eschmidbauer/DL-Art-School
1
0
Fork 0
forked from mrq/DL-Art-School
Code Issues Pull Requests Packages Projects Releases Wiki Activity
889 Commits 1 Branch 0 Tags 5.6 MiB
9c53314ea2
Commit Graph

644 Commits

Author SHA1 Message Date
James Betker
c3ef8a4a31 Stacked switches - return a tuple 2020-10-04 21:02:24 -06:00
James Betker
13f97e1e97 Add recursive loss 2020-10-04 20:48:15 -06:00
James Betker
ffd069fd97 Lots of SSG work 
- 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)
 2020-10-04 20:48:08 -06:00
James Betker
aca2c7ab41 Full checkpoint-ize SSG1 2020-10-04 18:24:52 -06:00
James Betker
e3294939b0 Revert "SSG: offer option to use BN-based attention normalization" 
Didn't work. Oh well.

This reverts commit 5cd2b37591.
 2020-10-03 17:54:53 -06:00
James Betker
5cd2b37591 SSG: offer option to use BN-based attention normalization 
Not sure how this is going to work, lets try it.
 2020-10-03 16:16:19 -06:00
James Betker
9b4ed82093 Get rid of unused convs in spsr7 2020-10-03 11:36:26 -06:00
James Betker
3561cc164d Fix up fea_loss calculator (for validation) 
Not sure how this was working in regular training mode, but it
was failing in DDP.
 2020-10-03 11:19:20 -06:00
James Betker
6c9718ad64 Don't log if you aren't 0 rank 2020-10-03 11:14:13 -06:00
James Betker
922b1d76df Don't record visuals when not on rank 0 2020-10-03 11:10:03 -06:00
James Betker
8197fd646f Don't accumulate losses for metrics when the loss isn't a tensor 2020-10-03 11:03:55 -06:00
James Betker
19a4075e1e Allow checkpointing to be disabled in the options file 
Also makes options a global variable for usage in utils.
 2020-10-03 11:03:28 -06:00
James Betker
dd9d7b27ac Add more sophisticated mechanism for balancing GAN losses 2020-10-02 22:53:42 -06:00
James Betker
39865ca3df TOTAL_loss, dumbo 2020-10-02 21:06:10 -06:00
James Betker
4e44fcd655 Loss accumulator fix 2020-10-02 20:55:33 -06:00
James Betker
567b4d50a4 ExtensibleTrainer - don't compute backward when there is no loss 2020-10-02 20:54:06 -06:00
James Betker
146a9125f2 Modify geometric & translational losses so they can be used with embeddings 2020-10-02 20:40:13 -06:00
James Betker
e30a1443cd Change sw2 refs 2020-10-02 09:01:18 -06:00
James Betker
e38716925f Fix spsr8 class init 2020-10-02 09:00:18 -06:00
James Betker
35469f08e2 Spsr 8 2020-10-02 08:58:15 -06:00
James Betker
aa4fd89018 resnext with groupnorm 2020-10-01 15:49:28 -06:00
James Betker
8beaa47933 resnext discriminator 2020-10-01 11:48:14 -06:00
James Betker
55f2764fef Allow fixup50 to be used as a discriminator 2020-10-01 11:28:18 -06:00
James Betker
7986185fcb Change 'mod_step' to 'every' 2020-10-01 11:28:06 -06:00
James Betker
d9ae970fd9 SSG update 2020-10-01 11:27:51 -06:00
James Betker
e3053e4e55 Exchange SpsrNet for SpsrNetSimplified 2020-09-30 17:01:04 -06:00
James Betker
66d4512029 Fix up translational equivariance loss so it's ready for prime time 2020-09-30 12:01:00 -06:00
James Betker
896b4f5be2 Revert "spsr7 adjustments" 
This reverts commit 9fee1cec71.
 2020-09-29 18:30:41 -06:00
James Betker
9fee1cec71 spsr7 adjustments 2020-09-29 17:19:59 -06:00
James Betker
dc8f3b24de Don't let duplicate keys be used for injectors and losses 2020-09-29 16:59:44 -06:00
James Betker
0b5a033503 spsr7 + cleanup 
SPSR7 adds ref onto spsr6, makes more "common sense" mods.
