Use TensorBoard from torch.utils.tensorboard to track the progress of the discriminator to make sure that it doesn't suddenly spike.

Write a GAN that generates handwritten digits.

ssh -YC -L 16007:127.0.0.1:6007 workstation

Allows me to access the remote Tensorboard session from http://localhost:16007/

Create an animation of the generator output as we interpolate from 0 to 1, 1 to 2, ..., all the way back to 0.

The current idea with conditional generation is to pass in a one-hot vector and use that to determine what kind of number the neural network will generate. What if we wanted to pass in multiple 1's in that vector?

• First, experiment to see what the generator actually generates in this situation.
• It would be nice to have the GAN be able to generate variable-sized outputs that correspond to a sequence of numbers like 567 or 97324 or something along those lines. We'd pass in a matrix with the column representing the different numbers to generate and in what sequence.

https://distill.pub/2017/aia/

Even with just the unconditional GAN.

• Cluster analysis
• What numbers have embeddings that are closer together in space?
• Use PCA to visualize the embeddings space
• Are there multiple cluster centroids for the same number?

https://pytorch.org/docs/stable/cuda.html#torch.cuda.is_available

After #2 is done, experiment with controllable generation so we can select what kind of digit we want to see.

It's hard to evaluate to the performance of a GAN. How do you even do so? Look into the literature to see what kind of stuff they're doing and why they work.

1. Fake 0 - Fake 9
2. Fake bucket

One common problem people encounter when training GANs is that the discriminator gets too good too fast because its task is much easier than the generator's. Did we notice that in the course of our experiments?