Hello Community,
I am working on adding Apache MXNet backend for Keras2. Jiajie Chen, MXNet contributor, was actively involved and has contributed many operators and functionalities. Thanks to early contributors for Keras1.2 which we have extended in this work. I hereby request for Apache MXNet community to engage, contribute and early testing the in-progress work. *Current State:* You can train a Keras MLP and CNNs on CPU, GPU, multi-GPU using MXNet as backend. However, a lot of functionalities to add (see next section). Here is a quick install guide to try it out - https://github.com/deep-learning-tools/keras/wiki/Installation-Guide---Keras-with-MXNet-backend *Planned Next Steps:* - Create PR with keras-team/keras Github repo and submit the Apache MXNet backend contributions. I am working fixing couple of operators and broken keras test to create PR. Hopefully, in 2 weeks, the first PR with MLP and CNN support will be created. *Call for contributions:* There are many upcoming pending works and I request the community to engage and contribute. At a high level, following are the features we need to work on: - Enable multi-machine distributed training for Keras with MXNet backend. - Support RNNs for Keras with MXNet backend. - Support for Sparse Tensors for Keras with MXNet backend. - Other operators. - Try and provide your feedback or open Github issues. *Links:* I have created Github issues listing known issues, TODOs and more, in the work in progress repo - MXNet backend Work in progress repository - https://github.com/deep-learning-tools/keras/tree/keras2_mxnet_backend - https://github.com/deep-learning-tools/keras/issues - MXNet Tracking Issue - https://github.com/apache/incubator-mxnet/issues/8941 - Keras Tracking Issue - https://github.com/keras-team/keras/issues/8697 Best Regards, Sandeep Krishnamurthy