 2020-09-29 16:59:26 -06:00
James Betker
f9b83176f1 Fix bugs in extensibletrainer 2020-09-28 22:09:42 -06:00
James Betker
db52bec4ab spsr6 
This is meant to be a variant of SPSR5 that harkens
back to the simpler earlier architectures that do not
have embeddings or ref_ inputs, but do have deep
multiplexers. It does, however, use some of the new
conjoin mechanisms.
 2020-09-28 22:09:27 -06:00
James Betker
7e240f2fed Recurrent / teco work 2020-09-28 22:06:56 -06:00
James Betker
aeaf185314 Add RCAN 2020-09-27 16:00:41 -06:00
James Betker
4d29b7729e Model arch cleanup 2020-09-27 11:18:45 -06:00
James Betker
31641d7f63 Add ImagePatchInjector and TranslationalLoss 2020-09-26 21:25:32 -06:00
James Betker
d8621e611a BackboneSpineNoHead takes ref 2020-09-26 21:25:04 -06:00
James Betker
5a27187c59 More mods to accomodate new dataset 2020-09-25 22:45:57 -06:00
James Betker
6d0490a0e6 Tecogan implementation work 2020-09-25 16:38:23 -06:00
James Betker
ce4613ecb9 Finish up single_image_dataset work 
Sweet!
 2020-09-25 16:37:54 -06:00
James Betker
ea565b7eaf More fixes 2020-09-24 17:51:52 -06:00
James Betker
553917a8d1 Fix torchvision import bug 2020-09-24 17:38:34 -06:00
James Betker
58886109d4 Update how spsr arches do attention to conform with sgsr 2020-09-24 16:53:54 -06:00
James Betker
9a50a7966d SiLU doesnt support inplace 2020-09-23 21:09:13 -06:00
James Betker
eda0eadba2 Use custom SiLU 
Torch didnt have this before 1.7
 2020-09-23 21:05:06 -06:00
James Betker
05963157c1 Several things 
- Fixes to 'after' and 'before' defs for steps (turns out they werent working)
- Feature nets take in a list of layers to extract. Not fully implemented yet.
- Fixes bugs with RAGAN
- Allows real input into generator gan to not be detached by param
 2020-09-23 11:56:36 -06:00
James Betker
4ab989e015 try again.. 2020-09-22 18:27:52 -06:00
James Betker
3b6c957194 Fix? it again? 2020-09-22 18:25:59 -06:00
James Betker
7b60d9e0d8 Fix? cosine loss 2020-09-22 18:18:35 -06:00
James Betker
2e18c4c22d Add CosineEmbeddingLoss to F 2020-09-22 17:10:29 -06:00
James Betker
f40beb5460 Add 'before' and 'after' defs to injections, steps and optimizers 2020-09-22 17:03:22 -06:00
James Betker
419f77ec19 Some new backbones 2020-09-21 12:36:49 -06:00
James Betker
9429544a60 Spinenet: implementation without 4x downsampling right off the bat 2020-09-21 12:36:30 -06:00
James Betker
53a5657850 Fix SSGR 2020-09-20 19:07:15 -06:00
James Betker
17c569ea62 Add geometric loss 2020-09-20 16:24:23 -06:00
James Betker
17dd99b29b Fix bug with discriminator noise addition 
It wasn't using the scale and was applying the noise to the
underlying state variable.
 2020-09-20 12:00:27 -06:00
James Betker
3138f98fbc Allow discriminator noise to be injected at the loss level, cleans up configs 2020-09-19 21:47:52 -06:00
James Betker
e9a39bfa14 Recursively detach all outputs, even if they are nested in data structures 2020-09-19 21:47:34 -06:00
James Betker
fe82785ba5 Add some new architectures to ssg 2020-09-19 21:47:10 -06:00
James Betker
b83f097082 Get rid of get_debug_values from RRDB, rectify outputs 2020-09-19 21:46:36 -06:00
James Betker
e0bd68efda Add ImageFlowInjector 2020-09-19 10:07:00 -06:00
James Betker
e2a146abc7 Add in experiments hook 2020-09-19 10:05:25 -06:00
James Betker
9a17ade550 Some convenience adjustments to ExtensibleTrainer 2020-09-17 21:05:32 -06:00
James Betker
9963b37200 Add a new script for loading a discriminator network and using it to filter images 2020-09-17 13:30:32 -06:00
James Betker
723754c133 Update attention debugger outputting for SSG 2020-09-16 13:09:46 -06:00
James Betker
0918430572 SSG network 
This branches off of SPSR. It is identical but substantially reduced
in complexity. It's intended to be my long term working arch.
