Adding GAN support to TensorGraph

I think our current model with TensorGraph can’t nicely support GANs. In order to train GANs, you need to train both a discriminator D and a generator G. The training of these two models is threaded together, with each updated for a few minibatches at a time. While updating the generator, the discriminator weights should be frozen. While updating the discriminator, the generator weights need to be frozen.

I have an idea for a new TensorGraph API to support this sort of training. We add a new layers kwarg to fit:

• model.fit(feed_generator, layers=None)

The idea is that fit can be instructed to train only some layers at a time. Note that there’s now a semantic change:

• Layer is responsible for tracking the variables it introduces.
  • layer.variables should be a list of all trainable variables created by that Layer.

@peastman and @lilleswing Does this basic design make sense?

Hat tip to @enf for explaining GAN training to me in detail!

[peastman]

How about this as a candidate API:

submodel = model.create_submodel(layers, loss, optimizer)
model.fit(dataset, submodel=submodel)

So a submodel specifies a list of layers, a loss function, and an optimizer. When you perform fitting, you can optionally specify a submodel to do the fitting on. It then uses a different train_op, but everything else works exactly the same.

[pierrebaque]

In the abstract of SeqGAN, they also say the following, which could be very important in my view:

" A major reason lies in that the discrete outputs from the generative model make it difficult to pass the gradient update from the discriminative model to the generative model."

I think that the model that should be used really depends on the kind of data we want to generate. Maybe would it be better to focus on a specific chemical application first?