Merge branch 'develop' of https://github.com/Channingss/X2Paddle into develop

.travis.yml

+language: python
+
+python:
+  - '2.7'
+  - '3.5'
+  - '3.6'
+
+script:
+  - if [[ $TRAVIS_PYTHON_VERSION != 2.7 ]]; then /bin/bash ./scripts/check_code_style.sh; fi
+
+notifications:
+  email:
+    on_success: change
+    on_failure: always

FAQ.md

 ## 常见问题
 
 **Q1. TensorFlow模型转换过程中，提示『Unknown shape for input tensor[tensor name: "input"]， Please define shape of input here』？**  
-A：该提示信息表示无法从TensorFlow的pb模型中获取到输入tensor(tensor名为"input:)的shape信息，所以需要用户手动在提示后输入详细的shape信息，如None,224,224,3 其中None表示Batch 
+A：该提示信息表示无法从TensorFlow的pb模型中获取到输入tensor(tensor名为"input:)的shape信息，所以需要用户手动在提示后输入详细的shape信息，如None,224,224,3 其中None表示Batch
+
+
+**Q2. TensorFlow模型转换失败怎么解决?**  
+A: 如果并非是由缺少OP导致，那可能是由于TensorFlow模型转换时（NHWC->NCHW格式转换导致），在这种情况下，采用如下方式进行转换，同时固化输入大小的方式，继续尝试转换，见如下命令，转换过程中，根据提示，输入相应tensor的固化shape大小
+```
+x2paddle -f tensorflow -m tf.pb -s pd-model --without_data_format_optimization --define_input_shape
+```
+
+> 1. 目前Tensorflow的CV模型大部分均为`NHWC`的输入格式，而Paddle的默认输入格式为`NCHW`，因此X2Paddle在转换过程中，会对如`axis`， `shape`等参数进行转换，适应Paddle的NCHW格式。但在这种情况下，可能会由于TensorFlow模型太复杂，导致出错。  指定`--without_data_format_optimization`后，会停止对`axis`，`shape`等参数的优化（这可能会带来一定数量的transpose操作）

README.md

@@ -4,22 +4,38 @@
 X2Paddle支持将其余深度学习框架训练得到的模型，转换至PaddlePaddle模型。  
 X2Paddle is a toolkit for converting trained model to PaddlePaddle from other deep learning frameworks.
 
+## 转换模型库
+X2Paddle在多个主流的CV模型上，测试过TensorFlow/Caffe/ONNX模型的转换，可以在[X2Paddle-Model-Zoo](x2paddle_model_zoo.md)查看我们的模型测试列表。如果你在新的模型上进行了测试转换，也欢迎继续补充该列表；如若无法转换，可通过ISSUE反馈给我们，我们会尽快跟进。
+
 ## 环境依赖
 
 python >= 3.5  
 paddlepaddle >= 1.5.0  
 
-**以下依赖只需对应安装自己需要的即可**  
-转换tensorflow模型 ： tensorflow == 1.14.0  
-转换caffe模型 ： caffe == 1.0.0  
-转换onnx模型 ： onnx == 1.5.0  pytorch == 1.1.0
+**按需安装以下依赖**  
+tensorflow ： tensorflow == 1.14.0  
+caffe ： 无  
+onnx ： onnx == 1.5.0  pytorch == 1.1.0
+
 ## 安装
+### 安装方式一（推荐）
+使用最新的代码版本，可使用如下方式进行安装  
+```
+pip install git+https://github.com/PaddlePaddle/X2Paddle.git@develop
+```
+
+### 安装方式二
+我们会定期更新pip源上的x2paddle版本
 ```
 pip install x2paddle
 ```
-如果需要使用最新的代码版本，可使用如下方式进行安装  
+
+### 安装方式三
 ```
-pip install git+https://github.com/PaddlePaddle/X2Paddle.git@develop
+git clone https://github.com/PaddlePaddle/X2Paddle.git
+cd X2Paddle
+git checkout develop
+python setup.py install
 ```
 
 ## 使用方法
@@ -38,30 +54,42 @@ x2paddle --framework=onnx --model=onnx_model.onnx --save_dir=pd_model
 ### 参数选项
 | 参数 | |
 |----------|--------------|
-|--framework | 源模型类型 (tensorflow、caffe) |
+|--framework | 源模型类型 (tensorflow、caffe、onnx) |
 |--prototxt | 当framework为caffe时，该参数指定caffe模型的proto文件路径 |
 |--weight | 当framework为caffe时，该参数指定caffe模型的参数文件路径 |
 |--save_dir | 指定转换后的模型保存目录路径 |
-|--model | 当framework为tensorflow时，该参数指定tensorflow的pb模型文件路径 |
-|--caffe_proto | [可选]由caffe.proto编译成caffe_pb2.py文件的存放路径，当没有安装caffe或者使用自定义Layer时使用，默认为None |
+|--model | 当framework为tensorflow/onnx时，该参数指定tensorflow的pb模型文件或onnx模型路径 |
+|--caffe_proto | **[可选]** 由caffe.proto编译成caffe_pb2.py文件的存放路径，当存在自定义Layer时使用，默认为None |
+|--without_data_format_optimization | **[可选]** For TensorFlow, 当指定该参数时，关闭NHWC->NCHW的优化，见[文档Q2](FAQ.md) |
+|--define_input_shape | **[可选]** For TensorFlow, 当指定该参数时，强制用户输入每个Placeholder的shape，见[文档Q2](FAQ.md) |
+
 
 ## 使用转换后的模型
 转换后的模型包括`model_with_code`和`inference_model`两个目录。  
 `model_with_code`中保存了模型参数，和转换后的python模型代码  
 `inference_model`中保存了序列化的模型结构和参数，可直接使用paddle的接口进行加载，见[load_inference_model](https://www.paddlepaddle.org.cn/documentation/docs/zh/1.5/api_guides/low_level/inference.html#api-guide-inference)
 
+## 小工具
+X2Paddle提供了工具解决如下问题，详见[tools/README.md](tools/README.md)
+1. 检测模型是否在PaddleLite中支持  
+2. 合并模型参数文件
+
 ## 相关文档
 1. [X2Paddle使用过程中常见问题](FAQ.md)  
 2. [如何导出TensorFlow的pb模型](export_tf_model.md)
-3. [X2Paddle测试模型库](x2paddle_model_zoo.md)
+3. [X2Paddle测试模型库](x2paddle_model_zoo.md)  
+4. [PyTorch模型导出为ONNX模型](pytorch_to_onnx.md)
+5. [X2Paddle内置的Caffe自定义层](caffe_custom_layer.md)
 
 ## 更新历史
 2019.08.05  
-1. 统一tensorflow/caffe(onnx的代码即将release)模型转换代码和对外接口
+1. 统一tensorflow/caffe/onnx模型转换代码和对外接口
 2. 解决上一版caffe2fluid无法转换多分支模型的问题
 3. 解决Windows上保存模型无法加载的问题
 4. 新增optimizer，优化代码结构，合并conv、batch_norm的bias和激活函数  
-> X2Paddle 0.3及之前版本各子转换工具相互独立，在0.4版本，我们将tensorflow/caffe/onnx三个工具的代码整合，提供更易用的对外使用方式。用户可继续在[分支release-0.3](https://github.com/PaddlePaddle/X2Paddle/tree/release-0.3)中访问和使用之前的版本。
+
+**如果你需要之前版本的tensorflow2fluid/caffe2fluid/onnx2fluid，可以继续访问release-0.3分支，获取之前版本的代码使用。**
+
 
