Merge pull request #13 from PaddlePaddle/develop

docs

Merge pull request #13 from PaddlePaddle/develop
docs
e19d3442 · SunAhong1993 · GitHub · 1df218f2 · e829670f · e19d3442
33 changed file
--- a/README.md
+++ b/README.md
@@ -5,12 +5,12 @@ X2Paddle支持将其余深度学习框架训练得到的模型，转换至Paddle
 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，可以在[OP-LIST](op_list.md)中查看目前X2Paddle支持的OP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。
+X2Paddle在多个主流的CV模型上，测试过TensorFlow/Caffe/ONNX/PyTorch模型的转换，可以在[X2Paddle-Model-Zoo](./docs/introduction/x2paddle_model_zoo.md)查看我们的模型测试列表，可以在[OP-LIST](./docs/introduction/op_list.md)中查看目前X2Paddle支持的OP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 == 2.7 | python >= 3.5  
-paddlepaddle >= 2.0.0  
+paddlepaddle 2.0-rc 或者 develop  
 **按需安装以下依赖**  
 tensorflow ： tensorflow == 1.14.0  
@@ -30,7 +30,7 @@ python setup.py install
 ### 安装方式二
 我们会定期更新pip源上的x2paddle版本
 ```
-pip install x2paddle --index https://pypi.Python.org/simple/
+pip install x2paddle==1.0.0rc0 --index https://pypi.Python.org/simple/
 ```
 ## 使用方法
 ### TensorFlow
@@ -47,7 +47,7 @@ x2paddle --framework=onnx --model=onnx_model.onnx --save_dir=pd_model
 ```
 ### PyTorch
-> PyTorch不支持命令行使用方式，详见[PyTorch2Paddle](pytorch2paddle.md)
+> PyTorch不支持命令行使用方式，详见[PyTorch2Paddle](./docs/user_guides/pytorch2paddle.md)
 ### Paddle2ONNX
 > Paddle2ONNX功能已迁移至新的github: https://github.com/PaddlePaddle/paddle2onnx, 欢迎大家去新的代码仓库查看详细介绍以及新功能。
@@ -62,15 +62,21 @@ x2paddle --framework=onnx --model=onnx_model.onnx --save_dir=pd_model
 |--save_dir | 指定转换后的模型保存目录路径 |
 |--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 |
-|--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) |
+|--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](./docs/user_guides/FAQ.md) |
 |--params_merge | **[可选]** 当指定该参数时，转换完成后，inference_model中的所有模型参数将合并保存为一个文件__params__ |
 ## 使用转换后的模型
+- 静态图：
 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  
+`model_with_code`中保存了模型参数，和转换后的python模型静态图代码。  
-`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)
+`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[paddle.static.load_inference_model](https://www.paddlepaddle.org.cn/documentation/docs/zh/2.0-rc/api/paddle/static/load_inference_model_cn.html#load-inference-model)。
+- 动态图：
+转换后的模型包括`model.pdparams`和`x2paddle_code.py`两个文件，以及`inference_model`一个目录。  
+`model.pdparams`中保存了模型参数。  
+`x2paddle_code.py`是转换后的python模型动态图代码。  
+`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[paddle.static.load_inference_model](https://www.paddlepaddle.org.cn/documentation/docs/zh/2.0-rc/api/paddle/static/load_inference_model_cn.html#load-inference-model)。
 ## 小工具
 X2Paddle提供了工具解决如下问题，详见[tools/README.md](tools/README.md)
@@ -78,11 +84,12 @@ X2Paddle提供了工具解决如下问题，详见[tools/README.md](tools/README
 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)  
+1. [X2Paddle使用过程中常见问题](./docs/user_guides/FAQ.md)  
-2. [如何导出TensorFlow的pb模型](export_tf_model.md)
+2. [如何导出TensorFlow的pb模型](./docs/user_guides/export_tf_model.md)
-3. [X2Paddle测试模型库](x2paddle_model_zoo.md)  
+3. [X2Paddle测试模型库](./docs/introduction/x2paddle_model_zoo.md)  
-4. [PyTorch模型导出为ONNX模型](pytorch_to_onnx.md)
+4. [X2Paddle支持的op列表](./docs/introduction/op_list.md) 
-5. [X2Paddle内置的Caffe自定义层](caffe_custom_layer.md)
+5. [PyTorch模型导出为ONNX模型](./docs/user_guides/pytorch2onnx.md)
+6. [X2Paddle添加内置的Caffe自定义层](./docs/user_guides/add_caffe_custom_layer.md)
 ## 更新历史
 2019.08.05  
@@ -91,7 +98,15 @@ X2Paddle提供了工具解决如下问题，详见[tools/README.md](tools/README
 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  
-**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。**
+2020.12.09
+1. 新增PyTorch2Paddle转换方式，转换得到Paddle动态图代码，并动转静获得inference_model。
+方式一：trace方式，转换后的代码有模块划分，每个模块的功能与PyTorch相同。
+方式二：script方式，转换后的代码按执行顺序逐行出现。
+2. 新增Caffe/ONNX/Tensorflow到Paddle动态图的转换。
+3. 新增TensorFlow op（14个）：Neg、Greater、FloorMod、LogicalAdd、Prd、Equal、Conv3D、Ceil、AddN、DivNoNan、Where、MirrorPad、Size、TopKv2
+4. 新增Optimizer模块，主要包括op融合、op消除功能，转换后的代码可读性更强，进行预测时耗时更短。
+**如果你需要之前版本的tensorflow2fluid/caffe2fluid/onnx2fluid，可以继续访问release-0.9分支，获取之前版本的代码使用。**
 ## Acknowledgements

--- a/caffe_custom_layer.md
+++ b/caffe_custom_layer.md
-目前，代码中已经提供了10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) |
-| ReLU6 | [code](https://github.com/chuanqi305/ssd/blob/ssd/src/caffe/layers/relu6_layer.cpp) |
-| Upsample | [code](https://github.com/eric612/MobileNet-YOLO/blob/master/src/caffe/layers/upsample_layer.cpp) |
--- a/docs/README.md
+++ b/docs/README.md
+# 一、 introduction
+1. op_list.md：当前转换各个框架支持的op。
+2. x2paddle_model_zoo.md：测试过的模型列表。
+# 二、 user_guides
+1. FQA.md：常见问题集合。
+2. add_caffe_custom_layer.md：添加caffe自定义Layer的方法，以及当前支持的自定义Layer列表。
+3. export_tf_model.md：导出本工具支持的TensorFlow模型。
+4. pytorch2onnx.md：将PyTorch导出为ONNX。
+5. pytorch2paddle.md：将PyTorch模型转换为Paddle模型。
--- a/op_list.md
+++ b/op_list.md
@@ -24,7 +24,11 @@
 | 57 | Sqrt             | 58 | Softplus       | 59 | Erf            | 60 | AddV2                 |
 | 61 | LessEqual        | 62 | BatchMatMul    | 63 | BatchMatMulV2  | 64 | ExpandDims            |
 | 65 | BatchToSpaceND   | 66 | SpaceToBatchND | 67 | OneHot         | 68 | Pow                   |
-| 69 | All              | 70 | GatherV2       | 71 | IteratorV2     |    |                       |
+| 69 | All              | 70 | GatherV2       | 71 | IteratorV2     | 72 | Neg |
+| 73 | Greater | 74 | FloorMod | 75 | LogicalAdd | 76 | Prod |
+| 77 | Equal | 78 | Conv3D | 79 | Ceil | 80 | AddN |
+| 81 | DivNoNan | 82 | Where | 83 | MirrorPad | 84 | Size |
+| 85 | TopKv2 |  |  |  |  |  |  |
 ## Caffe
@@ -58,3 +62,44 @@
 | 45 | Squeeze  | 46 | Equal     | 47 | Identity  | 48 | GlobalAveragePool  |
 | 49 | MaxPool  | 50 | Conv      | 51 | Gemm      | 52 | NonZero            |
 | 53 | Abs      | 54 | Floor     |
+## PyTorch
+Aten:
+| 序号 | OP | 序号 | OP | 序号 | OP | 序号 | OP |
+|------|------|------|------|------|------|------|------|
+| 1  | aten::abs | 2  | aten::adaptive_avg_pool2d | 3  | aten::addmm | 4  | aten::add |
+| 5  | aten::add\_ | 6  | aten::\_\_and\_\_ | 7  | aten::append | 8  | aten::arange |
+| 9  | aten::avg\_pool2d | 10 | aten::avg\_pool3d | 11 | aten::avg_pool1d | 12 | aten::batch_norm |
+| 13 | aten::cat | 14 | aten::chunk | 15 | aten::clamp | 16 | aten::\_\_contains\_\_ |
+| 17 | aten::constant\_pad\_nd | 18 | aten::contiguous | 19 | aten::conv2d | 20 | aten::\_convolution |
+| 21 | aten::conv_transpose2d | 22 | aten::cos | 23 | aten::cumsum | 24 | aten::detach |
+| 25 | aten::dict | 26 | aten::dim | 27 | aten::div\_ | 28 | aten::div   |
+| 29 | aten::dropout | 30 | aten::dropout_ | 31 | aten::embedding | 32 | aten::eq     |
+| 33 | aten::exp | 34 | aten::expand | 35 | aten::expand_as | 36 | aten::eye |
+| 37 | aten::feature_dropout | 38 | aten::flatten | 39 | aten::Float | 40 | aten::floor |
+| 41 | aten::floordiv | 42 | aten::floor_divide | 43 | aten::full_like | 44 | aten::gather |
+| 45 | aten::gelu | 46 | aten::\_\_getitem\_\_ | 47 | aten::gt | 48 | aten::hardtanh\_ |
+| 49 | aten::index\_select | 50 | aten::Int | 51 | aten::\_\_is\_\_ | 52 | aten::\_\_isnot\_\_ |
+| 53 | aten::layer\_norm | 54 | aten::le |55|aten::leaky\_relu\_|56|aten::len|
+| 57 | aten::log | 58 | aten::lt |59|aten::masked\_fil\l_|60|aten::masked\_fill|
+| 61 | aten::max | 62 | aten::max\_pool2d |63|aten::matmul|64|aten\_min|
+| 65 | aten::mean | 66 | aten::meshgrid |67|aten::mul|68|aten::mul\_|
+| 69 | aten::ne | 70 | aten::neg |71|aten::\_\_not\_\_|72|aten::ones|
+| 73 | aten::permute | 74 | aten::pow |75|aten::relu|76|aten::relu\_|
+| 77 | aten::relu6 | 78 | aten::repeat |79|aten::reshape|80|aten::rsub|
+| 81 | aten::ScalarImplicit | 82 | aten::select |83|aten::\_set\_item|84|aten::sigmoid|
+| 85 | aten::sin | 86 | aten::size |87|aten::slice|88|aten::softmax|
+| 89 | aten::softplus | 90 | aten::sqrt |91|aten::squeeze|92|aten::stack|
+| 93 | aten::sub | 94 | aten::t |95|aten::tanh|96|aten::split|
+| 97 | aten::transpose | 98 | aten::to |99|aten::type\_as|100|aten::unsqueeze|
+| 101 | aten::upsample\_bilinear2d | 102 | aten::values |103|aten::view|104|aten::warn|
