Convert pytorch to Caffe by ONNX

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

Update

20230315: change pool roundmethod to floor, which pytorch used. add centernet example, pretrained from mmdetection.

Dependencies

• caffe (with python support)
• pytorch 0.4 (optional if you only want to convert onnx)
• onnx

install

git clone https://github.com/chenjun2hao/onnx2caffe.git
python setup.py install

or install on develop mode
python setup.py develop

How to use

just run in terminal

onnx2caffe ./model/MobileNetV2.onnx ./model/MobileNetV2.prototxt ./model/MobileNetV2.caffemodel

Current support operation

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

TODO List

• support all onnx operations (which is impossible)
• merge batchnormization to convolution
• merge scale to convolution