 ## Acknowledgements
 

add_caffe_custom_layer.md

+## 如何转换Caffe自定义Layer
+
+本文档介绍如何将Caffe自定义Layer转换为PaddlePaddle模型中的对应实现, 用户可根据自己需要，添加代码实现自定义层，从而支持模型的完整转换。  
+***步骤一 下载代码***  
+此处涉及修改源码，应先卸载x2paddle，并且下载源码，主要有以下两步完成：
+```
+pip uninstall x2paddle
+pip install git+https://github.com/PaddlePaddle/X2Paddle.git@develop
+```
+
+***步骤二 编译caffe.proto***  
+该步骤依赖protobuf编译器，其安装过程有以下两种方式：
+> 选择一：pip install protobuf  
+> 选择二：使用[官方源码](https://github.com/protocolbuffers/protobuf)进行编译
+
+使用脚本./tools/compile.sh将caffe.proto（包含所需的自定义Layer信息）编译成我们所需的目标语言（Python）  
+使用方式：
+```
+bash ./toos/compile.sh /home/root/caffe/src/caffe/proto
+# /home/root/caffe/src/caffe/proto为caffe.proto的存放路径，生成的caffe_pb2.py也将保存在该路径下
+```
+
+***步骤三 添加自定义Layer的实现代码***
+- 进入./x2paddle/op_mapper/caffe_custom_layer，创建.py文件，例如mylayer.py
+- 仿照./x2paddle/op_mapper/caffe_custom_layer中的其他文件，在mylayer.py中主要需要实现3个函数，下面以roipooling.py为例分析代码：
+  1. `def roipooling_shape(input_shape, pooled_w=None, pooled_h=None)`  
+     参数：
+     1. input_shape（list）：其中每个元素代表该层每个输入数据的shape，为必须传入的参数  
+     2. pooled_w（int）：代表ROI Pooling的kernel的宽，其命名与.prototxt中roi_pooling_param中的key一致
+     3. pooled_h（int）：代表ROI Pooling的kernel的高，其命名与.prototxt中roi_pooling_param中的key一致  
+
+     功能：计算出进行ROI Pooling后的shape  
+     返回：一个list，其中每个元素代表每个输出数据的shape，由于ROI Pooling的输出数据只有一个，所以其list长度为1  
+
+  2. `def roipooling_layer(inputs, input_shape=None, name=None, pooled_w=None, pooled_h=None, spatial_scale=None)`
+
+     参数：
+     1. inputs（list）：其中每个元素代表该层每个输入数据，为必须传入的参数
+     2. input_shape（list）：其中每个元素代表该层每个输入数据的shape，为必须传入的参数  
+     3. name（str）：ROI Pooling层的名字，为必须传入的参数  
+     4. pooled_w（int）：代表ROI Pooling的kernel的宽，其命名与.prototxt中roi_pooling_param中的key一致
+     5. pooled_h（int）：代表ROI Pooling的kernel的高，其命名与.prototxt中roi_pooling_param中的key一致
+     6. spatial_scale（float）：用于将ROI坐标从输入比例转换为池化时使用的比例，其命名与.prototxt中roi_pooling_param中的key一致  
+
+     功能：运用PaddlePaddle完成组网来实现`roipooling_layer`的功能  
+     返回：一个Variable，为组网后的结果
+
+  3. `def roipooling_weights(name, data=None)`  
+
+     参数：
+     1. name（str）：ROI Pooling层的名字，为必须传入的参数
+     2. data（list）：由Caffe模型.caffemodel获得的关于roipooling的参数，roipooling的参数为None
+
+     功能：为每个参数（例如kernel、bias等）命名；同时，若Caffe中该层参数与PaddlePaddle中参数的格式不一致，则变换操作也在该函数中实现。  
+     返回：一个list，包含每个参数的名字。
+
+- 在roipooling.py中注册`roipooling`，主要运用下述代码实现：
+  ```
+  register(kind='ROIPooling', shape=roipooling_shape, layer=roipooling_layer, weights=roipooling_weights)
+  # kind为在model.prototxt中roipooling的type
+  ```
+- 在./x2paddle/op_mapper/caffe_custom_layer/\_\_init\_\_.py中引入该层的使用
+  ```
+  from . import roipooling
+  ```
+
+
+***步骤四 运行转换代码***
+```
+# 在X2Paddle目录下安装x2paddle
+python setup.py install
+# 运行转换代码
+x2paddle --framework=caffe
+         --prototxt=deploy.proto
+         --weight=deploy.caffemodel
+         --save_dir=pd_model
+         --caffe_proto=/home/root/caffe/src/caffe/proto/caffe_pb2.py
+```

caffe_custom_layer.md

+目前，代码中已经提供了8个非官方op（不在[官网](http://caffe.berkeleyvision.org/tutorial/layers)上的op）的转换，这些op对应的Caffe实现源码如下：
+
+| op | 该版本实现源码 |
+|-------|--------|
+| PriorBox | [code](https://github.com/weiliu89/caffe/blob/ssd/src/caffe/layers/prior_box_layer.cpp) |
+| DetectionOutput | [code](https://github.com/weiliu89/caffe/blob/ssd/src/caffe/layers/detection_output_layer.cpp) |
+| ConvolutionDepthwise | [code](https://github.com/farmingyard/caffe-mobilenet/blob/master/conv_dw_layer.cpp) |
+| ShuffleChannel | [code](https://github.com/farmingyard/ShuffleNet/blob/master/shuffle_channel_layer.cpp) |
+| Permute | [code](https://github.com/weiliu89/caffe/blob/ssd/src/caffe/layers/permute_layer.cpp) |
+| Normalize | [code](https://github.com/weiliu89/caffe/blob/ssd/src/caffe/layers/normalize_layer.cpp) |
+| ROIPooling | [code](https://github.com/rbgirshick/caffe-fast-rcnn/blob/0dcd397b29507b8314e252e850518c5695efbb83/src/caffe/layers/roi_pooling_layer.cpp) |
+| Axpy | [code](https://github.com/hujie-frank/SENet/blob/master/src/caffe/layers/axpy_layer.cpp) |

export_tf_model.md

@@ -28,7 +28,7 @@ def freeze_model(sess, output_tensor_names, freeze_model_path):
         f.write(out_graph.SerializeToString())
 
     print("freeze model saved in {}".format(freeze_model_path))
-    
+
 # 加载模型参数
 sess = tf.Session()
 inputs = tf.placeholder(dtype=tf.float32,

pytorch_to_onnx.md

+## PyTorch模型导出为ONNX模型
+
+目前onnx2paddle主要支持onnx operator version 9。 用户可通过如下示例代码，将torchvison或者自己开发写的模型转换成onnx model:
+```
+#coding: utf-8
+import torch
+import torchvision
+
+# 指定输入大小的shape
+dummy_input = torch.randn(1, 3, 224, 224)
+
+# 构建pytorch model，并载入模型参数
+resnet18 = torchvision.models.resnet18(pretrained=True)
+
+# 导出resnet18.onnx模型文件
+torch.onnx.export(resnet18, dummy_input, "resnet18.onnx",verbose=True)
+
+```

requirements.txt

+pre-commit
+yapf == 0.28.0

scripts/check_code_style.sh

+#!/bin/bash
+function abort(){
+    echo "Your change doesn't follow X2Paddle's code style." 1>&2
+    echo "Please use pre-commit to check what is wrong." 1>&2
+    exit 1
+}
+
+trap 'abort' 0
+set -e
+cd $TRAVIS_BUILD_DIR
+export PATH=/usr/bin:$PATH
+pre-commit install
+
+if ! pre-commit run -a ; then
+    git diff
+    exit 1
+fi
+
+trap : 0

tools/README.md

+### 一、PaddleLite部署
+使用X2Paddle转换后的模型均可以使用Paddle Fluid进行预测。但对于PaddleLite上的部署，则需要检查模型中的OP是否都在PaddleLite中支持。使用`check_for_lite.py`可以进行检查。
+
+```
+python tools/check_for_lite.py paddle_model/inference_model/__model__
+```
+
+
+### 二、模型参数合并
+X2Paddle转换后产出的路径下包括两个目录，  
+1. `model_with_code`: 包含保存的参数文件和模型python代码文件，供用户debug  
+2. `inference_model`: 参数文件和序列化的模型结构文件，供用户预测部署  
+
+其中在`inference_model`中，X2Paddle将每个参数独立保存在不同的文件中（文件名和参数名一致），用户可使用`merge_params.py`将参数文件合并成一个文件使用
+```
+python tools/merge_params.py paddle_model/inference_model  new_model_dir
+```
+合并参数后的模型保存在`new_model_dir`中

tools/check_for_lite.py

+import urllib
+import sys
+from paddle.fluid.framework import Program
+
+ops_h = "https://raw.githubusercontent.com/PaddlePaddle/Paddle-Lite/develop/lite/api/_paddle_use_ops.h"
+try:
+    fetch = urllib.urlretrieve(ops_h, "./_paddle_use_ops.h")
+except:
+    fetch = urllib.request.urlretrieve(ops_h, "./_paddle_use_ops.h")
+
+ops = list()
+with open("./_paddle_use_ops.h") as f:
+    for line in f:
+        if "USE_LITE_OP" in line:
+            op = line.strip().split('(')[1].split(')')[0]
+            ops.append(op)
+
+model_file = sys.argv[1]
+with open(model_file, 'rb') as f:
+    program = Program.parse_from_string(f.read())
+
+unsupported_ops = set()
+for op in program.blocks[0].ops:
+    if op.type not in ops:
+        unsupported_ops.add(op.type)
+
+nums = len(unsupported_ops)
+if len(unsupported_ops) > 0:
+    print("========= {} OPs are not supported in Paddle-Lite=========".format(
+        nums))
+    for op in unsupported_ops:
+        print("========= {} ========".format(op))
+else:
+    print("\n========== Good News! ========")
+    print("Good! All ops in this model are supported in Paddle-Lite!\n")

tools/compile.sh

+#!/bin/bash
+
+#function:
+#   script used to generate caffe_pb2.py from caffe.proto using protoc
+#
+
+PROTOC=`which protoc`
+if [[ -z $PROTOC ]];then
+    echo "not found protoc, you should first install it following this[https://github.com/google/protobuf/releases]"
+    exit 1
+fi
+
+WORK_ROOT=$1
+PY_NAME="$WORK_ROOT/caffe_pb2.py"
+$PROTOC --proto_path=$WORK_ROOT --python_out=$WORK_ROOT $WORK_ROOT/caffe.proto
+ret=$?
+
+if [ -e "$PY_NAME" ];then
+    echo "succeed to generate [$PY_NAME]"
+    exit 0
+else
+    echo "failed to generate [$PY_NAME]"
+fi
+exit $ret

tools/merge_params.py

+import paddle.fluid as fluid
+import sys
+
+model_dir = sys.argv[1]
+new_model_dir = sys.argv[2]
+exe = fluid.Executor(fluid.CPUPlace())
+[inference_program, feed_target_names,
+ fetch_targets] = fluid.io.load_inference_model(dirname=model_dir, executor=exe)
+
+print(feed_target_names)
+fluid.io.save_inference_model(dirname=new_model_dir,
+                              feeded_var_names=feed_target_names,
+                              target_vars=fetch_targets,
+                              executor=exe,
+                              main_program=inference_program,
+                              params_filename="__params__")

x2paddle/__init__.py

-__version__ = "0.4.1"
+__version__ = "0.4.5"

