BatchNorm, Convolution and InnerProduct layers are supported.
To compile, change the paths from config.mk to point to your MXNet and Caffe libraries. In the Makefile, maybe some caffe dependencies must also be changed. Then, run "make" from the terminal.
The program needs these parameters:
./mxnet_to_caffe <mxnet_json> <mxnet_model> <caffe_prototxt> <caffe_model_output>
As an example, you can download the Inception21K model from:
http://data.dmlc.ml/mxnet/models/imagenet/inception-21k.tar.gz
And its corresponding Caffe deploy.prototxt from:
https://github.com/pertusa/InceptionBN-21K-for-Caffe
Then, run:
./mxnet_to_caffe Inception-symbol.json Inception-0009.params deploy.prototxt Inception21k.caffemodel
And the file "Inception21k.caffemodel" will be generated with the Caffe weights.
This code is based on the CXXNet to Caffe converter (https://github.com/n3011/cxxnet_converter).
License: GNU Public license