Add a flag to tfjs-converter: --split_weights_by_layer

[dsmilkov]

Problem
We ran across a Keras model that has 174 tiny layers, resulting in 174 weight files of the form group*-shard1of1. This slows down model loading.

Proposed solution
By default, we should create only one weight group and add an optional boolean flag to the tfjs_converter script --split_files_per_layer.

tjs_converter --split_files_per_layer ...
# Will create a weight group for each layer

This flag will only be valid for a Keras model. Other suggestions welcome!

[nsthorat]

I actually wonder if this should be the default (not putting each layer in their own group). If someone wants to break up weights by layer, they can choose to do that explicitly. I think the case of downloading a subset of weights is much less important than not shipping 174 separate files.

And relevant link where this needs to be changed:
https://github.com/tensorflow/tfjs-converter/blob/master/python/tensorflowjs/converters/keras_h5_conversion.py#L172-L177

[dsmilkov]

I agree with you! You have to be an advanced user with a unique use-case to want to split per layer :). Modified the suggestion.

[caisq]

@dsmilkov Woah - 174 files is obviously too many. I agree this needs to be fixed. Thanks for filing this issue. We'll prioritize this.