+| 105 | aten::where | 106 | aten::zeros |107|aten::zeros\_like|||
+Prim:
+| 序号 | OP | 序号 | OP | 序号 | OP | 序号 | OP |
+|------|------|------|------|------|------|------|------|
+| 1  | prim::Constant | 2  | prim::data | 3  | prim::DictConstruct | 4  | prim::GetAttr |
+| 5  | prim::If | 6  | prim::ListConstruct | 7  | prim::ListUnpack     | 8  | prim::Loop     |
+| 9  | prim::min | 10 | prim::NumToTensor | 11 | prim::RaiseException | 12 | prim::requires\_grad |
+| 13 | prim::SetAttr | 14 | prim::shape | 15 | prim::TupleConstruct | 16 | prim::TupleUnpack |
+| 17 | prim::unchecked\_cast | 18 | prim::Uninitialized | ||||
--- a/x2paddle_model_zoo.md
+++ b/x2paddle_model_zoo.md
@@ -74,3 +74,25 @@
 |Ultra-Light-Fast-Generic-Face-Detector-1MB| [onnx_model](https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB/tree/master/models/onnx)|9 |
 |BERT| [pytorch(huggingface)](https://github.com/huggingface/transformers/blob/master/notebooks/04-onnx-export.ipynb)|11|转换时需指定input shape，见[文档Q3](FAQ.md)|
 |GPT2| [pytorch(huggingface)](https://github.com/huggingface/transformers/blob/master/notebooks/04-onnx-export.ipynb)|11|转换时需指定input shape，见[文档Q3](FAQ.md)|
+## PyTorch
+| 模型 | 代码 | 备注 |
+|------|----------|------|
+| AlexNet | [code](https://github.com/pytorch/vision/blob/master/torchvision/models/alexnet.py)|-|
+| MNasNet | [code](https://github.com/pytorch/vision/blob/master/torchvision/models/mnasnet.py) |-|
+| MobileNetV2 | [code](https://github.com/pytorch/vision/blob/master/torchvision/models/mobilenet.py) |-|
+| ResNet18 | [code](https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py) |-|
+| ShuffleNetV2 | [code](https://github.com/pytorch/vision/blob/master/torchvision/models/shufflenet.py) |-|
+| SqueezeNet | [code](https://github.com/pytorch/vision/blob/master/torchvision/models/squeezenet.py) |-|
+| VGG16 | [code](https://github.com/pytorch/vision/blob/master/torchvision/models/vgg.py) |-|
+| InceptionV3 | [code](https://github.com/pytorch/vision/blob/master/torchvision/models/inception.py) |-|
+| DeepLabv3_ResNet50 | [code](https://github.com/pytorch/vision/blob/master/torchvision/models/segmentation/deeplabv3.py) |-|
+| FCN_ResNet50 | [code](https://github.com/pytorch/vision/blob/master/torchvision/models/segmentation/fcn.py) |-|
+| CamembertForQuestionAnswering | [code](https://huggingface.co/transformers/model_doc/camembert.html) |只支持trace模式|
+| DPRContextEncoder | [code](https://huggingface.co/transformers/model_doc/dpr.html) |只支持trace模式|
+| ElectraModel | [code](https://huggingface.co/transformers/model_doc/electra.html ) |只支持trace模式|
+| FlaubertModel | [code](https://huggingface.co/transformers/model_doc/flaubert.html) |只支持trace模式|
+| Roberta| [code](https://huggingface.co/transformers/model_doc/roberta.html)  |只支持trace模式|
+| XLMRobertaForTokenClassification|[code](https://huggingface.co/transformers/model_doc/xlmroberta.html)  |只支持trace模式|
--- a/FAQ.md
+++ b/FAQ.md
@@ -23,3 +23,20 @@ A: 此提示为错误信息，表示该模型的转换需要固定的输入大
 > 1. 模型来源于PaddleX导出，可以在导出的命令中，指定--fixed_input_shape=[Height,Width]，详情可见：[PaddleX模型导出文档](https://github.com/PaddlePaddle/PaddleX/blob/develop/docs/deploy/export_model.md)。
 > 2. 模型来源于PaddleDetection导出，可以在导出模型的时候，指定 TestReader.inputs_def.image_shape=[Channel,Height,Width], 详情可见：[PaddleDetection模型导出文档](https://github.com/PaddlePaddle/PaddleDetection/blob/master/docs/advanced_tutorials/deploy/EXPORT_MODEL.md#设置导出模型的输入大小)。
 > 3. 模型来源于自己构建，可在网络构建的`fluid.data(shape=[])`中，指定shape参数来固定模型的输入大小。
+**Q6. 进行动态图转换时，提示『Fail to generate inference model! Problem happend while export inference model from python code...』**
+A: 此提示为无法将动态图代码转换为静态图模型，有两种可能：
+> 使用动态图代码确认转换后的代码是否正确，可使用如下代码进行确认：
+```
+import paddle
+import numpy as np
+np.random.seed(6)
+# ipt为输入数据
+ipt = np.random.rand(1, 3, 224, 224).astype("float32")
+paddle.disable_static()
+# pd_model_dygraph为保存路径（其中的”/“用”.“替换）
+from pd_model_dygraph.x2paddle_code import main
+out =main(ipt)
+```
+> 若运行代码无误，则说明代码中有op不支持动转静，我们将会再未来支持；若报错，则说明pytorch2paddle转换出错，请提issue，我们将及时回复。
\ No newline at end of file
--- a/add_caffe_custom_layer.md
+++ b/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。  
+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。 
+目前，代码中已经提供了10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) |
+| ReLU6 | [code](https://github.com/chuanqi305/ssd/blob/ssd/src/caffe/layers/relu6_layer.cpp) |
+| Upsample | [code](https://github.com/eric612/MobileNet-YOLO/blob/master/src/caffe/layers/upsample_layer.cpp) |
+添加代码实现自定义层的步骤如下：
 ***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：
 ```
@@ -10,7 +27,7 @@ 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  
+> 选择一：pip install protobuf  (protobuf >= 3.6.0)
 > 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）  
@@ -21,8 +38,9 @@ bash ./toos/compile.sh /home/root/caffe/src/caffe/proto
 ```
 ***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
+**静态图方式：**
- 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：
+- 进入./x2paddle/op_mapper/static/caffe2paddle/caffe_custom_layer，创建.py文件，例如mylayer.py
+- 仿照./x2paddle/op_mapper/static/caffe2paddle/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  
@@ -63,7 +81,116 @@ bash ./toos/compile.sh /home/root/caffe/src/caffe/proto
  ```
  from . import roipooling
  ```
+**动态图方式：**
+> 【注意】若Caffe自定义layer与Paddle的op一一对应，使用方式一，否则使用方式二。
+- 方式一：
+1. 仿照./x2paddle/op_mapper/dygraph/caffe2paddle/caffe_op_mapper.py中的CaffeOpMapper类中的映射方法（输入为self和node），实现类似的映射方法，以下述的映射方法为例：
+```python
+def Permute(self, node):
+    assert len(
+        node.inputs) == 1, "The count of Permute node\'s input is not 1."
+    input = self.graph.get_input_node(node, idx=0, copy=True)
+    params = node.layer.permute_param
+    order = list(params.order)    
+    self.paddle_graph.add_layer(
+        "paddle.transpose",
+        inputs={"x": input.name},
+        outputs=[node.layer_name],
+        perm=order)
+```
+>需完成的步骤：
+>    a. 获取Caffe Layer的属性，并对应转换为Paddle的属性。
+>    b. 获取当前Layer的输入。
+>    c. 使用self.paddle_graph.add_layer为PaddleGraph添加layer。其中，第一个参数代表Paddle的kernel；inputs是一个字典，用于存储paddle中的输入的key与其输入名字；outputs是一个列表，用于存储输出的名字；其余参数为属性对应关系。
+2. 仿照./x2paddle/decoder/caffe_shape_inference.py中的shape_xx方法，实现获取当前Layer输出大小的函数，以下述方法为例：
+```python
+def shape_permute(layer, input_shape):
+    order = layer.permute_param.order
+    inshape = input_shape[0]
+    output_shape = []
+    order = list(order)
+    for ii in order:
+        assert ii < len(inshape), "invalid order for permute[%s]" % (name)
+        output_shape.append(inshape[ii])
+    return [output_shape]
+```
+>参数：
+>    layer (caffe_pb2.LayerParameter): caffe的Layer，可用于获取当前Layer的属性。
+>    input_shape (list): 其中每个元素代表该层每个输入数据的大小。
+- 方式二：
+1. 进入./x2paddle/op_mapper/dygraph/caffe2paddle/caffe_custom_layer，创建.py文件，例如mylayer.py
+2. 仿照./x2paddle/op_mapper/dygraph/caffe2paddle/caffe_custom_layer中的其他文件，在mylayer.py中主要需要实现1个类，下面以roipooling.py为例分析代码：
+```python
+class ROIPooling(object):
+    def __init__(self, pooled_height, pooled_width, spatial_scale):
+        self.roipooling_layer_attrs = {
+            "pooled_height": pooled_height,
+            "pooled_width": pooled_width,
+            "spatial_scale": spatial_scale}
+    def __call__(self, x0, x1):
+        slice_x1 = paddle.slice(input=x1, axes=[1], 
+                                starts=[1], ends=[5])
+        out = fluid.layers.roi_pool(input=x0, 
+                                    rois=slice_x1, 
+                                    **self.roipooling_layer_attrs)
+        return out
+```
+>\_\_init\_\_函数：用于初始化各个属性
+>\_\_call\_\_函数：用于组合实现当前Layer的前向，输入为当前Layer所需要的输入
+3. 仿照./x2paddle/op_mapper/dygraph/caffe2paddle/caffe_op_mapper.py中的CaffeOpMapper类中的映射方法（输入为self和node），实现类似的映射方法，以下述的映射方法为例：
+```python
+def ROIPooling(self, node):
+      roipooling_name = name_generator("roipooling", self.nn_name2id)
+      output_name = node.layer_name
+      layer_outputs = [roipooling_name, output_name]
+      assert len(
+          node.inputs) == 2, "The count of ROIPooling node\'s input is not 2."
