This tool converts caffe model convert to onnx model
only use for inference

This is the second version of converting caffe model to onnx model. In this version, all the parameters will be transformed to tensor and tensor value info when reading .caffemodel file and each operator node is constructed directly into the type of NodeProto in onnx.

• protobuf
• onnx==1.4.0

$ pip install -r requirements.txt

( caffe environment is not required! )

usage: convert2onnx.py [-h] [caffe_graph_path] [caffe_params_path] [onnx_name] [save_dir]
positional arguments:
  caffe_graph_path          caffe's prototxt file path
  caffe_params_path         caffe's caffemodel file path
  onnx_name                 onnx model name
  save_dir                  onnx model file saved path

Take ResNet-50 as an example, you can follow the instructions.

1. Download resnet50 .caffemodel file from BaiduDisk and put resnet-50-model.caffemodel to ./caffemodel/resnet-50/
   Link：https://pan.baidu.com/s/10YB42muAd0vGiNTCetvLsA
   Code：7az4

2. Convert resnet50 caffe model to onnx model

   $ python convert2onnx.py \
             caffemodel/resnet-50/resnet-50-model.prototxt \
             caffemodel/resnet-50/resnet-50-model.caffemodel \
             resnet50 onnxmodel

3. Visualize onnx model by netron

   $ netron onnxmodel/resnet50.onnx --host 0.0.0.0 --port 8008

4. Run test scripts

   $ python onnxmodel/test_resnet.py \
             --input_shape 224 224 \
             --img_path onnxmodel/airplane.jpg \
             --onnx_path onnxmodel/resnet50.onnx
   # you will get result 404 which is the class id of airplane in IMAGENET.

5. If you have custom layers in caffe which makes your caffe.proto is different than the one in the origin caffe code. The things you should do before convertion is:

   • First of all, compile your proto file with protoc

     # for example
     $ protoc /your/path/to/caffe_ssd.proto --python_out ./proto

   • Then specify the caffe proto file by replacing the line from proto import caffe_upsample_pb2 as caffe_pb2 with your module.

BatchNorm
Convolution
Deconvolution
Concat
Dropout
InnerProduct(Reshape+Gemm)
LRN
Pooling
Unpooling
ReLU
Softmax
Eltwise
Upsample
Scale

• Resnet50
• AlexNet
• Agenet
• Yolo V3
• vgg16

netron is recommended: https://github.com/lutzroeder/netron
netron Browser

See Develop Guide