A reimplementation of MixNet in Gluon.

[becauseofAI]

Add a new model named MixNet to gluoncv model-zoo.
It is a reimplementation of MixNet in Gluon. The paper is here. The model will be trained with ImageNet dataset when the machine is ready.

[mli]

Job PR-900-1 is done.
Docs are uploaded to http://gluon-vision-staging.s3-website-us-west-2.amazonaws.com/PR-900/1/index.html
Code coverage of this PR: vs. Master:

[bryanyzhu]

PR looks awesome. Minor things about removing commented lines. Looking forward to the numbers trained on ImageNet.

[mli]

Job PR-900-2 is done.
Docs are uploaded to http://gluon-vision-staging.s3-website-us-west-2.amazonaws.com/PR-900/2/index.html
Code coverage of this PR: vs. Master:

[hetong007]

I'd like to hold this implementation from merged into our incoming release, unless one of the following can be met:

1. the weights are transferred from the official source, or
2. the network can be trained to match the accuracy.

The reason behind is due to MXNet has a limited implementation of split, which cannot split the channel dimension unevenly. So this might not be an identical re-implementation. Still, I'm open with further suggestions.

[zhreshold]

@hetong007 split may not support uneven division, but @becauseofAI used workaround which uses F.slice_axis with pre-calculated brakets. So I guess it's matching the original implementation?
@becauseofAI Can you verify if you can convert weights and achieve same accuracy as reported in paper? 

[hetong007]

OK you are right, F.slice_axis makes it equivalent.

[zhreshold]

@becauseofAI any update on this?

[djaym7]

@becauseofAI any update on this?

[djaym7]

[total_filters // num_groups] * num_groups

Much faster