[Feature Request] Add support for grouping convolution filters

[mkaze]

Currently, the support for grouping convolution filters in convolution layers is missing. For more information, you can read the groups argument description of Conv2D layer in Keras here, or in PyTorch here.

[mkaze]

Though, currently the underlying TF Java API does not support grouped convolutions. Alternatively, this could be implemented by grouping the input channels, applying filters separately and then concatenating the results.

[zaleslaw]

Agree, that this functionality is missed, the initial idea to make MVP DL framework - to implement building blocks required to make parse and load CV models from the Keras.applications. As I understand, that groups are not used in the SOTA models, but maybe you could share architecture, competition or abstract/paper where it was used. I mean, do we really need this functionality?

[mkaze]

@zaleslaw I think ResNeXt used grouped convolutions; but I agree that it's not a crucial feature to have at this moment, especially for an MVP.

[zaleslaw]

If it is used in ResNeXt - it could make sense, the problem with ResNeXt is the following: it's not on the official list of pre-trained Keras models yet, but the model is created and the PR is proposed. I subscribed to these issues and PR to wait for the official release of the Resnext model.

But if you are interested in this task, feel free to share ideas and draft here.

[zaleslaw]

P.S Read the Contributing Guidelines.