One of the questions we frequently encounter in forums and emails is “I have caffe code that does this. How do I do the same thing in MXNet”. While it might be possible to go through the documentation and do the translation layer by layer manually, we thought a tool that automatically does it will be helpful for people who want to migrate from Caffe to MXNet. Here is the first version of the tool: https://github.com/apache/incubator-mxnet/tree/master/tools/caffe_translator. Thanks to Aaron (@aaronmarkham), Madan (@madjam), Pracheer (@pracheer) and Steffen (@srochel) for their help and feedback in building this.
Note that this is different from the caffe converter that already exists ( https://github.com/apache/incubator-mxnet/tree/master/tools/caffe_converter). Caffe converter converts Caffe Model to MXNet model so that user can run inference on MXNet or finetune on MXNet. The new translator on the other hand helps users to migrate their Caffe code to MXNet and continue development on MXNet. I think both tools are useful in different contexts. Caffe Translator can currently translate the layers mentioned in the README <https://github.com/apache/incubator-mxnet/blob/master/tools/caffe_translator/README.md>. Obviously this is only a humble beginning and there is much more layers that can be supported. Apart from adding support for more layers, following are some things that can make the tool more useful: - Eliminate the CaffeDataIter <https://mxnet.incubator.apache.org/how_to/caffe.html#use-io-caffedataiter> dependency to translate data layers. - Generate code that can do distributed training. This is one of main reasons I might want to move my Caffe code to MXNet. - Make the tool accessible through a website so that user can just copy paste a Caffe code snippet in the website and get translated code instantly without downloading and running anything locally. Any help in these efforts is greatly appreciated. Also let me know if there is some other improvement that could be done to make the tool more useful. Thanks, Indu
