mrq/DL-Art-School
1
3
Fork 5
Code Issues Pull Requests Packages Projects Releases Wiki Activity
388 Commits 1 Branch 0 Tags 5.6 MiB
d90c96e55e
Commit Graph

287 Commits

Author SHA1 Message Date
James Betker
30b16d5235 Update how branch GAN grad is disseminated 2020-08-06 11:13:02 -06:00
James Betker
1f21c02f8b Add cross-compare discriminator 2020-08-06 08:56:21 -06:00
James Betker
be272248af More RAGAN fixes 2020-08-05 16:47:21 -06:00
James Betker
26a6a5d512 Compute grad GAN loss against both the branch and final target, simplify pixel loss 
Also fixes a memory leak issue where we weren't detaching our loss stats when
logging them. This stabilizes memory usage substantially.
 2020-08-05 12:08:15 -06:00
James Betker
299ee13988 More RAGAN fixes 2020-08-05 11:03:06 -06:00
James Betker
b8a4df0a0a Enable RAGAN in SPSR, retrofit old RAGAN for efficiency 2020-08-05 10:34:34 -06:00
James Betker
3ab39f0d22 Several new spsr nets 2020-08-05 10:01:24 -06:00
James Betker
3c0a2d6efe Fix grad branch debug out 2020-08-04 16:43:43 -06:00
James Betker
ec2a795d53 Fix multistep optimizer (feeding from wrong config params) 2020-08-04 16:42:58 -06:00
James Betker
4bfbdaf94f Don't recompute generator outputs for D in standard operation 
Should significantly improve training performance with negligible
results differences.
 2020-08-04 11:28:52 -06:00
James Betker
11b227edfc Whoops 2020-08-04 10:30:40 -06:00
James Betker
6d25bcd5df Apply fixes to grad discriminator 2020-08-04 10:25:13 -06:00
James Betker
c7e5d3888a Add pix_grad_branch loss to metrics 2020-08-03 16:21:05 -06:00
James Betker
0d070b47a7 Add simplified SPSR architecture 
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.
 2020-08-03 10:25:37 -06:00
James Betker
47e24039b5 Fix bug that makes feature loss run even when it is off 2020-08-02 20:37:51 -06:00
James Betker
328afde9c0 Integrate SPSR into SRGAN_model 
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..
 2020-08-02 12:55:08 -06:00
James Betker
c8da78966b Substantial SPSR mods & fixes 
- Added in gradient accumulation via mega-batch-factor
- Added AMP
- Added missing train hooks
- Added debug image outputs
- Cleaned up including removing GradientPenaltyLoss, custom SpectralNorm
- Removed all the custom discriminators
 2020-08-02 10:45:24 -06:00
James Betker
f894ba8f98 Add SPSR_module 
This is a port from the SPSR repo, it's going to need a lot of work to be properly integrated
but as of this commit it at least runs.
 2020-08-01 22:02:54 -06:00
James Betker
f33ed578a2 Update how attention_maps are created 2020-08-01 20:23:46 -06:00
James Betker
c139f5cd17 More torch 1.6 fixes 2020-07-31 17:03:20 -06:00
James Betker
a66fbb32b6 Fix fixed_disc DataParallel issue 2020-07-31 16:59:23 -06:00
James Betker
8dd44182e6 Fix scale torch warning 2020-07-31 16:56:04 -06:00
James Betker
bcebed19b7 Fix pixdisc bugs 2020-07-31 16:38:14 -06:00
James Betker
eb11a08d1c Enable disjoint feature networks 
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.
 2020-07-31 16:29:47 -06:00
James Betker
6e086d0c20 Fix fixed_disc 2020-07-31 15:07:10 -06:00
James Betker
d5fa059594 Add capability to have old discriminators serve as feature networks 2020-07-31 14:59:54 -06:00
James Betker
6b45b35447 Allow multi_step_lr_scheduler to load a new LR schedule when restoring state 2020-07-31 11:21:11 -06:00
James Betker
e37726f302 Add feature_model for training custom feature nets 2020-07-31 11:20:39 -06:00
James Betker
7629cb0e61 Add FDPL Loss 
New loss type that can replace PSNR loss. Works against the frequency domain
and focuses on frequency features loss during hr->lr conversion.