x2paddle/convert.py

@@ -23,7 +23,7 @@ def arg_parser():
                         "-m",
                         type=_text_type,
                         default=None,
-                        help="model file path")
+                        help="define model file path for tensorflow or onnx")
     parser.add_argument("--prototxt",
                         "-p",
                         type=_text_type,
@@ -39,29 +39,47 @@ def arg_parser():
                         type=_text_type,
                         default=None,
                         help="path to save translated model")
-    parser.add_argument("--framework",
-                        "-f",
-                        type=_text_type,
-                        default=None,
-                        help="define which deeplearning framework")
+    parser.add_argument(
+        "--framework",
+        "-f",
+        type=_text_type,
+        default=None,
+        help="define which deeplearning framework(tensorflow/caffe/onnx)")
     parser.add_argument(
         "--caffe_proto",
         "-c",
         type=_text_type,
         default=None,
-        help="the .py file compiled by caffe proto file of caffe model")
+        help="optional: the .py file compiled by caffe proto file of caffe model"
+    )
     parser.add_argument("--version",
                         "-v",
                         action="store_true",
                         default=False,
                         help="get version of x2paddle")
+    parser.add_argument(
+        "--without_data_format_optimization",
+        "-wo",
+        action="store_true",
+        default=False,
+        help="tf model conversion without data format optimization")
+    parser.add_argument("--define_input_shape",
+                        "-d",
+                        action="store_true",
+                        default=False,
+                        help="define input shape for tf model")
 
     return parser
 
 
-def tf2paddle(model_path, save_dir):
+def tf2paddle(model_path,
+              save_dir,
+              without_data_format_optimization=False,
+              define_input_shape=False):
     # check tensorflow installation and version
     try:
+        import os
+        os.environ["TF_CPP_MIN_LOG_LEVEL"] = '3'
         import tensorflow as tf
         version = tf.__version__
         if version >= '2.0.0' or version < '1.0.0':
@@ -75,17 +93,30 @@ def tf2paddle(model_path, save_dir):
 
     from x2paddle.decoder.tf_decoder import TFDecoder
     from x2paddle.op_mapper.tf_op_mapper import TFOpMapper
+    from x2paddle.op_mapper.tf_op_mapper_nhwc import TFOpMapperNHWC
     from x2paddle.optimizer.tf_optimizer import TFOptimizer
 
     print("Now translating model from tensorflow to paddle.")
-    model = TFDecoder(model_path)
-    mapper = TFOpMapper(model)
-    optimizer = TFOptimizer(mapper)
-    # neccesary optimization
-    optimizer.delete_redundance_code()
-    # optimizer below is experimental
-    optimizer.merge_activation()
-    optimizer.merge_bias()
+    model = TFDecoder(model_path, define_input_shape=define_input_shape)
+    if not without_data_format_optimization:
+        mapper = TFOpMapper(model)
+        optimizer = TFOptimizer(mapper)
+        # neccesary optimization
+        optimizer.delete_redundance_code()
+        # optimizer below is experimental
+        optimizer.merge_activation()
+        optimizer.merge_bias()
+        optimizer.merge_batch_norm()
+        optimizer.merge_prelu()
+    else:
+        mapper = TFOpMapperNHWC(model)
+        optimizer = TFOptimizer(mapper)
+        optimizer.delete_redundance_code()
+        optimizer.strip_graph()
+        optimizer.merge_activation()
+        optimizer.merge_bias()
+        optimizer.make_nchw_input_output()
+        optimizer.remove_transpose()
     mapper.save_inference_model(save_dir)
 
 
@@ -95,10 +126,17 @@ def tf2paddle(model_path, save_dir):
 def caffe2paddle(proto, weight, save_dir, caffe_proto):
     from x2paddle.decoder.caffe_decoder import CaffeDecoder
     from x2paddle.op_mapper.caffe_op_mapper import CaffeOpMapper
-
+    from x2paddle.optimizer.caffe_optimizer import CaffeOptimizer
+    import google.protobuf as gpb
+    ver_str = gpb.__version__.replace('.', '')
+    ver_int = int(ver_str[0:2])
+    assert ver_int >= 36, 'The version of protobuf must be larger than 3.6.0!'
     print("Now translating model from caffe to paddle.")
     model = CaffeDecoder(proto, weight, caffe_proto)
     mapper = CaffeOpMapper(model)
+    optimizer = CaffeOptimizer(mapper)
+    optimizer.merge_bn_scale()
+    optimizer.merge_op_activation()
     mapper.save_inference_model(save_dir)
 
 
@@ -131,15 +169,22 @@ def onnx2paddle(model_path, save_dir):
 def main():
     if len(sys.argv) < 2:
         print("Use \"x2paddle -h\" to print the help information")
+        print("For more information, please follow our github repo below:)")
+        print("\nGithub: https://github.com/PaddlePaddle/X2Paddle.git\n")
         return
 
     parser = arg_parser()
     args = parser.parse_args()
 
     if args.version:
-        print("x2paddle-{} with python>=3.5\n".format(x2paddle.__version__))
+        import x2paddle
+        print("x2paddle-{} with python>=3.5, paddlepaddle>=1.5.0\n".format(
+            x2paddle.__version__))
         return
 
+    assert args.framework is not None, "--framework is not defined(support tensorflow/caffe/onnx)"
+    assert args.save_dir is not None, "--save_dir is not defined"
+
     try:
         import paddle
         v0, v1, v2 = paddle.__version__.split('.')
@@ -148,12 +193,17 @@ def main():
             return
     except:
         print("paddlepaddle not installed, use \"pip install paddlepaddle\"")
-    assert args.framework is not None, "--from is not defined(tensorflow/caffe)"
-    assert args.save_dir is not None, "--save_dir is not defined"
 
     if args.framework == "tensorflow":
         assert args.model is not None, "--model should be defined while translating tensorflow model"
-        tf2paddle(args.model, args.save_dir)
+        without_data_format_optimization = False
+        define_input_shape = False
+        if args.without_data_format_optimization:
+            without_data_format_optimization = True
+        if args.define_input_shape:
+            define_input_shape = True
+        tf2paddle(args.model, args.save_dir, without_data_format_optimization,
+                  define_input_shape)
 
     elif args.framework == "caffe":
         assert args.prototxt is not None and args.weight is not None, "--prototxt and --weight should be defined while translating caffe model"

x2paddle/core/fluid_code.py

@@ -14,6 +14,7 @@
 
 from x2paddle.core.graph import GraphNode
 import collections
+from x2paddle.core.util import *
 
 
 class Layer(object):
@@ -63,11 +64,8 @@ class Layer(object):
                     else:
                         layer_code = layer_code + key + "={}, ".format(
                             input.layer_name)
-                elif isinstance(input, str):
-                    layer_code = layer_code + key + "={}, ".format(input)
                 else:
-                    raise Exception(
-                        "Element of inputs should GraphNode or String")
+                    layer_code = layer_code + key + "={}, ".format(input)
         elif isinstance(self.inputs, GraphNode):
             if hasattr(self.inputs, "index"):
                 layer_code += (self.inputs.layer_name +
@@ -81,6 +79,8 @@ class Layer(object):
 
         param_attr = collections.OrderedDict(self.param_attr)
         for key, value in param_attr.items():
+            if '\n' in str(value):
+                value = string(str(value).replace('\n', ','))
             layer_code = layer_code + key + "={}, ".format(value)
         layer_code = layer_code.strip(", ")
 

x2paddle/core/graph.py

@@ -85,7 +85,7 @@ class Graph(object):
                 node.index = int(idx)
                 return node
             else:
-                raise Exception("Graph doesn't have node [%s]." % name)
+                return None
         else:
             if copy:
                 node = cp.copy(self.node_map[name])

x2paddle/core/op_mapper.py

@@ -11,6 +11,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import paddle.fluid as fluid
 from paddle.fluid.proto import framework_pb2
 from x2paddle.core.util import *
 import inspect
@@ -46,6 +47,28 @@ def export_paddle_param(param, param_name, dir):
     fp.close()
 
 
+# This func will copy to generate code file
+def run_net(param_dir="./"):
+    import os
+    inputs, outputs = x2paddle_net()
+    for i, out in enumerate(outputs):
+        if isinstance(out, list):
+            for out_part in out:
+                outputs.append(out_part)
+            del outputs[i]
+    exe = fluid.Executor(fluid.CPUPlace())
+    exe.run(fluid.default_startup_program())
+
+    def if_exist(var):
+        b = os.path.exists(os.path.join(param_dir, var.name))
+        return b
+
+    fluid.io.load_vars(exe,
+                       param_dir,
+                       fluid.default_main_program(),
+                       predicate=if_exist)
+
+
 class OpMapper(object):
     def __init__(self):
         self.paddle_codes = ""
@@ -97,6 +120,11 @@ class OpMapper(object):
         import model
         try:
             inputs, outputs = model.x2paddle_net()
+            for i, out in enumerate(outputs):
+                if isinstance(out, list):
+                    for out_part in out:
+                        outputs.append(out_part)
+                    del outputs[i]
             input_names = [input.name for input in inputs]
             exe = fluid.Executor(fluid.CPUPlace())
             exe.run(fluid.default_startup_program())
@@ -143,6 +171,8 @@ class OpMapper(object):
         for i in range(len(self.graph.topo_sort)):
             node_name = self.graph.topo_sort[i]
             node = self.graph.get_node(node_name)
+            if node is None:
+                continue
             if len(node.fluid_code.layers) == 0:
                 continue
             self.add_codes(node.fluid_code.gen_codes(), 1)

x2paddle/core/util.py

@@ -11,8 +11,6 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
-import paddle.fluid as fluid
 import numpy
 import math
 import os
@@ -20,20 +18,3 @@ import os
 
 def string(param):
     return "\'{}\'".format(param)
-
-
-# This func will copy to generate code file
-def run_net(param_dir="./"):
-    import os
-    inputs, outputs = x2paddle_net()
-    exe = fluid.Executor(fluid.CPUPlace())
-    exe.run(fluid.default_startup_program())
-
-    def if_exist(var):
-        b = os.path.exists(os.path.join(param_dir, var.name))
-        return b
-
-    fluid.io.load_vars(exe,
-                       param_dir,
-                       fluid.default_main_program(),
-                       predicate=if_exist)

x2paddle/decoder/caffe_decoder.py

@@ -47,7 +47,7 @@ class CaffeResolver(object):
 
