append(var_desc)
 

onnx2fluid/requirements.txt

 -e .
-onnx>=1.4.0
+onnx>=1.4
 paddlepaddle

onnx2fluid/setup.cfg

@@ -7,7 +7,7 @@ name = onnx2fluid
 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
+version = 0.1.1
 # 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 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