Discussion: does DCGAN really need a manifest.json?

[lindapaiste]

Related to #949

In DCGAN we require the user to provide a manifest.json file. It seems to me that the model.json is sufficient. The only values which we use are the model path and the modelLatentDim.

We need to know the modelLatentDim in order to get the correct shape for the random input (when no latentVector is provided). However that information can also be accessed by looking at the loaded model.

const modelLatentDim = this.model.input.shape[1];

---

I've also realized that I don't quite understand how the relative paths to the model in the manifest.json are supposed to work. I assumed that the model directory was intended to be self-contained and that a relative model path would be relative to the manifest.json file. That is, if both are in the same directory then it would be "model": "model.json".

Our CVAE examples work this way, but our DCGAN examples do not. They have a model path which is relative to the URL where the code is executed, ie. "model": "model/geo/model.json". That makes no sense to me.

My PR #1376 breaks the DGCAN examples due to the handling of the relative URLs. I can fix it so that the examples work, but is that really how it is supposed to be?