/MXNet2Caffe

Convert MXNet model to Caffe model

Primary LanguagePythonMIT LicenseMIT

MXNet2Caffe: Convert MXNet model to Caffe model

You are welcome to file issues either for bugs in the source code, feature requests!

Brief Guide

Given a MXNet Model, MXNet2Caffe can automatically generate both .prototxt and .caffemodel.

Before starting using MXNet2Caffe, you need to manually set the path in find_caffe.py and find_mxnet.py.

After that, simply run python json2prototxt.py to generate the corresponding .prototxt.

And then, using python mxnet2caffe.py to generate the corresponding .caffemodel.

TODO

[1] Since there is no Flatten layer in caffe, you have to manually modify the automatically generated .prototxt. In other words, you have to change the bottom of the layer just after the Flatten layer making it linking to the layer before the Flatten layer. Currently, this part has to be done manually.

[2] The converted model scores a little bit worse than the original MXNet model.

[3] Code for automatically reversing the weight (and bias) of the first layer to support BGR input.

[4] Better support for caffe's in-place feature.

[5] Several TODOs in prototxt_basic.py