+      input0 = self.graph.get_input_node(node, idx=0, copy=True)
+      input1 = self.graph.get_input_node(node, idx=1, copy=True)
+      inputs_dict = {}
+      inputs_dict["x0"] = input0.name
+      inputs_dict["x1"] = input1.name
+      params = node.layer.roi_pooling_param
+      layer_attrs = {
+          "pooled_height": params.pooled_h,
+          "pooled_width": params.pooled_w,
+          "spatial_scale": params.spatial_scale}
+      self.paddle_graph.add_layer(
+          "custom_layer:ROIPooling",
+          inputs=inputs_dict,
+          outputs=layer_outputs,
+          **layer_attrs)
+```
+>需完成的步骤：
+>    a. 获取Caffe Layer的属性，并对应转换为Paddle的属性。
+>    b. 获取当前Layer的输入。
+>    c. 使用self.paddle_graph.add_layer为PaddleGraph添加layer。其中，第一个参数代表Paddle的kernel（此处kernel必须以“custom_layer:“开头）；inputs是一个字典，用于存储paddle中的输入的key与其输入名字；outputs是一个列表，用于存储输出的名字；其余参数为属性对应关系。
+4. 仿照./x2paddle/decoder/caffe_shape_inference.py中的shape_xx方法，实现获取当前Layer输出大小的函数，以下述方法为例：
+```python
+def shape_roipooling(layer, input_shape):
+  pooled_w = layer.roi_pooling_param.pooled_w
+  pooled_h = layer.roi_pooling_param.pooled_h
+  base_fea_shape = input_shapes[0]
+  rois_shape = input_shapes[1]
+  output_shape = base_fea_shape
+  output_shape[0] = rois_shape[0]
+  output_shape[2] = pooled_h
+  output_shape[3] = pooled_w
+  return [output_shape]
+```
+>参数：
+>    layer (caffe_pb2.LayerParameter): caffe的Layer，可用于获取当前Layer的属性。
+>    input_shape (list): 其中每个元素代表该层每个输入数据的大小。
 ***步骤四 运行转换代码***
 ```

--- a/export_tf_model.md
+++ b/export_tf_model.md
 ## 如何导出TensorFlow模型
 本文档介绍如何将TensorFlow模型导出为X2Paddle支持的模型格式。
+TensorFlow目前一般分为3种保存格式（checkpoint、FrozenModel、SavedModel），X2Paddle支持的是FrozenModel（将网络参数和网络结构同时保存到同一个文件中，并且只保存指定的前向计算子图），下面示例展示了如何导出X2Paddle支持的模型格式。
-TensorFlow提供了接口可将网络参数和网络结构同时保存到同一个文件中，并且只保存指定的前向计算子图，下面示例展示了如何导出tensorflow/models下的VGG16模型
+***下列代码Tensorflow 1.X下使用***
+- checkpoint模型+代码
 步骤一 下载模型参数文件
 ```
 wget http://download.tensorflow.org/models/vgg_16_2016_08_28.tar.gz
@@ -42,3 +43,70 @@ load_model(sess)
 # 导出模型
 freeze_model(sess, ["vgg_16/fc8/squeezed"], "vgg16.pb")
 ```
+- 纯checkpoint模型
+文件结构：
+> |--- checkpoint  
+> |--- model.ckpt-240000.data-00000-of-00001  
+> |--- model.ckpt-240000.index  
+> |--- model.ckpt-240000.meta  
+加载和导出模型：
+```python
+#coding: utf-8
+from tensorflow.python.framework import graph_util
+import tensorflow as tf
+# 固化模型函数
+# output_tensor_names: list，指定模型的输出tensor的name
+# freeze_model_path: 模型导出的文件路径
+def freeze_model(sess, output_tensor_names, freeze_model_path):
+    out_graph = graph_util.convert_variables_to_constants(
+        sess, sess.graph.as_graph_def(), output_tensor_names)
+    with tf.gfile.GFile(freeze_model_path, 'wb') as f:
+        f.write(out_graph.SerializeToString())
+    print("freeze model saved in {}".format(freeze_model_path))
+# 加载模型参数
+# 此处需要修改input_checkpoint（checkpoint的前缀）和save_pb_file（模型导出的文件路径）
+input_checkpoint = "./tfhub_models/save/model.ckpt"
+save_pb_file = "./tfhub_models/save.pb"
+sess = tf.Session()
+saver = tf.train.import_meta_graph(input_checkpoint + '.meta', clear_devices=True)
+saver.restore(sess, input_checkpoint)
+# 此处需要修改freeze_model的第二个参数，指定模型的输出tensor的name
+freeze_model(sess, ["vgg_16/fc8/squeezed"], save_pb_file)
+```
+- SavedModel模型
+文件结构：
+> |-- variables  
+> |------ variables.data-00000-of-00001  
+> |------ variables.data-00000-of-00001  
+> |-- saved_model.pb  
+加载和导出模型：
+```python
+#coding: utf-8
+import tensorflow as tf
+sess = tf.Session(graph=tf.Graph())
+# tf.saved_model.loader.load最后一个参数代表saved_model的保存路径
+tf.saved_model.loader.load(sess, {}, "/mnt/saved_model")
+graph = tf.get_default_graph()
+from tensorflow.python.framework import graph_util
+# 固化模型函数
+# output_tensor_names: list，指定模型的输出tensor的name
+# freeze_model_path: 模型导出的文件路径
+def freeze_model(sess, output_tensor_names, freeze_model_path):
+    out_graph = graph_util.convert_variables_to_constants(
+        sess, sess.graph.as_graph_def(), output_tensor_names)
+    with tf.gfile.GFile(freeze_model_path, 'wb') as f:
+        f.write(out_graph.SerializeToString())
+    print("freeze model saved in {}".format(freeze_model_path))
+# 导出模型
+freeze_model(sess, ["logits"], "model.pb")
+```
--- a/pytorch_to_onnx.md
+++ b/pytorch_to_onnx.md
--- a/pytorch2paddle.md
+++ b/pytorch2paddle.md
@@ -46,7 +46,7 @@ torch_module.load_state_dict(torch_state_dict)
 torch_module.eval()
 # 进行转换
 from x2paddle.convert import pytorch2paddle
-pytorch2paddle(torch_model, 
+pytorch2paddle(torch_module, 
               save_dir="pd_model_trace", 
               jit_type="trace", 
               input_examples=[torch.tensor(input_data)])

--- a/requirements.txt
+++ b/requirements.txt
 pre-commit
 yapf == 0.28.0
+pandas
+treelib
--- a/tools/merge_params.py
+++ b/tools/merge_params.py
+import paddle
 import paddle.fluid as fluid
 import sys
 model_dir = sys.argv[1]
 new_model_dir = sys.argv[2]
-exe = fluid.Executor(fluid.CPUPlace())
+paddle.enable_static()
+exe = paddle.static.Executor(paddle.CPUPlace())
 [inference_program, feed_target_names,
- fetch_targets] = fluid.io.load_inference_model(
+ fetch_targets] = paddle.static.load_inference_model(
     dirname=model_dir, executor=exe)
 print(feed_target_names)
-fluid.io.save_inference_model(
+paddle.static.save_inference_model(
    dirname=new_model_dir,
    feeded_var_names=feed_target_names,
    target_vars=fetch_targets,

--- a/x2paddle/convert.py
+++ b/x2paddle/convert.py
@@ -132,18 +132,9 @@ def tf2paddle(model_path,
        graph_opt = GraphOptimizer(source_frame="tf", paddle_type=paddle_type)
        graph_opt.optimize(mapper.paddle_graph)
    else:
-        from x2paddle.optimizer.tensorflow.bias import BiasOpt
+        from x2paddle.optimizer.optimizer import GraphOptimizer
-        from x2paddle.optimizer.tensorflow.transpose import TransposeOpt
+        graph_opt = GraphOptimizer(source_frame="tf", paddle_type=paddle_type)
-        from x2paddle.optimizer.tensorflow.batch_norm import BatchNormOpt
+        graph_opt.optimize(mapper.paddle_graph)
-        from x2paddle.optimizer.tensorflow.prelu import PReLUOpt
-        bias_opt = BiasOpt()
-        transpose_opt = TransposeOpt()
-        batch_norm_opt = BatchNormOpt()
-        prelu_opt = PReLUOpt()
-        bias_opt.run(mapper.paddle_graph)
-        batch_norm_opt.run(mapper.paddle_graph)
-        prelu_opt.run(mapper.paddle_graph)
-        transpose_opt.run(mapper.paddle_graph)
    mapper.paddle_graph.gen_model(save_dir)

--- a/x2paddle/op_mapper/dygraph/tf2paddle/tf_op_mapper.py
+++ b/x2paddle/op_mapper/dygraph/tf2paddle/tf_op_mapper.py
@@ -284,7 +284,6 @@ class TFOpMapper(OpMapper):
            inputs["shape"] = dims.name
        layer_attrs["dtype"] = string(input_value.dtype)
        layer_attrs["fill_value"] = input_value.value
        self.paddle_graph.add_layer(
            "paddle.full",
@@ -578,6 +577,9 @@ class TFOpMapper(OpMapper):
                inputs={"x": node.name},
                outputs=[node.name],
                perm=[0, 2, 3, 1])
+    def FusedBatchNormV3(self, node):
+        self.FusedBatchNorm(node)
    def Mean(self, node):
        input = self.graph.get_input_node(node, 0)
@@ -930,6 +932,23 @@ class TFOpMapper(OpMapper):
            outputs=[node.name],
            axis=axis)
+    def Concat(self, node):
+        inputs_list = list()
+        for i in range(1, len(node.inputs)):
+            inputs_list.append(self.graph.get_input_node(node, i))
+        axis = self.graph.get_input_node(node, 0)
+        assert axis.layer_type == "Const", "axis for ConcatV2 must be type Const"
+        axis = axis.value
+        if axis < 0:
+            axis += len(inputs_list[0].out_shapes[0])
+        input_names = [i.name for i in inputs_list]
+        self.paddle_graph.add_layer(
+            kernel="paddle.concat",
+            inputs={"x": input_names},
+            outputs=[node.name],
+            axis=axis)
    def AddN(self, node):
        inputs_list = list()
        for i in range(len(node.inputs) - 1):
@@ -1400,6 +1419,7 @@ class TFOpMapper(OpMapper):
        inputs = {"x": x.name, "y": y.name}
        x_shape = x.out_shapes[0]
        y_shape = y.out_shapes[0]
+        # TODO(syf)
        layer_id = self.paddle_graph.add_layer(
            "fluid.layers.elementwise_sub", inputs=inputs, outputs=[node.name])
        self.paddle_graph.layers[layer_id].input_shapes = {"x": x_shape, "y": y_shape}

--- a/x2paddle/op_mapper/static/caffe2paddle/caffe_op_mapper.py
+++ b/x2paddle/op_mapper/static/caffe2paddle/caffe_op_mapper.py
@@ -457,7 +457,6 @@ class CaffeOpMapper(OpMapper):
    def ReLU(self, node):
        """
        :param node:
        :return:
        """
@@ -973,5 +972,4 @@ class CaffeOpMapper(OpMapper):
        self.paddle_graph.add_layer(
            kernel=op_info,
            inputs={"x": self.get_input_name(input)},
            outputs=[node.layer_name])
\ No newline at end of file
--- a/x2paddle/op_mapper/static/tf2paddle/tf_op_mapper.py
+++ b/x2paddle/op_mapper/static/tf2paddle/tf_op_mapper.py
--- a/x2paddle/optimizer/elimination/static/__init__.py
+++ b/x2paddle/optimizer/elimination/static/__init__.py
+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from .transpose_elimination import StaticTransposeElimination
+from .transpose_eliminate_pass import StaticTransposeEliminatePass
\ No newline at end of file
--- a/x2paddle/optimizer/elimination/static/transpose_eliminate_pass.py
+++ b/x2paddle/optimizer/elimination/static/transpose_eliminate_pass.py
+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from x2paddle.optimizer.pass_ import Pass
+from x2paddle.optimizer.elimination.static import StaticTransposeElimination
+from x2paddle.optimizer.pass_manager import pass_register
+@pass_register
+class StaticTransposeEliminatePass(Pass):
+    name = "static_transpose_eliminate_pass"
+    def __init__(self):
+        Pass.__init__(self)
+    def apply(self, graph):
+        fuser = StaticTransposeElimination()
+        fuser.operate(graph)
+# 用于注册
+static_transpose_eliminate_pass = StaticTransposeEliminatePass()
\ No newline at end of file
--- a/x2paddle/optimizer/tensorflow/transpose.py
+++ b/x2paddle/optimizer/tensorflow/transpose.py
+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 import copy
 import sys
+import numpy as np
+from x2paddle.optimizer.pattern_matcher import FuseBase
+from x2paddle.core.program import PaddleGraph, PaddleLayer
+from x2paddle.core.util import *
-class TransposeOpt:
+class StaticTransposeElimination(FuseBase):
    def __init__(self):
-        self.image_layers = [
+        super(StaticTransposeElimination, self).__init__(graph_type="static")
-            'fluid.layers.conv2d', 'fluid.layers.batch_norm',
-            'fluid.layers.conv2d_transpose', 'fluid.layers.resize_nearest',
-            'fluid.layers.resize_bilinear', 'fluid.layers.pool2d',
-            'fluid.layers.pad2d'
-        ]
        self.direct_layers = [
-            'fluid.layers.relu', 'fluid.layers.relu6', 'fluid.layers.abs',
+            'paddle.nn.functional.relu', 'paddle.nn.functional.relu6', 'paddle.abs',
-            'fluid.layers.sigmoid', 'fluid.layers.exp', 'fluid.layers.rsqrt',
+            'paddle.nn.functional.sigmoid', 'paddle.exp', 'paddle.rsqrt',
-            'fluid.layers.swish_f32', 'fluid.layers.tanh',
+            'paddle.nn.functional.swish', 'paddle.tanh',
-            'fluid.layers.softplus', 'fluid.layers.leaky_relu',
