caffe2onnx

This tool converts Caffe models to ONNX via command line (without Caffe environment).

Introduction

caffe2onnx is based on previous work by htshinichi.

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.

Installation

Install latest from github

pip install git+https://github.com/kumardesappan/caffe2onnx

Build and install latest from source

git clone https://github.com/kumardesappan/caffe2onnx
python setup.py install

Usage

To get started with caffe2onnx, run the caffe2onnx.convert command, providing:

• the path to your caffe prototxt,
• the path to your caffe model (not required),
• the output path of the onnx model (not required),
• frozen graph or not (not required).

python -m caffe2onnx.convert
    --prototxt          caffe's prototxt file path
    [--caffemodel       caffe's caffemodel file path]
    [--onnx             output onnx file path]
    [--frozen           frozen graph or not]

Operators

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

Models

• ResNet
• AlexNet
• AgeNet
• Yolo V3
• VGG
• GoogleNet
• CaffeNet
• Sphereface

License

BSD-3