Tracking issue for all layers

[keras-team]

Tracking issue for all layers

[AakashKumarNain]

I am working on Merging layers.

[fchollet]

@fchollet I have one question regarding the data adapters for Pytorch. Pytorch recommends overriding Dataset which is similar to our PyDataset. Should we leverage the same for Pytorch or do you recommend implementing with the base data_adapter interface?

The general idea is that it should be possible to take an existing PyTorch data pipeline (using PyTorch Dataset and PyTorch dataloaders) and pass it to a Keras model .fit() call.

So we should not assume that the user will need to change anything to their pipeline. It should work with what they currently have.

[AakashKumarNain]

Makes sense. Thanks. I will try to come up with a good solution

[AakashKumarNain]

Is anyone working on the Activations layers in this list? If not, I will take up this task

[fchollet]

Hi Aakash -- yes, activations are already assigned. All layers are currently assigned except preprocessing layers (and FeatureSpace, not quite a layer), which are a bit special because we're not going to reimplement them (they can only run in TF), rather we're going to wrap them.

You could be working on preprocessing layers, or alternatively you could be writing a DataAdapter for a PyTorch Dataset/Dataloader...

[AakashKumarNain]

I will start with the DataAdapters in that case. Thank you @fchollet

[AakashKumarNain]

@fchollet I have one question regarding the data adapters for Pytorch. Pytorch recommends overriding Dataset which is similar to our PyDataset. Should we leverage the same for Pytorch or do you recommend implementing with the base data_adapter interface?