 2020-07-30 20:47:57 -06:00
James Betker
85ee64b8d9 Turn down feadisc intensity 
Honestly - this feature is probably going to be removed soon, so backwards
compatibility is not a huge deal anymore.
 2020-07-27 15:28:55 -06:00
James Betker
ebb199e884 Get rid of safety valve (probably being encountered in val) 2020-07-26 22:51:59 -06:00
James Betker
d09ed4e5f7 Misc fixes 2020-07-26 22:44:24 -06:00
James Betker
c54784ae9e Fix feature disc log item error 2020-07-26 22:25:59 -06:00
James Betker
9a8f227501 Allow separate dataset to pushed in for GAN-only training 2020-07-26 21:44:45 -06:00
James Betker
b06e1784e1 Fix SRG4 & switch disc 
"fix". hehe.
 2020-07-25 17:16:54 -06:00
James Betker
e6e91a1d75 Add SRG4 
Back to the idea that maybe what we need is a hybrid
approach between pure switches and RDB.
 2020-07-24 20:32:49 -06:00
James Betker
3320ad685f Fix mega_batch_factor not set for test 2020-07-24 12:26:44 -06:00
James Betker
c50cce2a62 Add an abstract, configurabler weight scheduling class and apply it to the feature weight 2020-07-23 17:03:54 -06:00
James Betker
9ccf771629 Fix feature validation, wrong device 
Only shows up in distributed training for some reason.
 2020-07-23 10:16:34 -06:00
James Betker
bba283776c Enable find_unused_parameters for DistributedDataParallel 
attention_norm has some parameters which are not used to compute grad,
which is causing failures in the distributed case.
 2020-07-23 09:08:13 -06:00
James Betker
dbf6147504 Add switched discriminator 
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!
 2020-07-22 20:52:59 -06:00
James Betker
106b8da315 Assert that temperature is set properly in eval mode. 2020-07-22 20:50:59 -06:00
James Betker
c74b9ee2e4 Add a way to disable grad on portions of the generator graph to save memory 2020-07-22 11:40:42 -06:00
James Betker
e3adafbeac Add convert_model.py and a hacky way to add extra layers to a model 2020-07-22 11:39:45 -06:00
James Betker
7f7e17e291 Update feature discriminator further 
Move the feature/disc losses closer and add a feature computation layer.
 2020-07-20 20:54:45 -06:00
James Betker
46aa776fbb Allow feature discriminator unet to only output closest layer to feature output 2020-07-19 19:05:08 -06:00
James Betker
8a9f215653 Huge set of mods to support progressive generator growth 2020-07-18 14:18:48 -06:00
James Betker
47a525241f Make attention norm optional 2020-07-18 07:24:02 -06:00
James Betker
ad97a6a18a Progressive SRG first check-in 2020-07-18 07:23:26 -06:00
James Betker
b08b1cad45 Fix feature decay 2020-07-16 23:27:06 -06:00
James Betker
3e7a83896b Fix pixgan debugging issues 2020-07-16 11:45:19 -06:00
James Betker
a1bff64d1a More fixes 2020-07-16 10:48:48 -06:00
James Betker
240f254263 More loss fixes 2020-07-16 10:45:50 -06:00
James Betker
6cfa67d831 Fix featuredisc broadcast error 2020-07-16 10:18:30 -06:00
James Betker
8d061a2687 Add u-net discriminator with feature output 2020-07-16 10:10:09 -06:00
James Betker
0c4c388e15 Remove dualoutputsrg 
Good idea, didn't pan out.