 
 class CaffeGraphNode(GraphNode):
-    def __init__(self, layer, layer_name=None):
+    def __init__(self, layer, type_str, layer_name=None):
         if layer_name is None:
             super(CaffeGraphNode,
                   self).__init__(layer,
@@ -56,28 +56,18 @@ class CaffeGraphNode(GraphNode):
             super(CaffeGraphNode,
                   self).__init__(layer,
                                  layer_name.replace('/', '_').replace('-', '_'))
-        self.layer_type = layer.type
+        self.layer_type = type_str
         self.fluid_code = FluidCode()
         self.data = None
 
     def set_params(self, params):
         self.data = params
 
-    def set_output_shape(self, input_shape, is_input=True):
-        func_name = 'shape_' + self.layer_type.lower()
-        if is_input:
-            self.output_shape = getattr(caffe_shape, func_name)(self.layer,
-                                                                input_shape)
-        else:
-            self.output_shape = input_shape
-
-    def set_input_shape(self, input_shape):
-        self.input_shape = input_shape
-
 
 class CaffeGraph(Graph):
-    def __init__(self, model, params):
+    def __init__(self, model, params, caffe_pb):
         self.params = params
+        self.caffe_pb = caffe_pb
         super(CaffeGraph, self).__init__(model)
 
     def filter_layers(self, layers):
@@ -86,6 +76,9 @@ class CaffeGraph(Graph):
         filtered_layer_names = set()
         filtered_layers = []
         for layer in layers:
+            if hasattr(layer, 'input'):
+                continue
+            type_str = self.get_layer_type(layer)
             phase = 'test'
             if len(layer.include):
                 phase = phase_map[layer.include[0].phase]
@@ -96,7 +89,7 @@ class CaffeGraph(Graph):
             # test-time networks. These are just ignored. We'll
             # filter them out here.
             if (not exclude) and (phase == 'test'):
-                exclude = (layer.type == 'Dropout')
+                exclude = (type_str == 'Dropout')
             if not exclude:
                 filtered_layers.append(layer)
                 # Guard against dupes.
@@ -106,13 +99,85 @@ class CaffeGraph(Graph):
                 print('The filter layer:' + layer.name)
         return filtered_layers
 
+    def generate_input_layer(self, dims, index):
+        dim_str = ''
+        for dim in dims:
+            dim_str += 'dim: {}\n'.format(str(dim))
+        input_str = 'layer {\n'
+        input_str += 'name: \"{}\"\n '.format(str(self.model.input[index]))
+        input_str += 'type: "Input"\n'
+        input_str += 'top: \"{}\"\n'.format(str(self.model.input[index]))
+        input_str += 'input_param {\n'
+        input_str += 'shape {\n'
+        input_str += dim_str
+        input_str += '}}}'
+        input_str = str.encode(input_str)
+        net = self.caffe_pb.NetParameter()
+        text_format.Merge(input_str, net)
+        return net.layers or net.layer
+
+    def input2layers(self, input_layers=[]):
+        inputs_num = len(self.model.input)
+        if inputs_num != 0:
+            input_dims_num = len(self.model.input_dim)
+            if input_dims_num != 0:
+                if input_dims_num > 0 and input_dims_num != inputs_num * 4:
+                    raise Error('invalid input_dim[%d] param in prototxt' %
+                                (input_dims_num))
+                for i in range(inputs_num):
+                    dims = self.model.input_dim[i * 4:(i + 1) * 4]
+                    l = self.generate_input_layer(dims, i)
+                    input_layers.append(l[0])
+            else:
+                for i in range(inputs_num):
+                    dims = self.model.input_shape[i].dim[0:4]
+                    l = self.generate_input_layer(dims, i)
+                    input_layers.append(l[0])
+
+    def transform_input_layers(self, layers, input_layers=[]):
+        for layer in layers:
+            if hasattr(layer, 'input'):
+                input_dims_num = len(layers.input_dim)
+                if input_dims_num > 0 and input_dims_num != 4:
+                    raise Error('invalid input_dim[%d] param in prototxt' %
+                                (input_dims_num))
+                dims = self.model.input_dim[0:4]
+                l = self.generate_input_layer(dims, i)
+                input_layers.append(l[0])
+
+    def get_layer_type(self, layer):
+        if isinstance(layer.type, int):
+            enum_values = self.caffe_pb._V1LAYERPARAMETER_LAYERTYPE.values
+            vals = [val for val in enum_values if val.number == layer.type]
+            part = vals[0].name.split('_')
+            part = [s.capitalize() for s in part]
+            type_str = ''
+            type_str = type_str.join(part)
+            if 'relu' in type_str.lower():
+                type_str = type_str.replace('elu', 'eLU')
+            elif type_str.lower() == 'lrn':
+                type_str = 'LRN'
+            return type_str
+        else:
+            return layer.type
+
     def build(self):
         layers = self.model.layers or self.model.layer
+
         layers = self.filter_layers(layers)
 
+        input_layers = []
+
+        self.input2layers(input_layers)
+        self.transform_input_layers(layers, input_layers)
+        layers = input_layers + layers
+
         top_layer = {}
         for layer in layers:
-            self.node_map[layer.name] = CaffeGraphNode(layer)
+            if hasattr(layer, 'input'):
+                continue
+            type_str = self.get_layer_type(layer)
+            self.node_map[layer.name] = CaffeGraphNode(layer, type_str)
             for in_name in layer.bottom:
                 if in_name in top_layer:
                     self.connect(top_layer[in_name][-1], layer.name)
@@ -125,7 +190,6 @@ class CaffeGraph(Graph):
                     top_layer[out_name] = [layer.name]
                 else:
                     top_layer[out_name].append(layer.name)
-
         for layer_name, data in self.params:
             if layer_name in self.node_map:
                 node = self.node_map[layer_name]
@@ -157,19 +221,26 @@ class CaffeDecoder(object):
         self.resolver = CaffeResolver(caffe_proto=caffe_proto)
         self.net = self.resolver.NetParameter()
         with open(proto_path, 'rb') as proto_file:
-            proto_str = self.old2new(proto_file)
+            proto_str = proto_file.read()
             text_format.Merge(proto_str, self.net)
 
         self.load_using_pb()
-        self.caffe_graph = CaffeGraph(self.net, self.params)
+
+        self.caffe_graph = CaffeGraph(self.net, self.params,
+                                      self.resolver.caffepb)
         self.caffe_graph.build()
 
     def load_using_pb(self):
         data = self.resolver.NetParameter()
+
         data.MergeFromString(open(self.model_path, 'rb').read())
         pair = lambda layer: (layer.name, self.normalize_pb_data(layer))
         layers = data.layers or data.layer
+        import time
+        start = time.time()
         self.params = [pair(layer) for layer in layers if layer.blobs]
+        end = time.time()
+        print('cost:', str(end - start))
 
     def normalize_pb_data(self, layer):
         transformed = []
@@ -182,72 +253,8 @@ class CaffeDecoder(object):
                 c_i = blob.channels
                 h = blob.height
                 w = blob.width
-            data = np.array(blob.data, dtype=np.float32).reshape(c_o, c_i, h, w)
+            data = np.asarray(list(blob.data),
+                              dtype=np.float32).reshape(c_o, c_i, h, w)
+
             transformed.append(data)
         return transformed
-
-    def old2new(self, proto_file):
-        part1_str = ''
-        part2_str = ''
-        part3_str = ''
-        is_input = False
-        dims = []
-        line = proto_file.readline()
-        print('Check if it is a new style of caffe...')
-        while line:
-            l_str = bytes.decode(line)
-            if l_str.replace(' ', '').startswith('input:'):
-                part2_str += 'layer {\n'
-                part2_str += (
-                    '  name: ' +
-                    l_str.strip().replace(' ', '').split('input:')[-1] + '\n')
-                part2_str += '  type: \"Input\"\n'
-                part2_str += (
-                    '  top: ' +
-                    l_str.strip().replace(' ', '').split('input:')[-1] + '\n')
-                is_input = True
-                line = proto_file.readline()
-                continue
-            elif l_str.replace(' ', '').startswith('input_dim:'):
-                dims.append(
-                    int(l_str.strip().replace(' ', '').split('input_dim:')[-1]))
-                if len(dims) == 4:
-                    part2_str += '  input_param { shape: { dim: ' + str(dims[0]) + \
-                                               ' dim: ' + str(dims[1]) + \
-                                               ' dim: ' + str(dims[2]) + \
-                                               ' dim: ' + str(dims[3]) + ' } }\n'
-                    dims = []
-                    part2_str += '}\n'
-                line = proto_file.readline()
-                if bytes.decode(line).replace(' ', '').startswith('}'):
-                    line = proto_file.readline()
-                continue
-            elif l_str.replace(' ', '').startswith('input_shape'):
-                part2_str += l_str.replace('input_shape',
-                                           'input_param { shape: ')
-                l_str = bytes.decode(proto_file.readline())
-                while l_str:
-                    if '}' in l_str:
-                        part2_str += l_str + '\n}\n}'
-                        break
-                    else:
-                        part2_str += l_str
-                    l_str = bytes.decode(proto_file.readline())
-                line = proto_file.readline()
-                continue
-            if not is_input:
-                part1_str += bytes.decode(line)
-            else:
-                part3_str += bytes.decode(line)
-            line = proto_file.readline()
-        out = part1_str + part2_str + part3_str
-        layer_str = 'layer{'
-        part = out.split(layer_str)
-        if len(part) == 1:
-            layer_str = 'layer {'
-            part = out.split(layer_str)
-        for i in range(len(part)):
-            if part[i].strip().replace(' ', '') == '' or part[i].count(':') > 1:
-                continue
-            out = out.replace(layer_str + part[i], part[i].replace(' ', ''))
-        return str.encode(out)

x2paddle/decoder/tf_decoder.py

@@ -15,7 +15,6 @@
 from x2paddle.core.graph import GraphNode, Graph
 from x2paddle.core.fluid_code import FluidCode
 from tensorflow.python.framework import tensor_util
-from tensorflow.python.platform import gfile
 from tensorflow.core.framework import attr_value_pb2
 import tensorflow as tf
 import copy as cp
@@ -39,7 +38,7 @@ class TFGraphNode(GraphNode):
         self.pd_data_format = "NCHW"
         self.fluid_code = FluidCode()
 
