Update to latest MXnet (v1.8.x)

[tjad]

I'd like to get involved in updating this project.

Could a list of known tasks be compiled here please.

[tjad]

@mrkn I can confirm that all existing functionality still works on 1.7.0. I have run some training using 1.7.0

This in preparation for 1.8.x update. I guess all that really needs to be done is extending the gluon api modules

[tjad]

I have further bumped versions in the ci. Perhaps you could publish a new docker image to facilitate those versions

[mrkn]

@tjad Do you want to manage this project? If you can do it, I want to migrate this to red-data-tools organization and give you admin permission.

[tjad]

@mrkn Sounds like a plan. I'm eager to push the Ruby ML community forward with regards to tools.

As an aside, do you have considerations for proposing incubator-mxnet adopt this project, to officially support Ruby? I think they'll be pleased to have a 9th language aboard, if we can get the project quite updated (especially if it supports Gluon well - seeing as they want Gluon to be the framework of choice for "all languages" )

[toddsundsted]

i worked on parts of the gluon support. i'd be happy to assist as well.

[mrkn]

@tjad Sounds great. I'm OK to donate this repository to incubator-mxnet. Are the things below only remaining work to be done before the donation?

• Catching up the latest MXNet version.
• Fixing GC-related bugs like Issue #45 : GC mark param keys and values #48.
• Gluon support.

[tjad]

@toddsundsted Awesome!

@mrkn Yeah, I suppose they'd still rely on Red Data team to keep this repository updated anyway.
As far as I can tell, yes completing above list would put this repository in great stead for MXNet adoption.

Still working towards proper solution for #48

[tjad]

@toddsundsted I think I mostly understood the transfer from python to ruby. It's simple and great work! And I hope we can push it forward for greater use.

I've added Adam optimiser in a optimiser branch. Perhaps we could further branch out into the relevant modules of Gluon such as nn, rnn, loss etc, then we can tackle outstanding functionality and merge back into master as completed.

[tjad]

I've opened #49 and #50 to help with coordination of tasks regarding coverage. When you're tackling a task, claim it by mentioning the item in the respective issue. As this is mostly a code transfer, we can emphasise on the implementations first, before test coverage is created. Other issues may arise out of implementing this coverage, such as stype support

[tjad]

I think we can put the KV store coverage into the above list too
WIP
6c8df1b

[tjad]

Contrib modules we can defer and prioritise as more users require. I think it'd be good to include ONNX support for completion.