+            'paddle.nn.functional.softplus', 'paddle.nn.functional.leaky_relu',
-            'fluid.layers.floor', 'fluid.layers.erf', 'fluid.layers.swish'
+            'paddle.floor', 'paddle.erf', 'paddle.square'
        ]
        self.elementwise_layers = [
-            'fluid.layers.elementwise_add', 'fluid.layers.elementwise_sub',
+            'paddle.add', 'fluid.layers.elementwise_sub',
-            'fluid.layers.elementwise_mul', 'fluid.layers.elementwise_div'
+            'paddle.multiply', 'paddle.divide'
        ]
        self.reduce_layers = [
-            'fluid.layers.reduce_mean', 'fluid.layers.reduce_all',
+            'paddle.mean', 'paddle.all',
-            'fluid.layers.reduce_max', 'fluid.layers.reduce_any',
+            'paddle.max', 'paddle.any',
-            'fluid.layers.reduce_sum', 'fluid.layers.reduce_prod'
+            'paddle.sum', 'paddle.prod'
        ]
    def get_transpose_num(self, graph):
        count = 0
        for layer_id, layer in graph.layers.items():
-            if layer.kernel == "fluid.layers.transpose":
+            if layer.kernel == "paddle.transpose":
                count += 1
        return count
-    def run(self, graph):
+    def operate(self, graph):
-        print("Optimize: TransposeOpt...")
        total_layer_num = len(graph.layers)
        scanned_layers = set()
        optimized_transpose_layers = list()
        optimized_reduce_layers = list()
        optimized_concat_layers = list()
        optimized_elementwise_layers = list()
+        def get_index(layer):
+            if layer.kernel.startswith("paddle.nn") and "functional" not in layer.kernel:
+                return 1
+            else:
+                return 0 
        def strip_transpose(_graph):
            layers = copy.deepcopy(_graph.layers)
@@ -53,7 +71,7 @@ class TransposeOpt:
                sys.stderr.write("\rOptimize Transpose Layers...{}%".format(
                    percent))
-                if layer.kernel != "fluid.layers.transpose":
+                if layer.kernel != "paddle.transpose":
                    continue
                if layer.attrs["perm"] != [0, 2, 3, 1]:
                    continue
@@ -65,7 +83,7 @@ class TransposeOpt:
                elementwise_layers = list()
                can_be_optimized = True
                for out in _graph.edges_out.get(layer_id, []):
-                    if _graph.layers[out].kernel == "fluid.layers.transpose":
+                    if _graph.layers[out].kernel == "paddle.transpose":
                        if _graph.layers[out].attrs["perm"] != [0, 3, 1, 2]:
                            can_be_optimized = False
                            break
@@ -73,21 +91,24 @@ class TransposeOpt:
                    elif _graph.layers[out].kernel in self.elementwise_layers:
                        propagate_layers.append(out)
                    elif _graph.layers[out].kernel in self.direct_layers:
-                        if _graph.layers[out].outputs[0] in _graph.outputs:
+                        ouput_index = get_index(_graph.layers[out])
+                        if _graph.layers[out].outputs[ouput_index] in _graph.outputs:
                            can_be_optimized = False
                            break
                        propagate_layers.append(out)
                    elif _graph.layers[out].kernel in self.reduce_layers:
-                        if _graph.layers[out].outputs[0] in _graph.outputs:
+                        ouput_index = get_index(_graph.layers[out])
+                        if _graph.layers[out].outputs[ouput_index] in _graph.outputs:
                            can_be_optimized = False
                            break
-                        if not _graph.layers[out].attrs.get('keep_dim', False):
+                        if not _graph.layers[out].attrs.get('keepdim', False):
                            can_be_optimized = False
                            break
                        propagate_layers.append(out)
                        reduce_layers.append(out)
-                    elif _graph.layers[out].kernel == "fluid.layers.concat":
+                    elif _graph.layers[out].kernel == "paddle.concat":
-                        if _graph.layers[out].outputs[0] in _graph.outputs:
+                        ouput_index = get_index(_graph.layers[out])
+                        if _graph.layers[out].outputs[ouput_index] in _graph.outputs:
                            can_be_optimized = False
                            break
                        propagate_layers.append(out)
@@ -102,37 +123,41 @@ class TransposeOpt:
                    visited_layers.add(current_id)
                    for out in _graph.edges_out.get(current_id, []):
                        if _graph.layers[
-                                out].kernel == "fluid.layers.transpose":
+                                out].kernel == "paddle.transpose":
                            if _graph.layers[out].attrs["perm"] != [0, 3, 1, 2]:
                                can_be_optimized = False
                                break
                            transpose_layers.append(out)
                        elif _graph.layers[
                                out].kernel in self.elementwise_layers:
-                            if _graph.layers[out].outputs[0] in _graph.outputs:
+                            output_index = get_index(_graph.layers[out])
+                            if _graph.layers[out].outputs[output_index] in _graph.outputs:
                                can_be_optimized = False
                                break
                            if out not in visited_layers:
                                propagate_layers.append(out)
                        elif _graph.layers[out].kernel in self.direct_layers:
-                            if _graph.layers[out].outputs[0] in _graph.outputs:
+                            output_index = get_index(_graph.layers[out])
+                            if _graph.layers[out].outputs[output_index] in _graph.outputs:
                                can_be_optimized = False
                                break
                            if out not in visited_layers:
                                propagate_layers.append(out)
                        elif _graph.layers[out].kernel in self.reduce_layers:
-                            if _graph.layers[out].outputs[0] in _graph.outputs:
+                            output_index = get_index(_graph.layers[out])
+                            if _graph.layers[out].outputs[output_index] in _graph.outputs:
                                can_be_optimized = False
                                break
-                            if not _graph.layers[out].attrs.get('keep_dim',
+                            if not _graph.layers[out].attrs.get('keepdim',
                                                                False):
                                can_be_optimized = False
                                break
                            if out not in visited_layers:
                                propagate_layers.append(out)
                                reduce_layers.append(out)
-                        elif _graph.layers[out].kernel == "fluid.layers.concat":
+                        elif _graph.layers[out].kernel == "paddle.concat":
-                            if _graph.layers[out].outputs[0] in _graph.outputs:
+                            output_index = get_index(_graph.layers[out])
+                            if _graph.layers[out].outputs[output_index] in _graph.outputs:
                                can_be_optimized = False
                                break
                            if out not in visited_layers:
@@ -149,14 +174,15 @@ class TransposeOpt:
                                    current_id].input_shapes['x']
                                y_shape = _graph.layers[
                                    current_id].input_shapes['y']
+                                output_index = get_index(_graph.layers[ipt])
                                if _graph.layers[ipt].outputs[
-                                        0] == _graph.layers[current_id].inputs[
+                                        output_index] == _graph.layers[current_id].inputs[
                                            'x']:
                                    if len(x_shape) <= 1:
                                        elementwise_layers.append(current_id)
                                        continue
                                elif _graph.layers[ipt].outputs[
-                                        0] == _graph.layers[current_id].inputs[
+                                        output_index] == _graph.layers[current_id].inputs[
                                            'y']:
                                    if len(y_shape) <= 1:
                                        elementwise_layers.append(current_id)
@@ -168,8 +194,9 @@ class TransposeOpt:
                            except Exception as e:
                                can_be_optimized = False
                                break
+                        output_index = get_index(_graph.layers[ipt])
                        if _graph.layers[
-                                ipt].kernel == "fluid.layers.transpose":
+                                ipt].kernel == "paddle.transpose":
                            if _graph.layers[ipt].attrs["perm"] != [0, 2, 3, 1]:
                                can_be_optimized = False
                                break
@@ -177,30 +204,30 @@ class TransposeOpt:
                                transpose_layers.append(ipt)
                        elif _graph.layers[
                                ipt].kernel in self.elementwise_layers:
-                            if _graph.layers[ipt].outputs[0] in _graph.outputs:
+                            if _graph.layers[ipt].outputs[output_index] in _graph.outputs:
                                can_be_optimized = False
                                break
                            if ipt not in visited_layers:
                                propagate_layers.append(ipt)
                        elif _graph.layers[ipt].kernel in self.direct_layers:
-                            if _graph.layers[ipt].outputs[0] in _graph.outputs:
+                            if _graph.layers[ipt].outputs[output_index] in _graph.outputs:
                                can_be_optimized = False
                                break
                            if ipt not in visited_layers:
                                propagate_layers.append(ipt)
                        elif _graph.layers[ipt].kernel in self.reduce_layers:
-                            if _graph.layers[ipt].outputs[0] in _graph.outputs:
+                            if _graph.layers[ipt].outputs[output_index] in _graph.outputs:
                                can_be_optimized = False
                                break
-                            if not _graph.layers[ipt].attrs.get('keep_dim',
+                            if not _graph.layers[ipt].attrs.get('keepdim',
                                                                False):
                                can_be_optimized = False
                                break
                            if ipt not in visited_layers:
                                propagate_layers.append(ipt)
                                reduce_layers.append(ipt)
