Add functionality to give custom input to model (interactive mode)

googleinterns / amaranth

Apache License 2.0

2 stars 0 forks source link

Add functionality to give custom input to model (interactive mode) #10

Closed tommylau-exe closed 4 years ago

tommylau-exe commented 4 years ago

This enhancement will make it possible to see the ML model's output for any given string of text. This would be incredibly useful for sanity checks and future interactivity.

tommylau-exe commented 4 years ago

Relevant (originally from #12):

Not sure how to encode strings into a format the model will accept after it has been saved and reloaded in a different environment. May have to find some way to serialize this as well, through the model or otherwise.

tommylau-exe commented 4 years ago

Keras has an experimental pre-processing layer called TextVectorization that may work here. The TextVectorization layer has some advantages that make it a great fit for this problem:

Would reduce complex custom logic in amaranth/ml/train.py
Tightly couples text vectorization with model
Looks to be serializable, like all other Keras layers, fixing the reloading problem

Of course it's not a perfect solution:

This layer is experimental
May not perform well
May have bugs

It's worth noting that the layer has been experimental since at least November of 2019, so it may be somewhat stable by now.

tommylau-exe commented 4 years ago

After some testing, the experimental TextVectorization layer from Keras is definitely serializable, and works exceptionally well in this project. One issue I ran into was a ValueError being thrown when the model is read back in after serialization. Looks like it's already been solved here. Unfortunately though, this fix isn't available in the latest stable Tensorflow release. We may have to switch to one of the nightly builds to get this to work properly.