Adds support for GD models, courtesy of some maths from openai.
Also:
- Fixes requirement for eval{} even when it isn't being used
- Adds support for denormalizing an imagenet norm
- Allow image_folder_dataset to normalize inbound images
- ExtensibleTrainer can denormalize images on the output path
- Support .webp - an output from LSUN
- Support logistic GAN divergence loss
- Support stylegan2 TF weight extraction for discriminator
- New injector that produces latent noise (with separated paths)
- Modify FID evaluator to be operable with rosinality-style GANs
- Turns out my custom convolution was RIDDLED with backwards bugs, which is
why the existing implementation wasn't working so well.
- Implements the switch logic from both Mixture of Experts and Switch Transformers
for testing purposes.
Changes VQVAE as so:
- Reverts back to smaller codebook
- Adds an additional conv layer at the highest resolution for both the encoder & decoder
- Uses LeakyReLU on trunk
I'm being really lazy here - these nets are not really different from each other
except at which layer they terminate. This one terminates at 2x downsampling,
which is simply indicative of a direction I want to go for testing these pixpro networks.
- The pixpro latent now rescales the latent space instead of using a "coordinate vector", which
**might** have performance implications.
- The latent against which the pixel loss is computed can now be a small, randomly sampled patch
out of the entire latent, allowing further memory/computational discounts. Since the loss
computation does not have a receptive field, this should not alter the loss.
- The instance projection size can now be separate from the pixel projection size.
- PixContrast removed entirely.
- ResUnet with full resolution added.