-        self.dtype_map = {1: "float32", 3: "int32", 4: "int8", 9: "int64"}
+        self.dtype_map = {1: "float32", 3: "int32", 4: "uint8", 9: "int64"}
 
     @property
     def out_shapes(self):
@@ -52,7 +51,11 @@ class TFGraphNode(GraphNode):
 
     @property
     def dtype(self):
-        dtype = self.layer.attr["dtype"].type
+        keys = ['dtype', 'Tidx', 'T', 'DstT']
+        for k in keys:
+            dtype = self.layer.attr[k].type
+            if dtype > 0:
+                break
         if dtype not in self.dtype_map:
             raise Exception("Dtype[{}] not in dtype_map".format(dtype))
         return self.dtype_map[dtype]
@@ -125,6 +128,8 @@ class TFGraph(Graph):
             items[0] = self.identity_map[items[0]]
         new_node_name = ":".join(items)
         node = super(TFGraph, self).get_node(new_node_name, copy)
+        if node is None:
+            return None
         if len(items) == 1 and node.layer_type in self.multi_out_ops:
             node.index = 0
         return node
@@ -134,7 +139,7 @@ class TFGraph(Graph):
             raise Exception("Node[{}] not in graph".format(node_name))
         inputs = self.node_map[node_name].inputs
         outputs = self.node_map[node_name].outputs
-        assert len(inputs) == 1
+        #        assert len(inputs) == 1
         input_node = self.node_map[inputs[0]]
         idx = input_node.outputs.index(node_name)
         del input_node.outputs[idx]
@@ -171,7 +176,7 @@ class TFGraph(Graph):
     def _remove_identity_node(self):
         identity_node = list()
         for node_name, node in self.node_map.items():
-            if node.layer_type == "Identity":
+            if node.layer_type == "Identity" or node.layer_type == "StopGradient":
                 identity_node.append(node_name)
 
         for node_name in identity_node:
@@ -198,18 +203,29 @@ class TFGraph(Graph):
 
 
 class TFDecoder(object):
-    def __init__(self, pb_model, data_format="NHWC"):
-        self.sess = tf.Session()
+    def __init__(self, pb_model, data_format="NHWC", define_input_shape=False):
+        try:
+            self.sess = tf.compat.v1.Session()
+        except:
+            self.sess = tf.Session()
         self.input_info = dict()
-        with gfile.FastGFile(pb_model, 'rb') as f:
-            graph_def = tf.GraphDef()
+        self.define_input_shape = define_input_shape
+        with open(pb_model, 'rb') as f:
+            try:
+                graph_def = tf.compat.v1.GraphDef()
+            except:
+                graph_def = tf.GraphDef()
             graph_def.ParseFromString(f.read())
             input_map = self._check_input_shape(graph_def)
             self._fix_output_shape(graph_def)
             self.sess.graph.as_default()
             tf.import_graph_def(graph_def, name='', input_map=input_map)
 
-        self.sess.run(tf.global_variables_initializer())
+        try:
+            initializer = tf.compat.v1.global_variables_initializer()
+        except:
+            initializer = tf.global_variables_initializer()
+        self.sess.run(initializer)
 
         self.tf_graph = TFGraph(
             self.sess.graph._as_graph_def(add_shapes=True)[0], data_format)
@@ -229,10 +245,14 @@ class TFDecoder(object):
             if layer.op != "Placeholder":
                 continue
             graph_node = TFGraphNode(layer)
-            dtype = graph_node.dtype
+            dtype = graph_node.layer.attr['dtype'].type
 
             need_define_shape = 0
-            if not graph_node.get_attr("shape"):
+            if self.define_input_shape:
+                need_define_shape = 3
+            elif graph_node.layer.attr[
+                    'shape'].shape.unknown_rank or not graph_node.get_attr(
+                        "shape"):
                 need_define_shape = 1
             else:
                 value = graph_node.layer.attr["shape"].shape
@@ -241,13 +261,21 @@ class TFDecoder(object):
                     need_define_shape = 2
 
             if need_define_shape > 0:
+                shape = None
+                if graph_node.get_attr("shape"):
+                    value = value = graph_node.layer.attr["shape"].shape
+                    shape = [dim.size for dim in value.dim]
                 if need_define_shape == 1:
                     print("Unknown shape for input tensor[tensor name: \"{}\"]".
                           format(layer.name))
-                else:
+                elif need_define_shape == 2:
                     print(
                         "\nShape[now is {}] for input tensor[tensor name: \"{}\"] not support yet"
                         .format(shape, layer.name))
+                else:
+                    print(
+                        "Define shape[now is {}] for input tensor[tensor name: \"{}\']"
+                        .format(shape, layer.name))
                 print(
                     "Use your keyboard type the shape of input tensor below :)")
 
@@ -264,12 +292,20 @@ class TFDecoder(object):
                     for dim in shape.strip().split(',')
                 ]
                 assert shape.count(None) <= 1, "Only one dimension can be None"
-                x2paddle_input = tf.placeholder(dtype=dtype,
-                                                shape=shape,
-                                                name="x2paddle_{}".format(
-                                                    layer.name))
+                try:
+                    x2paddle_input = tf.compat.v1.placeholder(
+                        dtype=dtype,
+                        shape=shape,
+                        name="x2paddle_{}".format(layer.name))
+                except:
+                    x2paddle_input = tf.placeholder(dtype=dtype,
+                                                    shape=shape,
+                                                    name="x2paddle_{}".format(
+                                                        layer.name))
+
                 input_map["{}:0".format(layer.name)] = x2paddle_input
-                shape[shape.index(None)] = -1
+                if shape.count(None) > 0:
+                    shape[shape.index(None)] = -1
                 self.input_info["x2paddle_{}".format(layer.name)] = (shape,
                                                                      dtype)
             else:
@@ -282,7 +318,6 @@ class TFDecoder(object):
     # trick method
     # should be removed after PaddlePaddle V1.6 been released
     def infer_tensor(self, graph_node):
-        print("========== Use infer_tensor for tensor: ", graph_node.layer.name)
         if hasattr(graph_node, "index"):
             tensor_name = graph_node.layer.name + ":{}".format(graph_node.index)
         else:
@@ -298,8 +333,6 @@ class TFDecoder(object):
         return self.sess.run([output_tensor], feed)[0]
 
     def infer_shape_tensor(self, graph_node, out_shape=None):
-        print("========== Use infer_shape_tensor for tensor: ",
-              graph_node.layer.name)
         if hasattr(graph_node, "index"):
             tensor_name = graph_node.layer.name + ":{}".format(graph_node.index)
         else:
@@ -341,3 +374,38 @@ class TFDecoder(object):
             return results[0].tolist()
         else:
             raise Exception("Couldn't infer a stable shape shape tensor value")
+
+    def infer_tensor_shape(self, graph_node):
+        if hasattr(graph_node, "index"):
+            tensor_name = graph_node.layer.name + ":{}".format(graph_node.index)
+        else:
+            tensor_name = graph_node.layer.name + ":0"
+        feed = dict()
+        batch_size = [2, 3, 5]
+        shapes = list()
+        for b in batch_size:
+            for input_name, info in self.input_info.items():
+                (shape, dtype) = cp.deepcopy(info)
+                input_tensor = self.sess.graph.get_tensor_by_name(input_name +
+                                                                  ":0")
+                if shape.count(-1) > 0:
+                    shape[shape.index(-1)] = b
+                feed[input_tensor] = numpy.random.random_sample(shape)
+            output_tensor = self.sess.graph.get_tensor_by_name(tensor_name)
+            shape = self.sess.run([output_tensor], feed)[0].shape
+            shapes.append(numpy.array(shape))
+
+        compare01 = (shapes[0] == shapes[1])
+        compare12 = (shapes[1] == shapes[2])
+
+        if compare01.all() and compare12.all():
+            return shape[0].tolist()
+
+        if (compare01 == compare12).all():
+            index = numpy.argwhere(compare01 == False).flatten()
+            if index.shape[0] != 1:
+                raise Exception("There's not only one unstable dimension")
+            if index[0] != 0:
+                raise Exception("Batch size not in the first dimension")
+            shapes[0][0] = -1
+            return shapes[0].tolist()

