Support for AutoML Vision Edge models with sharded weights in React Native

TensorFlow.js version

@tensorflow/tfjs : 1.7.3 @tensorflow/tfjs-react-native : 0.2.3

Browser version

expo : 37.0.8

Describe the problem or feature request

Since you:

• Can’t load sharded weights using the bundleResourceIO function provided in tfjs-react-native.
• Can’t run your own tensorflowjs conversion to change the weight_shard_size_bytes on a TF Saved Model exported from AutoML.

You get caught at an impasse when trying to use an AutoML Vision Edge model through the React Native APIs. It seems to me the options for someone with this issue are:

• Attain the ability to turn weight sharding off within AutoML’s Cloud UI.
• Attain a way to reproduce the exact tensorflowjs_converter command that AutoML runs on their end. Or however they make that conversion.
• Attain a way to load multiple shards within the tfjs-react-native bundleResourceIO.
• Attain a way to rejoin shards. I’m out of my depth, would that even be possible? I’ve experimented with taking non-sharded weights from my own conversion runs and splicing them into the model.json produced by AutoML with no luck.

Code to reproduce the bug / link to feature request

There’s not too much relevant code to show since it’s explicitly unsupported in tfjs-react-native. However, fwiw, here’s the conversion code I attempted to use when I realized direct support for the artifacts produced by AutoML tfjs export were not usable.

# This produces significantly different results than what is produced through the AutoML tfjs export.
tensorflowjs_converter \
  --input_format=tf_saved_model \
  --output_format=tfjs_graph_model \
  --weight_shard_size_bytes=10485760 \
  --skip_op_check \
  00000123 \
  test

[tafsiri]

Closing, sharded models are now supported in tfjs-react-native https://github.com/tensorflow/tfjs/pull/4113

We also updates the instructions on the automl README to describe how to load a model with a custom IOHandler (i.e. something other than a string).

[pyu10055]

@tafsiri after thinking about this more, it might make sense to provide graph_model to graph_model conversion, the allowed features could include sharding and quantization. I will look into that.