Commit Graph

107 Commits

Author SHA1 Message Date
James Betker
47a525241f Make attention norm optional 2020-07-18 07:24:02 -06:00
James Betker
8d061a2687 Add u-net discriminator with feature output 2020-07-16 10:10:09 -06:00
James Betker
1b1431133b Add DualOutputSRG
Also removes the old multi-return mechanism that Generators support.
Also fixes AttentionNorm.
2020-07-14 09:28:24 -06:00
James Betker
0b7193392f Implement unet disc
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.
2020-07-10 16:24:42 -06:00
James Betker
5f2c722a10 SRG2 revival
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.
2020-07-09 17:34:51 -06:00
James Betker
8a4eb8241d SRG3 work
Operates on top of a pre-trained SpineNET backbone (trained on CoCo 2017 with RetinaNet)

This variant is extremely shallow.
2020-07-07 13:46:40 -06:00
James Betker
a47a5dca43 Fix pixdisc bug 2020-07-05 21:57:52 -06:00
James Betker
9048105b72 Break out SRG1 as separate network
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.
2020-07-04 13:28:50 -06:00
James Betker
703dec4472 Add SpineNet & integrate with SRG
New version of SRG uses SpineNet for a switch backbone.
2020-07-03 12:07:31 -06:00
James Betker
480d1299d7 Remove RRDB with switching
This idea never really panned out, removing it.
2020-07-01 12:08:32 -06:00
James Betker
978036e7b3 Add NestedSwitchGenerator
An evolution of SwitchedResidualGenerator, this variant nests attention
modules upon themselves to extend the representative capacity of the
model significantly.
2020-06-28 21:22:05 -06:00
James Betker
c8a670842e Missed networks.py in last commit 2020-06-25 18:36:06 -06:00
James Betker
4001db1ede Add ConfigurableSwitchComputer 2020-06-24 19:49:37 -06:00
James Betker
83c3b8b982 Add parameterized noise injection into resgen 2020-06-23 10:16:02 -06:00
James Betker
dfcbe5f2db Add capability to place additional conv into discriminator
This should allow us to support larger images sizes. May need
to add another one of these.
2020-06-23 09:40:33 -06:00
James Betker
68bcab03ae Add growth channel to switch_growths for flat networks 2020-06-22 10:40:16 -06:00
James Betker
61364ec7d0 Fix inverse temperature curve logic and add upsample factor 2020-06-19 09:18:30 -06:00
James Betker
778e7b6931 Add a double-step to attention temperature 2020-06-18 11:29:31 -06:00
James Betker
645d0ca767 ResidualGen mods
- 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
2020-06-17 17:18:28 -06:00
James Betker
7d541642aa Get rid of SwitchedResidualGenerator
Just use the configurable one instead..
2020-06-16 16:23:29 -06:00
James Betker
2def96203e Mods to SwitchedResidualGenerator_arch
- Increased processing for high-resolution switches
- Do stride=2 first in HalvingProcessingBlock
2020-06-16 14:19:12 -06:00
James Betker
70c764b9d4 Create a configurable SwichedResidualGenerator
Also move attention image generator out of repo
2020-06-16 13:24:07 -06:00
James Betker
df1046c318 New arch: SwitchedResidualGenerator_arch
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.
2020-06-16 11:23:50 -06:00
James Betker
0a714e8451 Fix initialization in mhead switched rrdb 2020-06-15 21:32:03 -06:00
James Betker
be7982b9ae Add skip heads to switcher
These pass through the input so that it can be selected by the attention mechanism.
2020-06-14 12:46:54 -06:00
James Betker
48532a0a8a Fix initial_stride on lowdim models 2020-06-14 11:02:16 -06:00
James Betker
532704af40 Multiple modifications for experimental RRDB architectures
- 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
2020-06-13 11:37:27 -06:00
James Betker
5ca53e7786 Add alternative first block for PixShuffleRRDB 2020-06-10 21:45:24 -06:00
James Betker
12e8fad079 Add serveral new RRDB architectures 2020-06-09 13:28:55 -06:00
James Betker
786a4288d6 Allow switched RRDBNet to record metrics and decay temperature 2020-06-08 11:10:38 -06:00