x2paddle/op_mapper/caffe_custom_layer/convolutiondepthwise.py

@@ -19,20 +19,29 @@ def convolutiondepthwise_shape(input_shape,
     if isinstance(kernel_size, numbers.Number):
         [k_h, k_w] = [kernel_size] * 2
     elif len(kernel_size) > 0:
-        k_h = kernel_h if kernel_h else kernel_size[0]
-        k_w = kernel_w if kernel_w else kernel_size[len(kernel_size) - 1]
+        k_h = kernel_h if kernel_h > 0 else kernel_size[0]
+        k_w = kernel_w if kernel_w > 0 else kernel_size[len(kernel_size) - 1]
+    elif kernel_h > 0 or kernel_w > 0:
+        k_h = kernel_h
+        k_w = kernel_w
     [s_h, s_w] = [1, 1]
     if isinstance(stride, numbers.Number):
         [s_h, s_w] = [stride] * 2
     elif len(stride) > 0:
-        s_h = stride_h if stride_h else stride[0]
-        s_w = stride_w if stride_w else stride[len(stride) - 1]
+        s_h = stride_h if stride_h > 0 else stride[0]
+        s_w = stride_w if stride_w > 0 else stride[len(stride) - 1]
+    elif stride_h > 0 or stride_w > 0:
+        s_h = stride_h
+        s_w = stride_w
     [p_h, p_w] = [0, 0]
     if isinstance(pad, numbers.Number):
         [p_h, p_w] = [pad] * 2
     elif len(pad) > 0:
-        p_h = pad_h if pad_h else pad[0]
-        p_w = pad_w if pad_w else pad[len(pad) - 1]
+        p_h = pad_h if pad_h > 0 else pad[0]
+        p_w = pad_w if pad_w > 0 else pad[len(pad) - 1]
+    elif pad_h > 0 or pad_w > 0:
+        p_h = pad_h
+        p_w = pad_w
     dila_len = len(dilation)
     dila_h = 1
     dila_w = 1
@@ -74,20 +83,29 @@ def convolutiondepthwise_layer(inputs,
     if isinstance(kernel_size, numbers.Number):
         [k_h, k_w] = [kernel_size] * 2
     elif len(kernel_size) > 0:
-        k_h = kernel_h if kernel_h else kernel_size[0]
-        k_w = kernel_w if kernel_w else kernel_size[len(kernel_size) - 1]
+        k_h = kernel_h if kernel_h > 0 else kernel_size[0]
+        k_w = kernel_w if kernel_w > 0 else kernel_size[len(kernel_size) - 1]
+    elif kernel_h > 0 or kernel_w > 0:
+        k_h = kernel_h
+        k_w = kernel_w
     [s_h, s_w] = [1, 1]
     if isinstance(stride, numbers.Number):
         [s_h, s_w] = [stride] * 2
     elif len(stride) > 0:
-        s_h = stride_h if stride_h else stride[0]
-        s_w = stride_w if stride_w else stride[len(stride) - 1]
+        s_h = stride_h if stride_h > 0 else stride[0]
+        s_w = stride_w if stride_w > 0 else stride[len(stride) - 1]
+    elif stride_h > 0 or stride_w > 0:
+        s_h = stride_h
+        s_w = stride_w
     [p_h, p_w] = [0, 0]
     if isinstance(pad, numbers.Number):
         [p_h, p_w] = [pad] * 2
     elif len(pad) > 0:
-        p_h = pad_h if pad_h else pad[0]
-        p_w = pad_w if pad_w else pad[len(pad) - 1]
+        p_h = pad_h if pad_h > 0 else pad[0]
+        p_w = pad_w if pad_w > 0 else pad[len(pad) - 1]
+    elif pad_h > 0 or pad_w > 0:
+        p_h = pad_h
+        p_w = pad_w
     input = inputs[0]
     dila_len = len(dilation)
     dila_h = 1

x2paddle/op_mapper/caffe_custom_layer/detectionoutput.py

@@ -14,6 +14,18 @@ def detectionoutput_layer(inputs,
                           confidence_threshold=0.1,
                           input_shape=None,
                           name=None):
+    nms_param_str = nms_param
+    nms_param = {}
+    part = nms_param_str.split(',')
+    for s in part:
+        if s == '':
+            break
+        else:
+            name, obj = s.split(': ')
+            if name == 'top_k':
+                nms_param[name] = int(obj)
+            else:
+                nms_param[name] = float(obj)
     if nms_param is None:
         nms_param = {"nms_threshold": 0.3, "top_k": 10, "eta": 1.0}
     mbox_conf_flatten = inputs[1]
@@ -24,20 +36,21 @@ def detectionoutput_layer(inputs,
     pb = fluid.layers.reshape(x=pb, shape=[-1, 4])
     pbv = fluid.layers.reshape(x=pbv, shape=[-1, 4])
     mbox_loc = inputs[0]
-    mbox_loc = fluid.layers.reshape(x=mbox_loc,
-                                    shape=[-1, mbox_conf_flatten.shape[1], 4])
+    mbox_loc = fluid.layers.reshape(x=mbox_loc, shape=[-1, pb.shape[0], 4])
+    mbox_conf_flatten = fluid.layers.reshape(x=mbox_conf_flatten,
+                                             shape=[0, pb.shape[0], -1])
 
     default = {"nms_threshold": 0.3, "top_k": 10, "eta": 1.0}
     fields = ['eta', 'top_k', 'nms_threshold']
     for f in default.keys():
-        if not nms_param.has_key(f):
+        if f not in nms_param:
             nms_param[f] = default[f]
     out = fluid.layers.detection_output(
         scores=mbox_conf_flatten,
         loc=mbox_loc,
         prior_box=pb,
         prior_box_var=pbv,
-        background_label=background_label,
+        background_label=background_label_id,
         nms_threshold=nms_param["nms_threshold"],
         nms_top_k=nms_param["top_k"],
         keep_top_k=keep_top_k,

x2paddle/op_mapper/caffe_custom_layer/normalize.py

@@ -18,6 +18,8 @@ def normalize_layer(inputs,
         shape=[1] if channel_shared else [input_shape[0][1]],
         dtype=input.dtype,
         attr=name + '_scale')
+    scale_param = fluid.layers.reshape(x=scale_param, \
+                  shape=[1] if channel_shared else [input_shape[0][1]])
     out = fluid.layers.elementwise_mul(x=l2_norm,
                                        y=scale_param,
                                        axis=-1 if channel_shared else 1)

x2paddle/op_mapper/caffe_custom_layer/priorbox.py

@@ -3,7 +3,7 @@ from x2paddle.core.util import *
 
 
 def priorbox_shape(input_shape, max_size=None, aspect_ratio=None):
-    fc_shape = input_shapes[0]
+    fc_shape = input_shape[0]
     N = 1
     if not max_size == None:
         N += 1
@@ -18,26 +18,27 @@ def priorbox_layer(inputs,
                    step=0.0,
                    offset=0.5,
                    min_size=None,
-                   max_size=None,
+                   max_size=[],
                    aspect_ratio=[1.0],
                    flip=False,
                    clip=False,
                    variance=[0.1, 0.1, 0.2, 0.2],
                    input_shape=None,
                    name=None):
-    input = input_shape[0]
-    image = input_shape[1]
+    input = inputs[0]
+    image = inputs[1]
     steps = tuple(step) if type(step) is list or type(step) is tuple else (step,
                                                                            step)
+
     box, variance_ = fluid.layers.prior_box(input,
                                             image,
-                                            min_sizes=list(min_size),
-                                            max_sizes=list(max_size),
-                                            aspect_ratios=list(aspect_ratio),
-                                            variance=list(variance),
+                                            min_sizes=min_size,
+                                            max_sizes=max_size,
+                                            aspect_ratios=aspect_ratio,
+                                            variance=variance,
                                             flip=flip,
                                             clip=clip,
-                                            steps=step,
+                                            steps=steps,
                                             offset=offset,
                                             name=name,
                                             min_max_aspect_ratios_order=True)

x2paddle/op_mapper/caffe_custom_layer/shufflechannel.py

@@ -9,12 +9,12 @@ def shufflechannel_shape(input_shape):
 def shufflechannel_layer(inputs, group=None, input_shape=None, name=None):
     input = inputs[0]
     c_fm = fluid.layers.split(input, num_or_sections=input_shape[0][1], dim=1)
-    size = int(input_shape[0][1]/group)
+    size = int(input_shape[0][1] / group)
     new_c_fm = []
     for i in range(size):
         for j in range(group):
             new_c_fm.append(c_fm[j * size + i])
-    out = fluid.layers.concat(new_c_fm, axis = 1)
+    out = fluid.layers.concat(new_c_fm, axis=1)
     return out
 
 

x2paddle/op_mapper/caffe_shape.py

@@ -13,104 +13,68 @@
 # limitations under the License.
 
