MMDNN

Pytorch To Caffe

1. mmdnn目前只支持pytorch0.4.0
2. 网络前向尽量不要使用中间变量（下面是打印的node）
   
3. 支持的层：

   layer_map = {
   'onnx::Conv': 'Conv',
   'onnx::Flatten': 'Flatten',
   'onnx::Gemm': 'FullyConnected',
   'onnx::MaxPool': 'Maxpool',
   'onnx::AveragePool': 'Avgpool',
   'onnx::Dropout': 'Dropout',
   'onnx::BatchNormalization': 'BatchNormalization',
   'onnx::Add': 'Add',
   'onnx::Concat': 'Concat',
   'onnx::Relu': 'Relu',
   'onnx::Tanh': 'Tanh',
   'onnx::Sigmoid': 'Sigmoid',
   'onnx::Mul': 'Mul'
   
   
   # TODO
   # 'max_pool2d': convert_maxpool,
   # 'onnx::Mul': convert_elementwise_mul,
   # 'onnx::Sub': convert_elementwise_sub,
   # 'onnx::ConvTranspose': convert_convtranspose,
   # 'onnx::LeakyRelu': convert_lrelu,
   # 'onnx::Sigmoid': convert_sigmoid,
   # 'onnx::Softmax': convert_softmax,
   # 'onnx::Selu': convert_selu,
   # 'onnx::Transpose': convert_transpose,
   # 'onnx::Reshape': convert_reshape,
   # 'onnx::MatMul': convert_matmul,
   # 'onnx::Gather': convert_gather,
   # 'onnx::ReduceSum': convert_reduce_sum,
   # 'onnx::Constant': convert_constant,
   # 'onnx::Upsample': convert_upsample,
   # 'onnx::Pad': convert_padding,
   }
   

4. 修改pytorch_parser.py，69-78行，如下：

   # if not os.path.exists(model_file_name):
   #     print("Pytorch model file [{}] is not found.".format(model_file_name))
   #     assert False
   # # test
   
   # # cpu: https://github.com/pytorch/pytorch/issues/5286
   # try:
   #     model = torch.load(model_file_name)
   # except:
   #     model = torch.load(model_file_name, map_location='cpu')
   
   model = model_file_name
   

5. 转IR

   size = 112
   from mmdnn.conversion.pytorch.pytorch_parser import PytorchParser
   pytorchparser = PytorchParser(net, [3, size, size])
   IR_file = 'PFLD'
   pytorchparser.run(IR_file)
   

6. convert IR to Caffe Code

 mmtocode -f caffe -n PFLD.pb -w PFLD.npy -d caffe_converted.py -dw caffe_converted.npy

7. get the caffe model

 mmtomodel -f caffe -in caffe_converted.py -iw caffe_converted.npy -o caffe_target

Caffe model files are saved as [caffe_target.prototxt] and [caffe_target.caffemodel], generated by [caffe_converted.py] and [caffe_converted.npy].