-                        elif _graph.layers[ipt].kernel == "fluid.layers.concat":
+                        elif _graph.layers[ipt].kernel == "paddle.concat":
-                            if _graph.layers[ipt].outputs[0] in _graph.outputs:
+                            if _graph.layers[ipt].outputs[output_index] in _graph.outputs:
                                can_be_optimized = False
                                break
                            if ipt not in visited_layers:
@@ -217,7 +244,8 @@ class TransposeOpt:
                transpose_layers.append(layer_id)
                transpose_layers = list(set(transpose_layers))
                for l in transpose_layers:
-                    if graph.layers[l].outputs[0] in graph.outputs:
+                    output_index = get_index(graph.layers[l])
+                    if graph.layers[l].outputs[output_index] in graph.outputs:
                        can_be_optimized = False
                        break
                if not can_be_optimized:
@@ -243,17 +271,19 @@ class TransposeOpt:
        for layer_id in list(set(optimized_transpose_layers)):
            graph.del_layer(layer_id)
        for layer_id in list(set(optimized_reduce_layers)):
-            dim = graph.layers[layer_id].attrs.get('dim', None)
+            dim = graph.layers[layer_id].attrs.get('axis', None)
            if dim is not None:
                for i in range(len(dim)):
                    dim[i] = [0, 2, 3, 1][dim[i]]
-                graph.layers[layer_id].attrs['dim'] = dim
+                graph.layers[layer_id].attrs['axis'] = dim
        for layer_id in list(set(optimized_concat_layers)):
            axis = graph.layers[layer_id].attrs.get('axis', 0)
            graph.layers[layer_id].attrs['axis'] = [0, 2, 3, 1][axis]
        for layer_id in list(set(optimized_elementwise_layers)):
            axis = graph.layers[layer_id].attrs.get('axis', -1)
            graph.layers[layer_id].attrs['axis'] = [0, 2, 3, 1][axis]
+            if graph.layers[layer_id].kernel == "paddle.add":
+                graph.layers[layer_id].kernel = "fluid.layers.elementwise_add"
        current_transpose_num = self.get_transpose_num(graph)
        print(

--- a/x2paddle/optimizer/fusion/dygraph/conv2d_add_fuser.py
+++ b/x2paddle/optimizer/fusion/dygraph/conv2d_add_fuser.py
@@ -105,10 +105,6 @@ class DygraphConv2DAddFuser(FuseBase):
            if layer.kernel == "paddle.nn.Conv2D":
                conv_id = layer_id
        for layer_id, layer in matches.items():
-            if layer.kernel == "paddle.nn.functional.conv2d_transpose":
-                layer.bias = bias_name
-                if not is_transpose:
-                    layer.outputs[0] = output_name
            if layer.kernel == "paddle.nn.Conv2D":
                layer.attrs["bias_attr"] = bias_name
                if not is_transpose:

--- a/x2paddle/optimizer/fusion/static/__init__.py
+++ b/x2paddle/optimizer/fusion/static/__init__.py
@@ -13,4 +13,11 @@
 # limitations under the License.
 from .bn_scale_fuser import Static_BNScaleFuser
 from .bn_scale_fuse_pass import Static_BNScaleFusePass
\ No newline at end of file
+from .conv2d_add_fuser import StaticConv2DAddFuser
+from .conv2d_add_fuse_pass import StaticConv2DAddFusePass
+from .prelu_fuser import StaticPReLUFuser
+from .prelu_fuse_pass import StaticPReLUFusePass
+from .tf_batchnorm_fuser import StaticTFBatchNormFuser
+from .tf_batchnorm_fuse_pass import StaticTFBatchNormFusePass
--- a/x2paddle/optimizer/fusion/static/bn_scale_fuser.py
+++ b/x2paddle/optimizer/fusion/static/bn_scale_fuser.py
@@ -79,7 +79,6 @@ class Static_BNScaleFuser(FuseBase):
        graph.layers[new_layer_id] = new_layer
        matches.pop(new_layer_id)
    def gen_new_layer(self, parameters, matches):
        layers_id = list(matches.keys())
        layer = matches[layers_id[0]]

--- a/x2paddle/optimizer/fusion/static/conv2d_add_fuse_pass.py
+++ b/x2paddle/optimizer/fusion/static/conv2d_add_fuse_pass.py
+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from x2paddle.optimizer.pass_ import Pass
+from x2paddle.optimizer.fusion.static import StaticConv2DAddFuser
+from x2paddle.optimizer.pass_manager import pass_register
+@pass_register
+class StaticConv2DAddFusePass(Pass):
+    name = "static_conv2d_add_fuse_pass"
+    def __init__(self):
+        Pass.__init__(self)
+    def apply(self, graph):
+        fuser = StaticConv2DAddFuser()
+        fuser.operate(graph, match_kind="edge")
+# 用于注册
+static_conv2d_add_fuse_pass = StaticConv2DAddFusePass()
\ No newline at end of file
--- a/x2paddle/optimizer/fusion/static/conv2d_add_fuser.py
+++ b/x2paddle/optimizer/fusion/static/conv2d_add_fuser.py
+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import copy
+import numpy as np
+from x2paddle.optimizer.pattern_matcher import FuseBase
+from x2paddle.core.program import PaddleGraph, PaddleLayer
+from x2paddle.core.util import *
+class StaticConv2DAddFuser(FuseBase):
+    def __init__(self):
+        super(StaticConv2DAddFuser, self).__init__(graph_type="static")
+        self.patterns = list()
+    def build_pattern(self):
+        """ 描述需要替换的conv2d+add图结构。
+        conv2d+add层模式python实现代码示例:
+        模式一：
+        MobilenetV1_Logits_Conv2d_1c_1x1_biases = paddle.static.create_parameter(dtype='float32', shape=[1001], name='MobilenetV1_Logits_Conv2d_1c_1x1_biases', default_initializer=paddle.nn.initializer.Constant(value=0.0))
+        conv2d_transpose_14 = paddle.transpose(x=MobilenetV1_Logits_AvgPool_1a_AvgPool, perm=[0, 3, 1, 2])
+        MobilenetV1_Logits_Conv2d_1c_1x1_Conv2D = paddle.nn.functional.conv2d(x=conv2d_transpose_14, weight=MobilenetV1_Logits_Conv2d_1c_1x1_weights, bias=None, stride=[1, 1], dilation=[1, 1], padding='SAME')
+        MobilenetV1_Logits_Conv2d_1c_1x1_Conv2D = paddle.transpose(x=MobilenetV1_Logits_Conv2d_1c_1x1_Conv2D, perm=[0, 2, 3, 1])
+        MobilenetV1_Logits_Conv2d_1c_1x1_BiasAdd = paddle.add(x=MobilenetV1_Logits_Conv2d_1c_1x1_Conv2D, y=MobilenetV1_Logits_Conv2d_1c_1x1_biases)
+        模式二：
+        MobilenetV1_Logits_Conv2d_1c_1x1_biases = paddle.static.create_parameter(dtype='float32', shape=[1001], name='MobilenetV1_Logits_Conv2d_1c_1x1_biases', default_initializer=paddle.nn.initializer.Constant(value=0.0))
+        MobilenetV1_Logits_Conv2d_1c_1x1_Conv2D = paddle.nn.functional.conv2d(x=conv2d_transpose_14, weight=MobilenetV1_Logits_Conv2d_1c_1x1_weights, bias=None, stride=[1, 1], dilation=[1, 1], padding='SAME')
+        MobilenetV1_Logits_Conv2d_1c_1x1_BiasAdd = paddle.add(x=MobilenetV1_Logits_Conv2d_1c_1x1_Conv2D, y=MobilenetV1_Logits_Conv2d_1c_1x1_biases)
+        """
+        def gen_name(id):
+            return "x" + str(id)
+        pattern = PaddleGraph(graph_type="dygraph")
+        pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(0)])
+        pattern.add_layer(
+            kernel="paddle.transpose",
+            inputs={"x": "conv-input-0"},
+            outputs=[gen_name(1)],
+            perm=[0, 3, 1, 2])
+        pattern.add_layer(
+            kernel="paddle.nn.functional.conv2d",
+            inputs={"input": gen_name(1),
+                    "weight": "conv-input-1"},
+            outputs=[gen_name(2)])
+        pattern.add_layer(
+            kernel="paddle.transpose",
+            inputs={"x": gen_name(2)},
+            outputs=[gen_name(2)],
+            perm=[0, 2, 3, 1])
+        pattern.add_layer(
+            kernel="paddle.add",
+            inputs={"x": gen_name(2),
+                    "y": gen_name(0)},
+            outputs=[gen_name(3)])
+        pattern.build(inputs={"input-0": "conv-input-0", 
+                              "input-1": "conv-input-1"})
+        self.patterns.append(pattern)
+        pattern = PaddleGraph(graph_type="dygraph")
+        pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(0)])
+        pattern.add_layer(
+            kernel="paddle.nn.functional.conv2d",
+            inputs={"input": "conv-input-0",
+                    "weight": "conv-input-1"},
+            outputs=[gen_name(1)])
+        pattern.add_layer(
+            kernel="paddle.add",
+            inputs={"x": gen_name(1),
+                    "y": gen_name(0)},
+            outputs=[gen_name(2)])
+        pattern.build(inputs={"input-0": "conv-input-0", 
+                              "input-1": "conv-input-1"})
+        self.patterns.append(pattern)
+    def insert_new_layer(self, graph, parameters, matches):
+        self.gen_new_layer(matches, graph)
+        matches_copy = copy.deepcopy(matches)
+        for layer_id, layer in matches_copy.items():
+            if layer.kernel not in ["paddle.add"]:
+                matches.pop(layer_id)
+    def gen_new_layer(self, matches, graph):
+        is_transpose = False
+        for layer_id, layer in matches.items():
+            if layer.kernel == "paddle.static.create_parameter":
+                bias_name = layer.attrs["name"][1: -1]
+            if layer.kernel == "paddle.transpose":
+                is_transpose = True
+            if layer.kernel == "paddle.add":
+                output_name = layer.outputs[0]
+            if layer.kernel == "paddle.nn.functional.conv2d":
+                conv_id = layer_id
+        for layer_id, layer in matches.items():
+            if layer.kernel == "paddle.nn.functional.conv2d":
+                layer.inputs["bias"] = bias_name
+                layer.attrs.pop("bias")
+                if not is_transpose:
+                    layer.outputs[0] = output_name
+            if layer.kernel == "paddle.transpose":
+                if conv_id in graph.edges_in[layer_id]:
+                    layer.outputs[0] = output_name
--- a/x2paddle/optimizer/fusion/static/prelu_fuse_pass.py
+++ b/x2paddle/optimizer/fusion/static/prelu_fuse_pass.py
+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from x2paddle.optimizer.pass_ import Pass
+from x2paddle.optimizer.fusion.static import StaticPReLUFuser
+from x2paddle.optimizer.pass_manager import pass_register
+@pass_register
+class StaticPReLUFusePass(Pass):
+    name = "static_prelu_fuse_pass"
+    def __init__(self):
+        Pass.__init__(self)
+    def apply(self, graph):
+        fuser = StaticPReLUFuser()
+        fuser.operate(graph, match_kind="edge")
+# 用于注册
+static_prelu_fuse_pass = StaticPReLUFusePass()
\ No newline at end of file
--- a/x2paddle/optimizer/fusion/static/prelu_fuser.py
+++ b/x2paddle/optimizer/fusion/static/prelu_fuser.py
+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import copy
+import numpy as np
+from collections import OrderedDict
+from x2paddle.optimizer.pattern_matcher import FuseBase
+from x2paddle.core.program import PaddleGraph, PaddleLayer
+from x2paddle.core.util import *
+class StaticPReLUFuser(FuseBase):
+    def __init__(self):
+        super(StaticPReLUFuser, self).__init__(graph_type="static")
+    def build_pattern(self):
+        """ 描述需要替换的prelu图结构。
+        prelu层模式python实现代码示例:
+            conv4_alphas = paddle.static.create_parameter(dtype='float32', shape=[128], name='conv4_alphas', default_initializer=paddle.nn.initializer.Constant(value=0.0))
+            conv4_mul_1_y = paddle.full(dtype='float32', shape=[1], fill_value=0.5)
+            conv4_Relu = paddle.nn.functional.relu(x=conv4_BiasAdd)
+            conv4_Abs = paddle.abs(x=conv4_BiasAdd)
+            conv4_sub = fluid.layers.elementwise_sub(x=conv4_BiasAdd, y=conv4_Abs)
+            conv4_mul = paddle.multiply(x=conv4_alphas, y=conv4_sub)
