Awesome

Torch2Caffe2

Python and shell script to convert Torch models into Caffe2 protobuf files. The conversion uses ONNX-Caffe2 backend and has three steps like: Torch (.t7) -> PyTorch (.pth) -> ONNX (.onnx.pb) -> Caffe2 (.pb)

IMPORTANT: You need to specify the input dimension of the model.

Source Download

git clone --recursive git@github.com:ca1773130n/Torch2Caffe2.git

Requirements

• Python 3.5
• PyTorch
• ONNX, ONNX-Caffe2
• Caffe2 (optional)

NOTE:

• If you don't have those in your environment yet, follow the steps below:
• https://github.com/caffe2/caffe2/issues/1204

Usage

torch_to_caffe2.sh <model_filename> <input_dimension>

• The model file must be in the same directory of the shell script
• The input dimension needs to be surrounded by quotes. It should look like this: "(1,3,512,512)"

Acknowledgement

This work is based on a GitHub project below.

• https://github.com/clcarwin/convert_torch_to_pytorch

Tested environment

• Ubuntu 16.04
• Anaconda Python 3.6
• A Torch VGG-19 model file

Known Issues

• PyTorch to ONNX conversion is imperfect (e.g. ONNX doesn't support ceil_mode in MaxPooling)