Model zoo for TensorFlow.js visualizer

[lutzroeder]

Model zoo for TensorFlow.js visualizer

[caisq]

@lutzroeder Thanks for this question. TensorFlow.js can currently load models from keras (hence model.json) and tensorflow (hence .pb). Are there any existing model zoos for (python) keras and tensorflow models? If so, I guess I don't fully understand why we need a separate zoo for TensorFlow.js.

On the flip side, it is currently possible to export the topology of TensorFlow.js models in the keras model.json format, but since those models are essentially keras models, they can just use whatever existing keras model zoo there is. Another question I have: does Netron load keras HDF5 files?

[lutzroeder]

Mostly looking for some typical tfjs models or corner cases. model.json files are not exactly the same as Keras .json files. For example, instead of model_config at the root there is another modelTopology object inserted at the root. Has there been some discussion to make them a strict superset of Keras .json files with model_config at the root and adding the weight information at the same level? Instead of a converter, wouldn't it make sense to use the Keras .json format directly or even save from Keras? Also, given other frameworks like MXNet use .json, some discussion around using file extensions or how to detect different formats would be helpful.

Yes, Netron does load Keras .h5 models.

[davidsoergel]

Yep, the question of standardizing the serialization format is under active discussion. The current model.json should be considered a TF.js implementation detail for now. Good point about the file extension, thanks!

[nsthorat]

Closing this out -- any discussions can move over to the community discussion mailing list.