 2020-07-16 10:09:24 -06:00
James Betker
4bcc409fc7 Fix loadSRG2 typo 2020-07-14 10:20:53 -06:00
James Betker
1e4083a35b Apply temperature mods to all SRG models 
(Honestly this needs to be base classed at this point)
 2020-07-14 10:19:35 -06:00
James Betker
7659bd6818 Fix temperature equation 2020-07-14 10:17:14 -06:00
James Betker
853468ef82 Allow legacy state_dicts in srg2 2020-07-14 10:03:45 -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
a2285ff2ee Scale anorm by transform count 2020-07-13 08:49:09 -06:00
James Betker
dd0bbd9a7c Enable AttentionNorm on SRG2 2020-07-13 08:38:17 -06:00
James Betker
4c0f770f2a Fix inverted temperature curve bug 2020-07-12 11:02:50 -06:00
James Betker
14d23b9d20 Fixes, do fake swaps less often in pixgan discriminator 2020-07-11 21:22:11 -06:00
James Betker
ba6187859a err5 2020-07-10 23:02:56 -06:00
James Betker
902527dfaa err4 2020-07-10 23:00:21 -06:00
James Betker
020b3361fa err3 2020-07-10 22:57:34 -06:00
James Betker
b3a2c21250 err2 2020-07-10 22:52:02 -06:00
James Betker
716433db1f err1 2020-07-10 22:50:56 -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
812c684f7d Update pixgan swap algorithm 
- 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.
 2020-07-10 15:56:14 -06:00
James Betker
33ca3832e1 Move ExpansionBlock to arch_util 
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.
 2020-07-10 15:53:41 -06:00
James Betker
5e8b52f34c Misc changes 2020-07-10 09:45:48 -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
12da993da8 More fixes... 2020-07-08 22:07:09 -06:00
James Betker
7d6eb28b87 More fixes 2020-07-08 22:00:57 -06:00
James Betker
b2507be13c Fix up pixgan loss and pixdisc 2020-07-08 21:27:48 -06:00
James Betker
26a4a66d1c Bug fixes and new gan mechanism 
- 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.
 2020-07-08 17:40:26 -06:00
James Betker
4305be97b4 Update log metrics 
They should now be universal regardless of job configuration
 2020-07-07 15:33:22 -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
0acad81035 More SRG2 adjustments.. 2020-07-06 22:40:40 -06:00
James Betker
086b2f0570 More bugs 2020-07-06 22:28:07 -06:00
James Betker
d4d4f85fc0 Bug fixes 2020-07-06 22:25:40 -06:00
James Betker
3c31bea1ac SRG2 architectural changes 2020-07-06 22:22:29 -06:00
James Betker
9a1c3241f5 Switch discriminator to groupnorm 2020-07-06 20:59:59 -06:00
James Betker
6beefa6d0c PixDisc - Add two more levels of losses coming from this gen at higher resolutions 2020-07-06 11:15:52 -06:00
James Betker
2636d3b620 Fix assertion error 2020-07-06 09:23:53 -06:00
James Betker
8f92c0a088 Interpolate attention well before softmax 2020-07-06 09:18:30 -06:00
James Betker
72f90cabf8 More pixdisc fixes 2020-07-05 22:03:16 -06:00
James Betker
909007ee6a Add G_warmup 
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.
 2020-07-05 21:58:35 -06:00
James Betker
a47a5dca43 Fix pixdisc bug 2020-07-05 21:57:52 -06:00
James Betker
d0957bd7d4 Alter weight initialization for transformation blocks 2020-07-05 17:32:46 -06:00
James Betker
16d1bf6dd7 Replace ConvBnRelus in SRG2 with Silus 2020-07-05 17:29:20 -06:00
James Betker
10f7e49214 Add ConvBnSilu to replace ConvBnRelu 
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.
 2020-07-05 13:39:08 -06:00
James Betker
9934e5d082 Move SRG1 to identical to new 2020-07-05 08:49:34 -06:00
James Betker
416538f31c SRG1 conjoined except ConvBnRelu 2020-07-05 08:44:17 -06:00
James Betker
c58c2b09ca Back to remove all biases (looks like a ConvBnRelu made its way in..) 2020-07-04 22:41:02 -06:00
James Betker
86cda86e94 Re-add biases, also add new init 
A/B testing where we lost our GAN competitiveness.
 2020-07-04 22:24:42 -06:00
James Betker
b03741f30e Remove all biases from generator 
Continuing to investigate loss of GAN competitiveness, this is a big difference
between "old" SRG1 and "new".
 2020-07-04 22:19:55 -06:00