+            conv4_mul_1 = paddle.multiply(x=conv4_mul, y=conv4_mul_1_y)
+            conv4_add = paddle.add(x=conv4_Relu, y=conv4_mul_1)
+        """
+        def gen_name(id):
+            return "x" + str(id)
+        self.pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(0)])
+        self.pattern.add_layer(
+            "paddle.full",
+            inputs={},
+            outputs=[gen_name(1)],
+            shape=[1],
+            fill_value=0.5)
+        self.pattern.add_layer(
+            "paddle.nn.functional.relu",
+            inputs={"x": "prelu-input-0"},
+            outputs=[gen_name(2)])
+        self.pattern.add_layer(
+            "paddle.abs",
+            inputs={"x": "prelu-input-0"},
+            outputs=[gen_name(3)])
+        self.pattern.add_layer(
+            "fluid.layers.elementwise_sub",
+            inputs={"x": "prelu-input-0",
+                    "y": gen_name(3)},
+            outputs=[gen_name(4)])
+        self.pattern.add_layer(
+            "paddle.multiply",
+            inputs={"x": gen_name(0),
+                    "y": gen_name(4)},
+            outputs=[gen_name(5)])
+        self.pattern.add_layer(
+            "paddle.multiply",
+            inputs={"x": gen_name(5),
+                    "y": gen_name(1)},
+            outputs=[gen_name(6)])
+        self.pattern.add_layer(
+            "paddle.add",
+            inputs={"x": gen_name(2),
+                    "y": gen_name(6)},
+            outputs=[gen_name(7)])
+        self.pattern.build(inputs={"input-0": "prelu-input-0", })
+    def insert_new_layer(self, graph, parameters, matches):
+        new_layers, last_layer_id = self.gen_new_layer(matches, parameters, graph)
+        matches_copy = copy.deepcopy(matches)
+        for layer_id, layer in matches_copy.items():
+            for i in range(4):
+                if layer_id == new_layers[i].id:
+                    matches.pop(new_layers[i].id)
+        prefix_layers = OrderedDict()
+        mid_layers = OrderedDict()
+        suffix_layers = OrderedDict()
+        is_need_id = False
+        for layer_id, layer in graph.layers.items():
+            if is_need_id:
+                suffix_layers[layer_id] = layer
+            else:
+                if layer_id == last_layer_id:
+                    for i in range(4):
+                        mid_layers[new_layers[i].id] = new_layers[i]
+                    is_need_id = True
+                prefix_layers[layer_id] = layer
+        prefix_layers.update(mid_layers)
+        prefix_layers.update(suffix_layers)
+        graph.layers = prefix_layers
+    def gen_new_layer(self, matches, parameters, graph):
+        layer_id_list = list(matches.keys())
+        layer_id_list.sort(key = int)
+        for layer_id, layer in matches.items():
+            if layer.kernel == "paddle.nn.functional.relu":
+                input_name = layer.inputs["x"]
+            if layer.kernel == "paddle.static.create_parameter":
+                param_layer = layer
+                param_name = layer.outputs[0]
+            if layer.kernel == "paddle.add":
+                output_name = layer.outputs[0]
+        transpose0 = PaddleLayer(
+            id=layer_id_list[-1] + "_1",
+            kernel="paddle.transpose",
+            inputs={"x": input_name},
+            outputs=["{}_transpose_for_prelu".format(input_name)],
+            perm=[0, 3, 1, 2])
+        param = parameters[param_name]
+        c = param.shape[0]
+        prelu = PaddleLayer(id=layer_id_list[-1] + "_2",
+                            kernel="paddle.nn.functional.prelu",
+                            inputs={"x": "{}_transpose_for_prelu".format(input_name),
+                                    "weight": param_name},
+                            outputs=["{}_prelu".format(input_name)])
+        transpose1 = PaddleLayer(
+            id=layer_id_list[-1] + "_3",
+            kernel="paddle.transpose",
+            inputs={"x": "{}_prelu".format(input_name)},
+            outputs=[output_name],
+            perm=[0, 2, 3, 1])
+        return [param_layer, transpose0, prelu, transpose1], layer_id_list[-1]
--- a/x2paddle/optimizer/fusion/static/tf_batchnorm_fuse_pass.py
+++ b/x2paddle/optimizer/fusion/static/tf_batchnorm_fuse_pass.py
+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from x2paddle.optimizer.pass_ import Pass
+from x2paddle.optimizer.fusion.static import StaticTFBatchNormFuser
+from x2paddle.optimizer.pass_manager import pass_register
+@pass_register
+class StaticTFBatchNormFusePass(Pass):
+    name = "static_tf_batchnorm_fuse_pass"
+    def __init__(self):
+        Pass.__init__(self)
+    def apply(self, graph):
+        fuser = StaticTFBatchNormFuser()
+        fuser.operate(graph, match_kind="edge")
+# 用于注册
+static_tf_batchnorm_fuse_pass = StaticTFBatchNormFusePass()
\ No newline at end of file
--- a/x2paddle/optimizer/fusion/static/tf_batchnorm_fuser.py
+++ b/x2paddle/optimizer/fusion/static/tf_batchnorm_fuser.py
+#   Copyright (c) 2020  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import copy
+import numpy as np
+from collections import OrderedDict
+from x2paddle.optimizer.pattern_matcher import FuseBase
+from x2paddle.core.program import PaddleGraph, PaddleLayer
+from x2paddle.core.util import *
+class StaticTFBatchNormFuser(FuseBase):
+    def __init__(self):
+        super(StaticTFBatchNormFuser, self).__init__(graph_type="static")
+        self.patterns = list()
+    def build_pattern(self):
+        """ 描述需要替换的batchnorm图结构。
+        batchnorm层模式python实现代码示例:
+        """
+        def gen_name(id):
+            return "x" + str(id)
+        pattern = PaddleGraph(graph_type="dygraph")
+        pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(0)])
+        pattern.add_layer(
+            "paddle.full",
+            inputs={},
+            outputs=[gen_name(1)],
+            shape=[1])
+        pattern.add_layer(
+            "paddle.add",
+            inputs={"x": gen_name(0), "y": gen_name(1)},
+            outputs=[gen_name(2)])
+        pattern.add_layer(
+            "paddle.rsqrt",
+            inputs={"x": gen_name(2)},
+            outputs=[gen_name(3)])
+        pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(4)])
+        pattern.add_layer(
+            "paddle.multiply",
+            inputs={"x": gen_name(3), "y": gen_name(4)},
+            outputs=[gen_name(5)])
+        pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(6)])
+        pattern.add_layer(
+            "paddle.multiply",
+            inputs={"x": gen_name(6), "y": gen_name(5)},
+            outputs=[gen_name(7)])
+        pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(8)])
+        pattern.add_layer(
+            "fluid.layers.elementwise_sub",
+            inputs={"x": gen_name(8), "y": gen_name(7)},
+            outputs=[gen_name(9)])
+        pattern.add_layer(
+            "paddle.multiply",
+            inputs={"x": "bn-input-0", "y": gen_name(5)},
+            outputs=[gen_name(10)])
+        pattern.add_layer(
+            "paddle.add",
+            inputs={"x": gen_name(10), "y": gen_name(9)},
+            outputs=[gen_name(11)])
+        pattern.build(inputs={"input-0": "bn-input-0", })
+        self.patterns.append(pattern)
+        pattern = PaddleGraph(graph_type="dygraph")
+        pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(0)])
+        pattern.add_layer(
+            "paddle.full",
+            inputs={},
+            outputs=[gen_name(1)],
+            shape=[1])
+        pattern.add_layer(
+            "paddle.add",
+            inputs={"x": gen_name(0), "y": gen_name(1)},
+            outputs=[gen_name(2)])
+        pattern.add_layer(
+            "paddle.rsqrt",
+            inputs={"x": gen_name(2)},
+            outputs=[gen_name(3)])
+        pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(4)])
+        pattern.add_layer(
+            "paddle.multiply",
+            inputs={"x": gen_name(3), "y": gen_name(4)},
+            outputs=[gen_name(5)])
+        pattern.add_layer(
+            "paddle.multiply",
+            inputs={"x": "bn-input-0", "y": gen_name(5)},
+            outputs=[gen_name(10)])
+        pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(6)])
+        pattern.add_layer(
+            "paddle.multiply",
+            inputs={"x": gen_name(6), "y": gen_name(5)},
+            outputs=[gen_name(7)])
+        pattern.add_layer(
+            "paddle.static.create_parameter",
+            inputs={},
+            outputs=[gen_name(8)])
+        pattern.add_layer(
+            "fluid.layers.elementwise_sub",
+            inputs={"x": gen_name(8), "y": gen_name(7)},
+            outputs=[gen_name(9)])
+        pattern.add_layer(
+            "paddle.add",
+            inputs={"x": gen_name(10), "y": gen_name(9)},
+            outputs=[gen_name(11)])
+        pattern.build(inputs={"input-0": "bn-input-0", })
+        self.patterns.append(pattern)
+    def insert_new_layer(self, graph, parameters, matches):
+        new_layers, last_layer_id = self.gen_new_layer(matches, parameters, graph)
+        matches_copy = copy.deepcopy(matches)
+        for layer_id, layer in matches_copy.items():
+            for i in range(7):
+                if layer_id == new_layers[i].id:
+                    matches.pop(new_layers[i].id)
+        prefix_layers = OrderedDict()
+        mid_layers = OrderedDict()
+        suffix_layers = OrderedDict()
+        is_need_id = False
+        for layer_id, layer in graph.layers.items():
+            if is_need_id:
+                suffix_layers[layer_id] = layer
+            else:
+                if layer_id == last_layer_id:
+                    for i in range(7):
+                        mid_layers[new_layers[i].id] = new_layers[i]
+                    is_need_id = True
+                prefix_layers[layer_id] = layer
+        prefix_layers.update(mid_layers)
+        prefix_layers.update(suffix_layers)
+        graph.layers = prefix_layers
+    def gen_new_layer(self, matches, parameters, graph):
+        layer_id_list = list(matches.keys())
+        layer_id_list.sort(key = int)
+        for layer_id, layer in matches.items():
+            if layer.kernel == "paddle.full":
+                full_layer = layer
+                out_layer_id = graph.edges_out[layer_id][0]
+                if matches[out_layer_id].kernel == "paddle.add":
+                    var_layer_id = graph.edges_in[out_layer_id][0]
+                    var_layer = matches[var_layer_id]
+            if layer.kernel == "paddle.rsqrt":
+                out_layer_id = graph.edges_out[layer_id][0]
+                if matches[out_layer_id].kernel == "paddle.multiply":
+                    gamma_layer_id = graph.edges_in[out_layer_id][1]
+                    gamma_layer = matches[gamma_layer_id]
+            if layer.kernel == "fluid.layers.elementwise_sub":
+                in_layer_id = graph.edges_in[layer_id][0]
+                beta_layer = matches[in_layer_id]
+                in_layer_id = graph.edges_in[layer_id][1]
+                in_layer_id = graph.edges_in[in_layer_id][0]
+                mean_layer = matches[in_layer_id]
+                out_layer_id = graph.edges_out[layer_id][0]
+                add_layer = matches[out_layer_id]
+            if layer.kernel == "paddle.multiply":
+                in_layer_id = graph.edges_in[layer_id][1]