James Betker
ae3301c0ea SwitchedRRDB work
Renames AttentiveRRDB to SwitchedRRDB. Moves SwitchedConv to
an external repo (neonbjb/switchedconv). Switchs RDB blocks instead
of conv blocks. Works good!
2020-06-08 08:47:34 -06:00
James Betker
063719c5cc Fix attention conv bugs 2020-06-06 18:31:02 -06:00
James Betker
edf0f8582e Fix rrdb bug 2020-06-02 11:15:55 -06:00
James Betker
dc17545083 Add RRDB Initial Stride
Allows downsampling immediately before processing, which reduces network complexity on
higher resolution images but keeps a higher filter count.
2020-06-02 10:47:15 -06:00
James Betker
57682ebee3 Separate feature extractors out, add resnet feature extractor 2020-05-28 20:26:30 -06:00
James Betker
3c2e5a0250 Apply fixes to resgen 2020-05-24 07:43:23 -06:00
James Betker
9b44f6f5c0 Add AssistedRRDB and remove RRDBNetXL 2020-05-23 21:09:21 -06:00
James Betker
af1968f9e5 Allow passthrough discriminator to have passthrough disabled from config 2020-05-19 09:41:16 -06:00
James Betker
9cde58be80 Make RRDB usable in the current iteration 2020-05-16 18:36:30 -06:00
James Betker
a33ec3e22b Fix skips & images samples
- 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.
2020-05-15 13:50:49 -06:00
James Betker
e36f22e14a Allow "corruptor" network to be specified
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.
2020-05-13 15:26:55 -06:00
James Betker
5d1b4caabf Allow noise to be injected at the generator inputs for resgen 2020-05-12 16:26:29 -06:00
James Betker
f217216c81 Implement ResGenv2
Implements a ResGenv2 architecture which slightly increases the complexity
of the final output layer but causes it to be shared across all skip outputs.
2020-05-12 10:09:15 -06:00
James Betker
ef48e819aa Allow resgen to have a conditional number of upsamples applied to it 2020-05-10 10:48:37 -06:00
James Betker
aa0305def9 Resnet discriminator overhaul
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.
2020-05-06 17:27:30 -06:00
James Betker
3cd85f8073 Implement ResGen arch
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.
2020-05-05 11:59:46 -06:00
James Betker
3b4e54c4c5 Add support for passthrough disc/gen
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.
2020-05-04 14:01:43 -06:00
James Betker
832f3587c5 Turn off EVDR (so we dont need the weird convs) 2020-05-02 17:47:14 -06:00
James Betker
9e1acfe396 Fixup upconv for the next attempt! 2020-05-01 19:56:14 -06:00
James Betker
7eaabce48d Full resnet corrupt, no BN
And it works! Thanks fixup..
2020-04-30 19:17:30 -06:00
James Betker
3781ea725c Add Resnet Discriminator with BN 2020-04-29 20:51:57 -06:00
James Betker
5b8a77f02c Discriminator part 1
New discriminator. Includes spectral norming.
2020-04-28 23:00:29 -06:00
James Betker
8ab595e427 Add FlatProcessorNet
After doing some thinking and reading on the subject, it occurred to me that
I was treating the generator like a discriminator by focusing the network
complexity at the feature levels. It makes far more sense to process each conv
level equally for the generator, hence the FlatProcessorNet in this commit. This
network borrows some of the residual pass-through logic from RRDB which makes
the gradient path exceptionally short for pretty much all model parameters and
can be trained in O1 optimization mode without overflows again.
2020-04-28 11:49:21 -06:00
James Betker
d95808f4ef Implement downsample GAN
This bad boy is for a workflow where you train a model on disjoint image sets to
downsample a "good" set of images like a "bad" set of images looks. You then
use that downsampler to generate a training set of paired images for supersampling.
2020-04-24 00:00:46 -06:00
James Betker
af5dfaa90d Change GT_size to target_size 2020-04-22 00:37:41 -06:00
James Betker
cc834bd5a3 Support >128px image squares 2020-04-21 16:32:59 -06:00
XintaoWang
037933ba66 mmsr 2019-08-23 21:42:47 +08:00