 import math
-
-
-def get_params_w_h(params):
+import numbers
+from functools import reduce
+
+
+def get_kernel_parameters(params):
+    [k_h, k_w] = [1, 1]
+    if isinstance(params.kernel_size, numbers.Number):
+        [k_h, k_w] = [params.kernel_size] * 2
+    elif len(params.kernel_size) > 0:
+        k_h = params.kernel_h if params.kernel_h > 0 else params.kernel_size[0]
+        k_w = params.kernel_w if params.kernel_w > 0 else params.kernel_size[
+            len(params.kernel_size) - 1]
+    elif params.kernel_h > 0 or params.kernel_w > 0:
+        k_h = params.kernel_h
+        k_w = params.kernel_w
+    [s_h, s_w] = [1, 1]
+    if isinstance(params.stride, numbers.Number):
+        [s_h, s_w] = [params.stride] * 2
+    elif len(params.stride) > 0:
+        s_h = params.stride_h if params.stride_h > 0 else params.stride[0]
+        s_w = params.stride_w if params.stride_w > 0 else params.stride[
+            len(params.stride) - 1]
+    elif params.stride_h > 0 or params.stride_w > 0:
+        s_h = params.stride_h
+        s_w = params.stride_w
+    [p_h, p_w] = [0, 0]
+    if isinstance(params.pad, numbers.Number):
+        [p_h, p_w] = [params.pad] * 2
+    elif len(params.pad) > 0:
+        p_h = params.pad_h if params.pad_h > 0 else params.pad[0]
+        p_w = params.pad_w if params.pad_w > 0 else params.pad[len(params.pad) -
+                                                               1]
+    elif params.pad_h > 0 or params.pad_w > 0:
+        p_h = params.pad_h
+        p_w = params.pad_w
+    dila_h = dila_w = 1
     if hasattr(params, 'dilation'):
-        if len(params.dilation) == 0:
-            dila_h = 1
-            dila_w = 1
-        elif len(params.dilation) == 1:
-            dila_h = params.dilation[0]
-            dila_w = params.dilation[0]
-        else:
+        dila_len = len(params.dilation)
+        if dila_len == 2:
             dila_h = params.dilation[0]
             dila_w = params.dilation[1]
-    else:
-        dila_h = 1
-        dila_w = 1
-
-    if not isinstance(getattr(params, 'pad'), int):
-        if len(params.pad) == 0:
-            pad_h = 0
-            pad_w = 0
-        elif len(params.pad) == 1:
-            pad_h = params.pad[0]
-            pad_w = params.pad[0]
-        else:
-            pad_h, pad_w, = params.pad[0]
-            pad_w = params.pad[1]
-        if params.pad_h != 0 or params.pad_w != 0:
-            pad_h = params.pad_h
-            pad_w = params.pad_w
-    else:
-        if params.pad_h != 0 or params.pad_w != 0:
-            pad_h = params.pad_h
-            pad_w = params.pad_w
-        else:
-            pad_h = getattr(params, 'pad')
-            pad_w = getattr(params, 'pad')
-
-    if not isinstance(getattr(params, 'kernel_size'), int):
-        if len(params.kernel_size) == 0:
-            kernel_h = 1
-            kernel_w = 1
-        elif len(params.kernel_size) == 1:
-            kernel_h = params.kernel_size[0]
-            kernel_w = params.kernel_size[0]
+        elif dila_len == 1:
+            dila_h = dila_w = params.dilation[0]
         else:
-            kernel_h = params.kernel_size[0]
-            kernel_w = params.kernel_size[1]
-        if params.kernel_h != 0 or params.kernel_w != 0:
-            kernel_h = params.kernel_h
-            kernel_w = params.kernel_w
-    else:
-        if params.kernel_h != 0 or params.kernel_w != 0:
-            kernel_h = params.kernel_h
-            kernel_w = params.kernel_w
-        else:
-            kernel_h = getattr(params, 'kernel_size')
-            kernel_w = getattr(params, 'kernel_size')
-    if not isinstance(getattr(params, 'stride'), int):
-        if len(params.stride) == 0:
-            stride_h = 1
-            stride_w = 1
-        elif len(params.stride) == 1:
-            stride_h = params.stride[0]
-            stride_w = params.stride[0]
-        else:
-            stride_h = params.stride[0]
-            stride_w = params.stride[1]
-        if params.stride_h != 0 or params.stride_w != 0:
-            stride_h = params.stride_h
-            stride_w = params.stride_w
-    else:
-        if params.stride_h != 0 or params.stride_w != 0:
-            stride_h = params.stride_h
-            stride_w = params.stride_w
-        else:
-            stride_h = getattr(params, 'stride')
-            stride_w = getattr(params, 'stride')
-    return dila_h, dila_w, pad_h, pad_w, kernel_h, kernel_w, stride_h, stride_w
+            assert dila_len == 0, "invalid length[%s] of dilation in convolution" % (
+                dila_len)
+    return dila_h, dila_w, p_h, p_w, k_h, k_w, s_h, s_w
 
 
-def get_filter_output_shape(i_h, i_w, params, round_func):
-    dila_h, dila_w, pad_h, pad_w, kernel_h, kernel_w, stride_h, stride_w = get_params_w_h(
+def get_strided_kernel_output_shape(params, input_shape, round_func):
+    i_h = input_shape[2]
+    i_w = input_shape[3]
+    dila_h, dila_w, pad_h, pad_w, kernel_h, kernel_w, stride_h, stride_w = get_kernel_parameters(
         params)
     o_h = (i_h + 2 * pad_h - (dila_h *
                               (kernel_h - 1) + 1)) / float(stride_h) + 1
     o_w = (i_w + 2 * pad_w - (dila_w *
                               (kernel_w - 1) + 1)) / float(stride_w) + 1
-    return (int(round_func(o_h)), int(round_func(o_w)))
-
-
-def get_strided_kernel_output_shape(params, input_shape, round_func):
-
-    o_h, o_w = get_filter_output_shape(input_shape[2], input_shape[3], params,
-                                       round_func)
+    o_h = int(round_func(o_h))
+    o_w = int(round_func(o_w))
     has_c_o = hasattr(params, 'num_output')
     c = params.num_output if has_c_o else input_shape[1]
-
     return [[input_shape[0], c, o_h, o_w]]
 
 
@@ -120,11 +84,12 @@ def shape_convolution(layer, input_shape):
 
 
 def shape_deconvolution(layer, input_shape):
-    h_i = input_shape[2]
-    w_i = input_shape[3]
+
+    h_i = input_shape[0][2]
+    w_i = input_shape[0][3]
 
     params = layer.convolution_param
-    dila_h, dila_w, pad_h, pad_w, kernel_h, kernel_w, stride_h, stride_w = get_params_w_h(
+    dila_h, dila_w, pad_h, pad_w, kernel_h, kernel_w, stride_h, stride_w = get_kernel_parameters(
         params)
 
     h_o = (h_i - 1) * stride_h - 2 * pad_h + dila_h * (kernel_h - 1) + 1
@@ -170,13 +135,25 @@ def shape_input(layer, input_shape):
     return [list(layer.input_param.shape[0].dim)]
 
 
+def shape_memorydata(layer, input_shape):
+    params = layer.memory_data_param
+    shape = []
+    shape.append(int(params.batch_size))
+    shape.append(int(params.channels))
+    shape.append(int(params.height))
+    shape.append(int(params.width))
+    return [shape]
+
+
 def shape_concat(layer, input_shape):
     params = layer.concat_param
     axis = params.axis
     output_shape = None
     for shape in input_shape:
         if output_shape is None:
-            output_shape = shape
+            output_shape = []
+            for i in range(len(shape)):
+                output_shape.append(shape[i])
         else:
             output_shape[axis] += shape[axis]
     return [output_shape]
@@ -184,14 +161,28 @@ def shape_concat(layer, input_shape):
 
 def shape_slice(layer, input_shape):
     inshape = input_shape[0]
+
+    top_len = len(layer.top)
     params = layer.slice_param
     axis = params.axis
-    count = inshape[axis]
+    slice_dim = params.slice_dim
+    if slice_dim != 1 and axis == 1:
+        axis = slice_dim
     points = list(params.slice_point)
+    count = inshape[axis]
+    if len(points) == 0:
+        assert count % top_len == 0, "the parameter of Slice is wrong"
+        part = count / top_len
+        t = part
+        while t < count:
+            points.append(int(t))
+            t += part
     points = [0] + points + [count]
     output_shape = []
     for i in range(len(points)):
-        shape = inshape
+        shape = []
+        for ii in range(len(inshape)):
+            shape.append(inshape[ii])
         size = points[i + 1] - points[i]
         shape[axis] = size
         output_shape.append(shape)
@@ -238,8 +229,8 @@ def shape_reshape(layer, input_shape):
 
     inshape = input_shape[0]
     params = layer.reshape_param
-    axis = params.axis if hasattr(params, axis) else 0
-    num_axes = params.num_axes if hasattr(params, num_axes) else -1
+    axis = params.axis if hasattr(params, 'axis') else 0
+    num_axes = params.num_axes if hasattr(params, 'num_axes') else -1
     if inshape[0] == -1:
         inshape[0] = 1
     input_count = count(inshape)
@@ -262,14 +253,14 @@ def shape_reshape(layer, input_shape):
 
     num_axes_replaced = end_axis - start_axis
     num_axes_retained = input_num_axes - num_axes_replaced
-    num_new_axes = len(shape['dim'])
+    num_new_axes = len(list(params.shape.dim))
     outshape = []
 
     for i in range(start_axis):
         outshape.append(inshape[i])
 
     for i in range(num_new_axes):
-        outshape.append(shape['dim'][i])
+        outshape.append(params.shape.dim[i])
 
     for i in range(end_axis, input_num_axes):
         outshape.append(inshape[i])
@@ -281,7 +272,7 @@ def shape_reshape(layer, input_shape):
     copy_axes = []
     constant_count = 1
     for i in range(num_new_axes):
-        top_dim = shape['dim'][i]
+        top_dim = params.shape.dim[i]
         if top_dim == 0:
             copy_axes.append(i)
             copy_axis_index = start_axis + i
@@ -297,24 +288,20 @@ def shape_reshape(layer, input_shape):
         l = inshape[0:start_axis]
         if len(l) > 0:
             explicit_count *= count(l)
-
         l = inshape[end_axis:]
         if len(l) > 0:
             explicit_count *= count(l)
-
         for i in range(len(copy_axes)):
             explicit_count *= outshape[start_axis + copy_axes[i]]
-
         assert input_count % explicit_count == 0, "[Reshape]botom count[%d] "\
                 "must be divisible by product of the specified dimensions[%d] "\
                 % (input_count, explicit_count)
-        outshape[start_axis + inferred_axis] = input_count / explicit_count
+        outshape[start_axis + inferred_axis] = int(input_count / explicit_count)
 
     output_count = count(outshape)
     assert output_count == input_count, "[Reshape]output count[%d] must match input count[%d]" % (
         output_count, input_count)
-    if inshape[0] == -1:
-        outshape[0] = -1
+    outshape[0] = -1
     return [outshape]
 
 
@@ -345,18 +332,22 @@ def shape_crop(layer, input_shape):
 
 def shape_flatten(layer, input_shape):
     assert len(input_shape) == 1, "the number of flatten's inputs must be 1"
+    inshape = input_shape[0]
     params = layer.flatten_param
     start_axis = params.axis
     end_axis = params.end_axis
     if start_axis < 0:
-        start_axis += len(input_shape[0])
+        start_axis += len(inshape)
     if end_axis < 0:
-        end_axis += len(input_shape[0]) + 1
+        end_axis += len(inshape) + 1
     assert start_axis <= end_axis, 'invalid axis[%d] or end_axis[%d] params'\
             % (start_axis, end_axis)
-    output_shape = [0] * (start_axis - 0) + [
-        -1
-    ] + [0] * (len(input_shape[0]) - end_axis)
+    output_shape = inshape[0:start_axis]
+    if len(inshape[start_axis:end_axis]) != 0:
+        flat_sz = reduce(lambda a, b: a * b, inshape[start_axis:end_axis])
+        output_shape += [flat_sz]
+    output_shape += inshape[end_axis:len(inshape)]
+    output_shape[0] = -1
     return [output_shape]
 
