- Checkpointed pretty much the entire model - enabling recurrent inputs
- Added two new models for test - adding depth (again) and removing SPSR (in lieu of the new losses)
Some lessons learned:
- Biases are fairly important as a relief valve. They dont need to be everywhere, but
most computationally heavy branches should have a bias.
- GroupNorm in SPSR is not a great idea. Since image gradients are represented
in this model, normal means and standard deviations are not applicable. (imggrad
has a high representation of 0).
- Don't fuck with the mainline of any generative model. As much as possible, all
additions should be done through residual connections. Never pollute the mainline
with reference data, do that in branches. It basically leaves the mode untrainable.
SPSR_model really isn't that different from SRGAN_model. Rather than continuing to re-implement
everything I've done in SRGAN_model, port the new stuff from SPSR over.
This really demonstrates the need to refactor SRGAN_model a bit to make it cleaner. It is quite the
beast these days..
This is done by pre-training a feature net that predicts the features
of HR images from LR images. Then use the original feature network
and this new one in tandem to work only on LR/Gen images.
The logic is that the discriminator may be incapable of providing a truly
targeted loss for all image regions since it has to be too generic
(basically the same argument for the switched generator). So add some
switches in! See how it works!
- Swap multiple blocks in the image instead of just one. The discriminator was clearly
learning that most blocks have one region that needs to be fixed.
- Relax block size constraints. This was in place to gaurantee that the discriminator
signal was clean. Instead, just downsample the "loss image" with bilinear interpolation.
The result is noisier, but this is actually probably healthy for the discriminator.
Big update to SRG2 architecture to pull in a lot of things that have been learned:
- Use group norm instead of batch norm
- Initialize the weights on the transformations low like is done in RRDB rather than using the scalar. Models live or die by their early stages, and this ones early stage is pretty weak
- Transform multiplexer to use u-net like architecture.
- Just use one set of configuration variables instead of a list - flat networks performed fine in this regard.