+                mul_layer = matches[in_layer_id]
+                if mul_layer.kernel == "paddle.multiply":
+                    in_layer_id = graph.edges_in[layer_id][0]
+                    if in_layer_id not in matches:
+                        input_name = layer.inputs["x"]
+        transpose0 = PaddleLayer(
+            id=layer_id_list[-1] + "_1",
+            kernel="paddle.transpose",
+            inputs={"x": input_name},
+            outputs=["{}_transpose_for_bn".format(input_name)],
+            perm=[0, 3, 1, 2])
+        params = parameters[gamma_layer.outputs[0]]
+        c = params.shape[0]
+        bn = PaddleLayer(
+            id=layer_id_list[-1] + "_2",
+            kernel="paddle.nn.functional.batch_norm",
+            inputs={"x": "{}_transpose_for_bn".format(input_name),
+                    "running_mean": mean_layer.outputs[0],
+                    "running_var": var_layer.outputs[0],
+                    "weight": gamma_layer.outputs[0],
+                    "bias": beta_layer.outputs[0]},
+            outputs=["{}_bn".format(input_name)],
+            epsilon=full_layer.attrs["fill_value"])
+        transpose1 = PaddleLayer(
+            id=layer_id_list[-1] + "_3",
+            kernel="paddle.transpose",
+            inputs={"x": "{}_bn".format(input_name)},
+            outputs=add_layer.outputs,
+            perm=[0, 2, 3, 1])
+        mean_layer.id = layer_id_list[-1] + "_01"
+        var_layer.id = layer_id_list[-1] + "_02"
+        gamma_layer.id = layer_id_list[-1] + "_03"
+        beta_layer.id = layer_id_list[-1] + "_04"
+        return [mean_layer, var_layer, gamma_layer, beta_layer, 
+                transpose0, bn, transpose1], layer_id_list[-1]
--- a/x2paddle/optimizer/optimizer.py
+++ b/x2paddle/optimizer/optimizer.py
@@ -16,13 +16,13 @@ from x2paddle.optimizer.pass_manager import PassManager
 from x2paddle.optimizer.fusion.dygraph import *
 from x2paddle.optimizer.fusion.static import *
 from x2paddle.optimizer.elimination.dygraph import *
+from x2paddle.optimizer.elimination.static import *
 class GraphOptimizer(object):
    def __init__(self, source_frame, paddle_type="dygraph", jit_type="trace"):
        if source_frame == "pytorch":
            if jit_type == "trace":
-                self.passes = ["dygraph_constant_fuse_pass", 
+                self.passes = ["trace_fc_fuse_pass"]
-                               "trace_fc_fuse_pass"]
            else:
                self.passes = [
                    "dygraph_constant_fuse_pass", 
@@ -39,12 +39,20 @@ class GraphOptimizer(object):
            else:
                self.passes = ["static_bn_scale_fuse_pass"]
        elif source_frame == "tf":
-            self.passes = [
+            if paddle_type == "dygraph":
-                "dygraph_conv2d_add_fuse_pass",
+                self.passes = [
-                "dygraph_tf_batchnorm_fuse_pass",
+                    "dygraph_conv2d_add_fuse_pass",
-                "dygraph_prelu_fuse_pass",
+                    "dygraph_tf_batchnorm_fuse_pass",
-                "transpose_eliminate_pass"
+                    "dygraph_prelu_fuse_pass",
-            ]
+                    "transpose_eliminate_pass"
+                ]
+            else:
+                self.passes = [
+                    "static_conv2d_add_fuse_pass",
+                    "static_tf_batchnorm_fuse_pass",
+                    "static_prelu_fuse_pass",
+                    "static_transpose_eliminate_pass"
+                ]
        else:
            self.passes = []

--- a/x2paddle/optimizer/tensorflow/__init__.py
+++ b/x2paddle/optimizer/tensorflow/__init__.py
--- a/x2paddle/optimizer/tensorflow/batch_norm.py
+++ b/x2paddle/optimizer/tensorflow/batch_norm.py
-import copy
-from collections import OrderedDict
-from x2paddle.core.program import PaddleLayer
-class BatchNormOpt:
-    def __init__(self):
-        pass
-    def run(self, graph):
-        print("Optimize: BatchNormOpt...")
-        layers = copy.deepcopy(graph.layers)
-        for layer_id, layer in layers.items():
-            if layer.kernel != "fluid.layers.elementwise_add":
-                continue
-            axis = layer.attrs.get('axis', -1)
-            if axis != -1 and axis != 3:
-                continue
-            input_ids0 = graph.edges_in[layer_id]
-            mul_layer0 = graph.layers[input_ids0[0]]
-            sub_layer0 = graph.layers[input_ids0[1]]
-            if mul_layer0.kernel != "fluid.layers.elementwise_mul":
-                continue
-            if sub_layer0.kernel != "fluid.layers.elementwise_sub":
-                continue
-            axis = mul_layer0.attrs.get('axis', -1)
-            if axis != -1 and axis != 3:
-                continue
-            axis = sub_layer0.attrs.get('axis', -1)
-            if axis != -1 and axis != 0:
-                continue
-            if len(graph.edges_out.get(input_ids0[0], [])) != 1:
-                continue
-            if len(graph.edges_out.get(input_ids0[1], [])) != 1:
-                continue
-            input_ids1 = graph.edges_in[input_ids0[0]]
-            nhwc_input = graph.layers[input_ids1[0]]
-            mul_layer1 = graph.layers[input_ids1[1]]
-            if mul_layer1.kernel != "fluid.layers.elementwise_mul":
-                continue
-            axis = mul_layer1.attrs.get('axis', -1)
-            if axis != -1 and axis != 0:
-                continue
-            if len(graph.edges_out.get(input_ids1[1], [])) != 2:
-                continue
-            input_ids2 = graph.edges_in[input_ids0[1]]
-            beta = graph.layers[input_ids2[0]]
-            mul_layer2 = graph.layers[input_ids2[1]]
-            if beta.kernel != "fluid.layers.create_parameter":
-                continue
-            axis = mul_layer2.attrs.get('axis', -1)
-            if axis != -1 and axis != 0:
-                continue
-            if len(graph.edges_out.get(input_ids2[0], [])) != 1:
-                continue
-            if len(graph.edges_out.get(input_ids2[1], [])) != 1:
-                continue
-            if beta.outputs[0] not in graph.parameters:
-                continue
-            beta_shape = graph.parameters[beta.outputs[0]].shape
-            if len(beta_shape) != 1:
-                continue
-            input_ids3 = graph.edges_in[input_ids2[1]]
-            mean = graph.layers[input_ids3[0]]
-            mul_layer3 = graph.layers[input_ids3[1]]
-            if mean.kernel != "fluid.layers.create_parameter":
-                continue
-            axis = mul_layer3.attrs.get('axis', -1)
-            if axis != -1 and axis != 0:
-                continue
-            if len(graph.edges_out.get(input_ids3[0], [])) != 1:
-                continue
-            if len(graph.edges_out.get(input_ids3[1], [])) != 2:
-                continue
-            if mul_layer3.id != mul_layer1.id:
-                continue
-            if mean.outputs[0] not in graph.parameters:
-                continue
-            mean_shape = graph.parameters[mean.outputs[0]].shape
-            if mean_shape != beta_shape:
-                continue
-            input_ids4 = graph.edges_in[input_ids3[1]]
-            rsqrt_layer = graph.layers[input_ids4[0]]
-            gamma = graph.layers[input_ids4[1]]
-            if rsqrt_layer.kernel != "fluid.layers.rsqrt":
-                continue
-            if gamma.kernel != "fluid.layers.create_parameter":
-                continue
-            if len(graph.edges_out.get(input_ids4[0], [])) != 1:
-                continue
-            if len(graph.edges_out.get(input_ids4[1], [])) != 1:
-                continue
-            if gamma.outputs[0] not in graph.parameters:
-                continue
-            gamma_shape = graph.parameters[gamma.outputs[0]].shape
-            if gamma_shape != beta_shape:
-                continue
-            input_ids5 = graph.edges_in[input_ids4[0]]
-            add_layer = graph.layers[input_ids5[0]]
-            if add_layer.kernel != "fluid.layers.elementwise_add":
-                continue
-            axis = add_layer.attrs.get('axis', -1)
-            if axis != -1 and axis != 0:
-                continue
-            if len(graph.edges_out.get(input_ids5[0], [])) != 1:
-                continue
-            input_ids6 = graph.edges_in[input_ids5[0]]
-            variance = graph.layers[input_ids6[0]]
-            other = graph.layers[input_ids6[1]]
-            if variance.kernel != "fluid.layers.create_parameter":
-                continue
-            if other.kernel != "fluid.layers.fill_constant":
-                continue
-            if len(graph.edges_out.get(input_ids6[0], [])) != 1:
-                continue
-            if len(graph.edges_out.get(input_ids6[1], [])) != 1:
-                continue
-            if variance.outputs[0] not in graph.parameters:
-                continue
-            variance_shape = graph.parameters[variance.outputs[0]].shape
-            if variance_shape != beta_shape:
-                continue
-            ids = set([
-                layer_id, mul_layer0.id, sub_layer0.id, mul_layer1.id, beta.id,
-                mul_layer2.id, mean.id, mul_layer2.id, rsqrt_layer.id, gamma.id,
-                add_layer.id, variance.id, other.id
-            ])
-            for id in ids:
-                del graph.layers[id]
-                if id in graph.edges_in:
-                    del graph.edges_in[id]
-                if id in graph.edges_out:
-                    del graph.edges_out[id]
-            copy_layers = copy.deepcopy(graph.layers)
-            graph.layers = OrderedDict()
-            for k, v in copy_layers.items():
-                if k != nhwc_input.id:
-                    graph.layers[k] = v
-                    continue
-                graph.layers[k] = v
-                transpose0 = PaddleLayer(
-                    id='{}_1'.format(k),
-                    kernel="fluid.layers.transpose",
-                    inputs={"x": v.outputs[0]},
-                    outputs=["transpose_for_bn"],
-                    perm=[0, 3, 1, 2])
-                bn = PaddleLayer(
-                    id='{}_2'.format(k),
-                    kernel="fluid.layers.batch_norm",
-                    inputs={"input": "transpose_for_bn"},
-                    outputs=layer.outputs,
-                    epsilon=other.attrs["value"],
-                    param_attr="'{}'".format(gamma.outputs[0]),
-                    bias_attr="'{}'".format(beta.outputs[0]),
-                    moving_mean_name="'{}'".format(mean.outputs[0]),
-                    moving_variance_name="'{}'".format(variance.outputs[0]))
-                transpose1 = PaddleLayer(
-                    id=layer_id,
-                    kernel="fluid.layers.transpose",
-                    inputs={"x": layer.outputs[0]},
-                    outputs=layer.outputs,
-                    perm=[0, 2, 3, 1])
-                graph.layers[transpose0.id] = transpose0
-                graph.layers[bn.id] = bn
-                graph.layers[transpose1.id] = transpose1
-        graph.build()
--- a/x2paddle/optimizer/tensorflow/bias.py
+++ b/x2paddle/optimizer/tensorflow/bias.py
-import copy
-class BiasOpt:
-    def __init__(self):
-        self.conv_layers = [
-            'fluid.layers.conv2d', 'fluid.layers.conv2d_transpose'
-        ]
-    def run(self, graph):
-        print("Optimize: BiasOpt...")
