🧹 Cleanup of DCGAN & CVAE

[lindapaiste]

Extends PR #1376 -- merge that first.
Fixes #949
Possible resolves #1386, subject to discussion.

CVAE

• generate() accepts an optional latentVector (inspired by DCGAN).
• Throw an error if the label cannot be found instead of returning nonsensical garbage [undefined, undefined].
• Remove hard-coded input size of [1, 16] as this will vary based on the model. Accesses the input shape by looking at the model instead.
• Remove unnecessary instance properties this.labelVector and this.latentDim. These are now local variables in generate().
• Make use of built tensorflow functions:
  • tf.oneHot() to encode the label input
  • .squeeze() instead of verbose .reshape([temp.shape[1], temp.shape[2], temp.shape[3]])
• Should use property raw instead of raws to match other models. Now returns both to be non-breaking.
• Better JSDoc -- I need to find a system that can sync these JSDoc types and description into our documentation website.
• Inline generateInternal() as an IIFE inside of generate() so that I didn't need to duplicate the JSDoc.
• Exhaustive unit tests.

DCGAN

• Can accept either a model.json or a manifest.json, since the manifest does not provide any necessary info. (see issue Discussion: does DCGAN really need a manifest.json? #1386)
• generate() can accept both a callback and a latentVector with both optional and in any order. Callbacks should always be last so the way that we had it did not make sense.
• Combine generate(), generateInternal() and compute() into one method.
• Offload the first 10 lines of compute() into a utility.
• Exhaustive unit tests.

tensorInput

• New utility file handles common logic in both CVAE and DCGAN (and potentially others in the future).
• modelInputShape figures out the correct tensor shape based on the model. This is really important and should be used in a few other models as well. Having hard-coded variables like IMAGE_SIZE will not work if the user provides a URL to a custom model which uses a different input size than what is expected by the code.
  • Argument layerIndex allows this to work for models like CVAE which take an array of tensors instead of a single tensor input.
• validateLatentInput includes the existing logic which I removed from DCGAN.compute() for creating either a random input tensor or a tensor from a provided array.
  • Adds an error message if the array size is wrong. This should be slightly more helpful than the TF error that would get thrown on the next line ("Uncaught Error: Based on the provided shape, [2,2], the tensor should have 4 values but has 6").
  • The user can now provide a Tensor in addition to an array.

Failed Tests

There are 2 failures in the new tests that I've added:

1. Loading a DCGAN from the current geo manifest.json does not work due to the relative URL issue discussed in Discussion: does DCGAN really need a manifest.json? #1386
2. Generating a P5 image fails. I added P5 as a devDependency so that we can do tests like this. However, calling p5Utils.setP5Instance() does not work as intended. I believe that there is an issue in the P5Util class.

[joeyklee]

✅ LGTM -- there's a lot in here!

Loading a DCGAN from the current geo manifest.json does not work due to the relative URL issue discussed in #1386

• It seems like you figured out a way to get back the info that we need from the manifest.json -- tbh, I'm not sure what other info we expect to be in there. I wonder if those were old files that were part of models exported from models trained in tensorflow or keras or something 😬
• on the point of relative imports -- the original intention was just to try to make things as easy as possible for people to load models in. I don't think we had a plan other than it seemed ok at the time. I think I'm fine to defer to you to roll out your plan!

Thank you!