 

x2paddle/optimizer/caffe_optimizer.py

+#   Copyright (c) 2019  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from x2paddle.decoder.caffe_decoder import CaffeGraph
+from x2paddle.core.util import *
+
+
+class CaffeOptimizer(object):
+    layers_with_act = ['Convolution', 'Deconvolution', 'InnerProduct']
+    activation_ops = ['ReLU', 'Sigmoid']
+
+    def __init__(self, mapper):
+        self.graph = mapper.graph
+
+    def merge_bn_scale(self):
+        for node_name in self.graph.topo_sort:
+            node = self.graph.get_node(node_name)
+            if node.layer_type == 'Scale':
+                parent_node = self.graph.get_bottom_node(node, idx=0)
+                if parent_node.layer_type == 'BatchNorm':
+                    is_delete_node = True if len(
+                        parent_node.outputs) == 1 else False
+                    parent_fluid_layer = parent_node.fluid_code.layers[0]
+                    input = parent_fluid_layer.inputs
+                    parent_param_attr = parent_fluid_layer.param_attr
+                    parent_param_attr['param_attr'] = string(node.layer_name +
+                                                             '_scale')
+                    parent_param_attr['bias_attr'] = string(node.layer_name +
+                                                            '_offset')
+                    if is_delete_node:
+                        parent_node.fluid_code.clear()
+                    node.fluid_code.clear()
+                    node.fluid_code.add_layer("batch_norm",
+                                              inputs=input,
+                                              output=node,
+                                              param_attr=parent_param_attr)
+
+    def merge_op_activation(self):
+        for node_name in self.graph.topo_sort:
+            node = self.graph.get_node(node_name)
+            if node.layer_type in self.activation_ops:
+                parent_node = self.graph.get_bottom_node(node, idx=0)
+                if parent_node.layer_type in self.layers_with_act:
+                    is_delete_node = True if len(
+                        parent_node.outputs) == 1 else False
+                    parent_fluid_layer = parent_node.fluid_code.layers[0]
+                    input = parent_fluid_layer.inputs
+                    parent_param_attr = parent_fluid_layer.param_attr
+                    parent_param_attr['act'] = string(node.layer_type.lower())
+                    op = parent_fluid_layer.op
+                    if is_delete_node:
+                        parent_node.fluid_code.clear()
+                    node.fluid_code.clear()
+                    node.fluid_code.add_layer(op,
+                                              inputs=input,
+                                              output=node,
+                                              param_attr=parent_param_attr)

x2paddle_model_zoo.md

-目前X2Paddle支持40+的TensorFlow OP，40+的Caffe Layer，覆盖了大部分CV分类模型常用的操作。我们在如下模型列表中测试了X2Paddle的转换
+# X2Paddle模型测试库
+> 目前X2Paddle支持40+的TensorFlow OP，40+的Caffe Layer，覆盖了大部分CV分类模型常用的操作。我们在如下模型列表中测试了X2Paddle的转换。
 
-# TensorFlow
+**注：** 受限于不同框架的差异，部分模型可能会存在目前无法转换的情况，如TensorFlow中包含控制流的模型，NLP模型等。对于CV常见的模型，如若您发现无法转换或转换失败，存在较大diff等问题，欢迎通过[ISSUE反馈](https://github.com/PaddlePaddle/X2Paddle/issues/new)的方式告知我们(模型名，代码实现或模型获取方式)，我们会及时跟进：）
 
-| 模型 | 代码 |
-|------|----------|
-| SqueezeNet | [code](https://github.com/tensorflow/tpu/blob/master/models/official/squeezenet/squeezenet_model.py)|
-| MobileNet_V1 | [code](https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.md) |
-| MobileNet_V2 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets/mobilenet) |
-| ShuffleNet | [code](https://github.com/TropComplique/shufflenet-v2-tensorflow) |
-| mNASNet | [code](https://github.com/tensorflow/tpu/tree/master/models/official/mnasnet) |
-| EfficientNet | [code](https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet) |
-| Inception_V4 | [code](https://github.com/tensorflow/models/blob/master/research/slim/nets/inception_v4.py) |
-| Inception_ResNet_V2 | [code](https://github.com/tensorflow/models/blob/master/research/slim/nets/inception_resnet_v2.py) |
-| VGG16 | [code](https://github.com/tensorflow/models/blob/master/research/slim/nets/vgg.py) |
-| ResNet_V1_101 | [code](https://github.com/tensorflow/models/blob/master/research/slim/nets/resnet_v1.py) |
-| ResNet_V2_101 | [code](https://github.com/tensorflow/models/blob/master/research/slim/nets/resnet_v2.py) |
+## TensorFlow
+
+| 模型 | 代码 | 备注 |
+|------|----------|------|
+| SqueezeNet | [code](https://github.com/tensorflow/tpu/blob/master/models/official/squeezenet/squeezenet_model.py)|-|
+| MobileNet_V1 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) |-|
+| MobileNet_V2 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) |-|
+| ShuffleNet | [code](https://github.com/TropComplique/shufflenet-v2-tensorflow) |-|
+| mNASNet | [code](https://github.com/tensorflow/tpu/tree/master/models/official/mnasnet) |-|
+| EfficientNet | [code](https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet) |-|
+| Inception_V4 | [code](https://github.com/tensorflow/models/blob/master/research/slim/nets/inception_v4.py) |-|
+| Inception_ResNet_V2 | [code](https://github.com/tensorflow/models/blob/master/research/slim/nets/inception_resnet_v2.py) |-|
+| VGG16 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) |-|
+| ResNet_V1_101 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) |-|
+| ResNet_V2_101 | [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) |-|
+| UNet | [code1](https://github.com/jakeret/tf_unet )/[code2](https://github.com/lyatdawn/Unet-Tensorflow) |-|
+|MTCNN | [code](https://github.com/AITTSMD/MTCNN-Tensorflow) |-|
+|YOLO-V3| [code](https://github.com/YunYang1994/tensorflow-yolov3) | 转换需要关闭NHWC->NCHW的优化，见[文档Q2](FAQ.md) |
+|Inception_V4| [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) | - |
+|Inception_ResNet_V2| [code](https://github.com/tensorflow/models/tree/master/research/slim/nets) | - |
 
-# Caffe
+## Caffe
 
 | 模型 | 代码 |
 |-------|--------|
 | SqueezeNet | [code](https://github.com/DeepScale/SqueezeNet/tree/master/SqueezeNet_v1.1) |
 | MobileNet_V1 | [code](https://github.com/shicai/MobileNet-Caffe) |
 | MobileNet_V2 | [code](https://github.com/shicai/MobileNet-Caffe) |
-| ShuffleNet | [code](https://github.com/miaow1988/ShuffleNet_V2_pytorch_caffe/releases/tag/v0.1.0) |
+| ShuffleNet_v2 | [code](https://github.com/miaow1988/ShuffleNet_V2_pytorch_caffe/releases/tag/v0.1.0) |
 | mNASNet | [code](https://github.com/LiJianfei06/MnasNet-caffe) |
 | MTCNN | [code](https://github.com/kpzhang93/MTCNN_face_detection_alignment/tree/master/code/codes/MTCNNv1/model) |
+| Mobilenet_SSD | [code](https://github.com/chuanqi305/MobileNet-SSD) |
+| ResNet18 | [code](https://github.com/HolmesShuan/ResNet-18-Caffemodel-on-ImageNet/blob/master/deploy.prototxt) |
+| ResNet50 | [code](https://github.com/soeaver/caffe-model/blob/master/cls/resnet/deploy_resnet50.prototxt) |
+| Unet | [code](https://github.com/jolibrain/deepdetect/blob/master/templates/caffe/unet/deploy.prototxt) |
+| VGGNet | [code](https://gist.github.com/ksimonyan/211839e770f7b538e2d8#file-vgg_ilsvrc_16_layers_deploy-prototxt) |
+| FaceDetection | [code](https://github.com/ShiqiYu/libfacedetection/blob/master/models/caffe/yufacedetectnet-open-v1.prototxt) |
+
+
+
+
 
-# ONNX
+## ONNX
+**注：** 部分模型来源于PyTorch，PyTorch的转换可参考[pytorch_to_onnx.md](pytorch_to_onnx.md)
 
 | 模型 | 来源 | operator version|
 |-------|--------|---------|
@@ -46,20 +65,3 @@
 | mNASNet | [pytorch(personal practice)](https://github.com/rwightman/gen-efficientnet-pytorch) |9|
 | EfficientNet | [pytorch(personal practice)](https://github.com/rwightman/gen-efficientnet-pytorch) |9|
 | SqueezeNet | [onnx official](https://s3.amazonaws.com/download.onnx/models/opset_9/squeezenet.tar.gz) |9|
-
-目前onnx2paddle主要支持onnx operator version 9；
-如何将torchvison或者个人开发者写的pytroch model转换成onnx model:
-```
-import torch
-import torchvision
-
-#根据不同模型调整输入的shape
-dummy_input = torch.randn(1, 3, 224, 224)
-
-#预训练后的pytorch model
-resnet18 = torchvision.models.resnet18(pretrained=True)
-
-#"resnet18.onnx"为onnx model的存储路径
-torch.onnx.export(resnet18, dummy_input, "resnet18.onnx",verbose=True)
-
-```