On 2018/03/23 05:49:07, sandeep krishnamurthy <sandeep.krishn...@gmail.com>
wrote:
> Hello MXNet Community,
>
> Along with Lai, Karan and other MXNet contributors, I am working on adding
> MXNet backend for Keras. Currently supporting around ~70% of Keras APIs
> across CNNs and RNNs.
> https://github.com/deep-learning-tools/keras/tree/keras2_mxnet_backend
>
> We wanted to gather the community feedback on the proposal for including
> this keras-mxnet package as a submodule in Apache MXNet. This will enable
> providing the Keras interface for MXNet users. MXNet users can choose Keras
> interface for building their Neural Networks in Symbolic Mode (Ex:
> mx.keras).
>
> *Advantages:*
>
> 1. Keras is widely popular interface that many DL practitioners are
> familiar. By including keras interface within MXNet natively, we enable
> many users to use MXNet with 0 learning curve.
>
> 2. Adding as submodule and exposing natively within MXNet pip package,
> would greatly enhance user experience and get more users as compared to
> releasing a fork repository independently.
>
> 3. Why submodule? - Helps in easily managing with patching the latest
> parent keras-team/keras developments and releases. Thereby helping us
> provide users the core keras experience. Operational management.
>
> 4. Other minor advantages - Operational maintenance, pip, CI and quality
> control.
>
> Please do share your comments on the proposal.
>
> Best,
> Sandeep
>
> *Note: *We tried merging with keras-team/keras and we created a PR
> <https://github.com/keras-team/keras/pull/9291> as well. However, due to
> multiple design incompatibility challenges, we need significant re-work on
> MXNet Module, KVStore, Optimizers to address keras-team design concerns.
> Since, we are adhering to keras API interface exposed to users, we are
> planning release on the forked repo for now. More details on the design
> challenges and workaround tried -
> https://docs.google.com/document/d/1Vn5ip5MzCKcN29KCCnwjB2d59y-VevdLrdn_eNd3nE4/edit?usp=sharing
> -1.
Submodules implies that they are to be used as dependencies. Apparently it is
not the case.
