This tool converts pytorch model to Caffe model by ONNX
only use for inference

• python 3.5.2
• pycaffe (with python3 support)
• pytorch 1.0.0
• onnx 1.4.1
• protobuf 3.6.1
• pyhocon 0.3.50 (for config)

To convert onnx model to caffe:

python convertCaffe.py --conf-path ${CONFIG}.hocon

• Conv
• ConvTranspose
• BatchNormalization
• MaxPool
• AveragePool
• Relu
• Sigmoid
• Dropout
• Gemm (InnerProduct only)
• Add
• Mul
• Reshape
• Upsample
• Flatten
• PRelu

• Remove Constant, Shape, Unsqueeze, Squeeze ops in onnx and concatenate prototxt after removing these ops
• Transfer attribute value of Constant to Reshape
• Concat layer
  • the common usage of pytorch view operation, the Concat before the Reshape is redundant for caffe
• merge batchnormization to convolution
• merge scale to convolutionv