-        layers = copy.deepcopy(graph.layers)
-        for layer_id, layer in layers.items():
-            if layer.kernel in self.conv_layers or layer.kernel == "fluid.layers.transpose":
-                if len(graph.edges_out.get(layer_id, [])) > 1:
-                    continue
-                if layer.outputs[0] in graph.outputs:
-                    continue
-                out_layer_id = graph.edges_out[layer_id][0]
-                if graph.layers[
-                        out_layer_id].kernel != "fluid.layers.elementwise_add":
-                    continue
-                if graph.layers[out_layer_id].attrs.get('axis', -1) != -1:
-                    continue
-                in_layer_id = graph.edges_in[out_layer_id]
-                bias_layer_id = in_layer_id[1 - in_layer_id.index(layer_id)]
-                if graph.layers[
-                        bias_layer_id].kernel != "fluid.layers.create_parameter":
-                    continue
-                bias_layer = graph.layers[bias_layer_id]
-                if len(bias_layer.attrs['shape']) != 1:
-                    continue
-                if len(graph.edges_out[bias_layer_id]) != 1:
-                    continue
-                if layer.kernel == "fluid.layers.transpose":
-                    if layer.attrs['perm'] != [0, 2, 3, 1]:
-                        continue
-                    in_layer_id = graph.edges_in[layer_id][0]
-                    if graph.layers[in_layer_id].kernel not in self.conv_layers:
-                        continue
-                    if graph.layers[in_layer_id].attrs['bias_attr'] != False:
-                        continue
-                    if graph.layers[in_layer_id].outputs[0] in graph.outputs:
-                        continue
-                    if len(graph.edges_out[in_layer_id]) != 1:
-                        continue
-                    graph.layers[in_layer_id].attrs[
-                        'bias_attr'] = bias_layer.attrs['name']
-                else:
-                    graph.layers[layer_id].attrs[
-                        'bias_attr'] = bias_layer.attrs['name']
-                bias_add_outs = graph.edges_out.get(out_layer_id, [])
-                bias_add_output = graph.layers[out_layer_id].outputs[0]
-                graph.del_layer(bias_layer_id)
-                graph.del_layer(out_layer_id)
-                for out in bias_add_outs:
-                    for k, v in graph.layers[out].inputs.items():
-                        if v == layer.outputs[0]:
-                            graph.layers[out].inputs[k] = bias_add_output
-                graph.layers[layer_id].outputs[0] = bias_add_output
-                if layer.kernel == "fluid.layers.transpose":
-                    in_layer_id = graph.edges_in[layer_id][0]
-                    graph.layers[in_layer_id].outputs[0] = bias_add_output
-                    graph.layers[layer_id].inputs['x'] = bias_add_output
--- a/x2paddle/optimizer/tensorflow/prelu.py
+++ b/x2paddle/optimizer/tensorflow/prelu.py
-import copy
-import numpy as np
-from collections import OrderedDict
-from x2paddle.core.program import PaddleLayer
-from x2paddle.core.util import *
-class PReLUOpt:
-    def __init__(self):
-        pass
-    def run(self, graph):
-        print("Optimize: PReLUOpt...")
-        layers = copy.deepcopy(graph.layers)
-        for layer_id, layer in layers.items():
-            if layer.kernel != "fluid.layers.elementwise_add":
-                continue
-            axis = layer.attrs.get('axis', -1)
-            if axis != -1 and axis != 3:
-                continue
-            input_ids0 = graph.edges_in[layer_id]
-            relu_layer0 = graph.layers[input_ids0[0]]
-            mul_layer0 = graph.layers[input_ids0[1]]
-            if relu_layer0.kernel != "fluid.layers.relu":
-                continue
-            if mul_layer0.kernel != "fluid.layers.elementwise_mul":
-                continue
-            axis = mul_layer0.attrs.get('axis', -1)
-            if axis != -1 and axis != 3:
-                continue
-            if len(graph.edges_out.get(input_ids0[0], [])) != 1:
-                continue
-            if len(graph.edges_out.get(input_ids0[1], [])) != 1:
-                continue
-            input_ids1_0 = graph.edges_in[input_ids0[0]]
-            input_ids1_1 = graph.edges_in[input_ids0[1]]
-            fill_layer = graph.layers[input_ids1_1[1]]
-            mul_layer1 = graph.layers[input_ids1_1[0]]
-            if fill_layer.kernel != "fluid.layers.fill_constant":
-                continue
-            if mul_layer1.kernel != "fluid.layers.elementwise_mul":
-                continue
-            axis = mul_layer1.attrs.get('axis', -1)
-            if axis != -1 and axis != 0:
-                continue
-            if len(graph.edges_out.get(input_ids1_1[1], [])) != 1:
-                continue
-            if len(graph.edges_out.get(input_ids1_0[0], [])) != 3:
-                continue     
-            input_ids2 = graph.edges_in[input_ids1_1[0]]    
-            alpha = graph.layers[input_ids2[0]]
-            sub_layer = graph.layers[input_ids2[1]]
-            if alpha.kernel != "fluid.layers.create_parameter":
-                continue
-            if sub_layer.kernel != "fluid.layers.elementwise_sub":
-                continue
-            axis = sub_layer.attrs.get('axis', -1)
-            if axis != -1 and axis != 3:
-                continue
-            if len(graph.edges_out.get(input_ids2[0], [])) != 1:
-                continue
-            if len(graph.edges_out.get(input_ids2[1], [])) != 1:
-                continue
-            if alpha.outputs[0] not in graph.parameters:
-                continue
-            input_ids3 = graph.edges_in[input_ids2[1]]
-            add_layer = graph.layers[input_ids3[0]]
-            abs_layer = graph.layers[input_ids3[1]]
-            if abs_layer.kernel != "fluid.layers.abs":
-                continue
-            if len(graph.edges_out.get(input_ids3[1], [])) != 1:
-                continue
-            ids = set([
-                layer.id, relu_layer0.id, mul_layer0.id, fill_layer.id, mul_layer1.id, alpha.id,
-                sub_layer.id, abs_layer.id])
-            for id in ids:
-                del graph.layers[id]
-                if id in graph.edges_in:
-                    del graph.edges_in[id]
-                if id in graph.edges_out:
-                    del graph.edges_out[id]
-            copy_layers = copy.deepcopy(graph.layers)
-            graph.layers = OrderedDict()
-            for k, v in copy_layers.items():
-                if k != add_layer.id:
-                    graph.layers[k] = v
-                    continue
-                graph.layers[k] = v
-                transpose0 = PaddleLayer(
-                    id='{}_1'.format(k),
-                    kernel="fluid.layers.transpose",
-                    inputs={"x": v.outputs[0]},
-                    outputs=["transpose_for_prelu"],
-                    perm=[0, 3, 1, 2])
-                prelu = PaddleLayer(
-                    id='{}_2'.format(k),
-                    kernel="fluid.layers.prelu",
-                    inputs={"x": "transpose_for_prelu"},
-                    outputs=layer.outputs,
-                    mode=string("channel"),
-                    param_attr="'{}'".format(alpha.outputs[0]))
-                transpose1 = PaddleLayer(
-                    id=layer_id,
-                    kernel="fluid.layers.transpose",
-                    inputs={"x": layer.outputs[0]},
-                    outputs=layer.outputs,
-                    perm=[0, 2, 3, 1])
-                graph.layers[transpose0.id] = transpose0
-                graph.layers[prelu.id] = prelu
-                graph.layers[transpose1.id] = transpose1
-                first_axis = graph.parameters[alpha.outputs[0]].shape[0]
-                graph.parameters[alpha.outputs[0]] = np.reshape(graph.parameters[alpha.outputs[0]], (1, first_axis, 1, 1))
-        graph.build()
\ No newline at end of file