Basically just cleaning up the code, removing some bad conventions,
and reducing complexity somewhat so that I can play around with
this arch a bit more easily.
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!
The latest discriminator architecture was already pretty much a unet. This
one makes that official and uses shared layers. It also upsamples one additional
time and throws out the lowest upsampling result.
The intent is to delete the old vgg pixdisc, but I'll keep it around for a bit since
I'm still trying out a few models with it.
- 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.
Also makes all processing blocks have a conformant signature.
Alters ExpansionBlock to perform a processing conv on the passthrough
before the conjoin operation - this will break backwards compatibilty with SRG2.
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.
- Removed a bunch of unnecessary image loggers. These were just consuming space and never being viewed
- Got rid of support of artificial var_ref support. The new pixdisc is what i wanted to implement then - it's much better.
- Add pixgan GAN mechanism. This is purpose-built for the pixdisc. It is intended to promote a healthy discriminator
- Megabatchfactor was applied twice on metrics, fixed that
Adds pix_gan (untested) which swaps a portion of the fake and real image with each other, then expects the discriminator
to properly discriminate the swapped regions.
Let the Generator get to a point where it is at least competing with the discriminator before firing off.
Backwards from most GAN architectures, but this one is a bit different from most.
Relu produced good performance gains over LeakyRelu, but
GAN performance degraded significantly. Try SiLU as an alternative
to see if it's the leaky-ness we are looking for or the smooth activation
curvature.
Something strange is going on. These networks do not respond to
discriminator gradients properly anymore. SRG1 did, however so
reverting back to last known good state to figure out why.
- Get rid of forwards(), it makes numeric_stability.py not work properly.
- Do stability auditing across layers.
- Upsample last instead of first, work in much higher dimensionality for transforms.
Move to a fully fixup residual network for the switch (no
batch norms). Fix a bunch of other small bugs. Add in a
temporary latent feed-forward from the bottom of the
switch. Fix several initialization issues.
- Just use resnet blocks for the multiplexer trunk of the generator
- Every block initializes itself, rather than everything at the end
- Cleans up some messy parts of the architecture, including unnecessary
kernel sizes and places where BN is not used properly.
An evolution of SwitchedResidualGenerator, this variant nests attention
modules upon themselves to extend the representative capacity of the
model significantly.
Got rid of the converged multiplexer bases but kept the configurable architecture. The
new multiplexers look a lot like the old one.
Took some queues from the transformer architecture: translate image to a higher filter-space
and stay there for the duration of the models computation. Also perform convs after each
switch to allow the model to anneal issues that arise.
Found out that batch norm is causing the switches to init really poorly -
not using a significant number of transforms. Might be a great time to
re-consider using the attention norm, but for now just re-enable it.
- Add filters_mid spec which allows a expansion->squeeze for the transformation layers.
- Add scale and bias AFTER the switch
- Remove identity transform (models were converging on this)
- Move attention image generation and temperature setting into new function which gets called every step with a save path
The concept here is to use switching to split the generator into two functions:
interpretation and transformation. Transformation is done at the pixel level by
relatively simple conv layers, while interpretation is computed at various levels
by far more complicated conv stacks. The two are merged using the switching
mechanism.
This architecture is far less computationally intensive that RRDB.
- Add LowDimRRDB; essentially a "normal RRDB" but the RDB blocks process at a low dimension using PixelShuffle
- Add switching wrappers around it
- Add support for switching on top of multi-headed inputs and outputs
- Moves PixelUnshuffle to arch_util
Renames AttentiveRRDB to SwitchedRRDB. Moves SwitchedConv to
an external repo (neonbjb/switchedconv). Switchs RDB blocks instead
of conv blocks. Works good!
- Makes skip connections between the generator and discriminator more
extensible by adding additional configuration options for them and supporting
1 and 0 skips.
- Places the temp/ directory with sample images from the training process appear
in the training directory instead of the codes/ directory.
Model swapout is a feature where, at specified intervals,
a random D and G model will be swapped in place for the
one being trained. After a short period of time, this model
is swapped back out. This is intended to increase training
diversity.
This network is just a fixed (pre-trained) generator
that performs a corruption transformation that the
generator-in-training is expected to undo alongside
SR.
It now injects noise directly into the input filters, rather than a
pure noise filter. The pure noise filter was producing really
poor results (and I'm honestly not quite sure why).
Implements a ResGenv2 architecture which slightly increases the complexity
of the final output layer but causes it to be shared across all skip outputs.
It's been a tough day figuring out WTH is going on with my discriminators.
It appears the raw FixUp discriminator can get into an "defective" state where
they stop trying to learn and just predict as close to "0" D_fake and D_real as
possible. In this state they provide no feedback to the generator and never
recover. Adding batch norm back in seems to fix this so it must be some sort
of parameterization error.. Should look into fixing this in the future.
This is a simpler resnet-based generator which performs mutations
on an input interspersed with interpolate-upsampling. It is a two
part generator:
1) A component that "fixes" LQ images with a long string of resnet
blocks. This component is intended to remove compression artifacts
and other noise from a LQ image.
2) A component that can double the image size. The idea is that this
component be trained so that it can work at most reasonable
resolutions, such that it can be repeatedly applied to itself to
perform multiple upsamples.
The motivation here is to simplify what is being done inside of RRDB.
I don't believe the complexity inside of that network is justified.
Add RRDBNetXL, which performs processing at multiple image sizes.
Add DiscResnet_passthrough, which allows passthrough of image at different sizes for discrimination.
Adjust the rest of the repo to allow generators that return more than just a single image.
This is a checkpoint of a set of long tests with reduced-complexity networks. Some takeaways:
1) A full GAN using the resnet discriminator does appear to converge, but the quality is capped.
2) Likewise, a combination GAN/feature loss does not converge. The feature loss is optimized but
the model appears unable to fight the discriminator, so the G-loss steadily increases.
Going forwards, I want to try some bigger models. In particular, I want to change the generator
to increase complexity and capacity. I also want to add skip connections between the
disc and generator.
