MXNet2Caffe: Convert MXNet model to Caffe model,And use Tensor-RT Plugin_Class to improve performance (SEnet)

• caffe_plugin_layer

• mxnet2caffe

• caffe2TensorRT

• TensorRT_plugin_layer

  Provide mxnet to caffe conversion tool,currently supports Conv、BN、Elemwise、Concat、Pooling、Flatten、 Cast、Fully、Slice、L2、Reshape、Broadcast etc. And then use the TensorRT(4.0) engine to parse the caffe model to improve performance.The project was successfully tested on SEnet.

Code

• json2prototxt.py prototxt_basic.py Read mxnet_json file and converte to prototxt
• mxnet2caffe.py Read mxnet_model params_dict and converte to .caffemodel
• mxnet_test.py Debug mxnet output and you can compare the result with the converted caffemodel
• load_param.py Print mxnet or caffe model param_dict
• caffe_plugin_layer Add caffe plugin_layer, but forward computing implementation in Tensorrt_RT Plugin layer
• Tensor_RT/Tensor_RT_Plugin.cpp Add Brocast_layer、Pooling and debug_layer in Tensor_RT by using IPluginExt API

caffe_plugin_layer

The caffe framework does't support many Layers_op. If you need to convert a special layer, you need to add a layer plugin in the caffe framework and register the Layer_param. Otherwise, you will get an error when building the network structure using this tool.

• Add layer implementations in the caffe/src/caffe/layers/ directory, such as broadcast_mul.cpp broadcast_mul.cu,You need to provide forward and reverse operations.
• Add a layer declaration in the caffe/include/caffe/layers/ directory, such as broadcast_mul.hpp
• Register Layer_param in caffe/src/caffe/proto/caffe.proto (read the parameters required for the Layer from prototxt)

message LayerParameter {
  optional string name = 1; // the layer name
  optional string type = 2; // the layer type
  repeated string bottom = 3; // the name of each bottom blob
  repeated string top = 4; // the name of each top blob
  ......
  
  optional BroadcastmulParameter broadcastmul_param = 230;
}

message BroadcastmulParameter {  
   
} 

• Re-make build libcaffe.so
• sudo make pycaffe recompile python interface ***

mxnet2caffe

• First,you should run json2prototxt.py to convert the structure(json) of the mxnet model to prototxt format.
• Then, run mxnet2caffe.py to read the prototxt network structure build using the API provided by pycaffe.
• Using mxnet_Api Read the param parameter saved by mxnet, and copy the parameter key value pair into the caffemodel file;

caffe2TensorRT and TensorRT_plugin_layer

The tensorrt engine can directly parse the caffe model. For unsupported ops, you can manually add them using the interface. This project adds a broadcast operation. In addition, it also tests the Pooling_layer，It also adds a test layer, which can separately print the parameters in the structure and assist the Debug.

• Firsy,Inherit the IPluginExt interface to create a custom layer class

• Create a PluginFactory function that will be used to add custom layer classes to the network.

  virtual int getNbOutputs() const = 0;

  virtual Dims getOutputDimensions(int index, const Dims* inputs, int nbInputDims) = 0;

  virtual void configure(const Dims* inputDims, int nbInputs, const Dims* outputDims, int nbOutputs, int maxBatchSize) = 0;

  virtual int initialize() = 0;

  virtual void terminate() = 0;

  virtual size_t getWorkspaceSize(int maxBatchSize) const = 0;

  virtual int enqueue(int batchSize, const void*const * inputs, void** outputs, void* workspace, cudaStream_t stream) = 0;

  virtual size_t getSerializationSize() = 0;

  virtual void serialize(void* buffer) = 0;

This project provide Pooling, broadcast, and test_layer,you can see the implementation in the code.

TODO:

• Tensor RT supported Se_Resnet