 2020-09-15 20:59:24 -06:00
James Betker
6deab85b9b Add BackboneEncoderNoRef 2020-09-15 16:55:38 -06:00
James Betker
d0321ca5de Don't load amp state dict if amp is disabled 2020-09-14 15:21:42 -06:00
James Betker
ccf8438001 SPSR5 
This is SPSR4, but the multiplexers have access to the output of the transformations
for making their decision.
 2020-09-13 20:10:24 -06:00
James Betker
5b85f891af Only log the name of the first network in the total_loss training set 2020-09-12 16:07:09 -06:00
James Betker
fb595e72a4 Supporting infrastructure in ExtensibleTrainer to train spsr4 
Need to be able to train 2 nets in one step: the backbone will be entirely separate
with its own optimizer (for an extremely low LR).

This functionality was already present, just not implemented correctly.
 2020-09-11 22:57:06 -06:00
James Betker
4e44bca611 SPSR4 
aka - return of the backbone! I'm tired of massively overparameterized generators
with pile-of-shit multiplexers. Let's give this another try..
 2020-09-11 22:55:37 -06:00
James Betker
19896abaea Clean up old SwitchedSpsr arch 
It didn't work anyways, so why not?
 2020-09-11 16:09:28 -06:00
James Betker
50ca17bb0a Feature mode -> back to LR fea 2020-09-11 13:09:55 -06:00
James Betker
1086f0476b Fix ref branch using fixed filters 2020-09-11 08:58:35 -06:00
James Betker
8c469b8286 Enable memory checkpointing 2020-09-11 08:44:29 -06:00
James Betker
5189b11dac Add combined dataset for training across multiple datasets 2020-09-11 08:44:06 -06:00
James Betker
313424d7b5 Add new referencing discriminator 
Also extend the way losses work so that you can pass
parameters into the discriminator from the config file
 2020-09-10 21:35:29 -06:00
James Betker
9e5aa166de Report the standard deviation of ref branches 
This patch also ups the contribution
 2020-09-10 16:34:41 -06:00
James Betker
668bfbff6d Back to best arch for spsr3 2020-09-10 14:58:14 -06:00
James Betker
992b0a8d98 spsr3 with conjoin stage as part of the switch 2020-09-10 09:11:37 -06:00
James Betker
e0fc5eb50c Temporary commit - noise 2020-09-09 17:12:52 -06:00
James Betker
00da69d450 Temporary commit - ref 2020-09-09 17:09:44 -06:00
James Betker
df59d6c99d More spsr3 mods 
- Most branches get their own noise vector now.
- First attention branch has the intended sole purpose of raw image processing
- Remove norms from joiner block
 2020-09-09 16:46:38 -06:00
James Betker
747ded2bf7 Fixes to the spsr3 
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.
 2020-09-09 15:28:14 -06:00
James Betker
0ffac391c1 SPSR with ref joining 2020-09-09 11:17:07 -06:00
James Betker
3027e6e27d Enable amp to be disabled 2020-09-09 10:45:59 -06:00
James Betker
c04f244802 More mods 2020-09-08 20:36:27 -06:00
James Betker
dffbfd2ec4 Allow SRG checkpointing to be toggled 2020-09-08 15:14:43 -06:00
James Betker
e6207d4c50 SPSR3 work 
SPSR3 is meant to fix whatever is causing the switching units
inside of the newer SPSR architectures to fail and basically
not use the multiplexers.
 2020-09-08 15:14:23 -06:00
James Betker
5606e8b0ee Fix SRGAN_model/fullimgdataset compatibility 1 2020-09-08 11:34:35 -06:00
James Betker
22c98f1567 Move MultiConvBlock to arch_util 2020-09-08 08:17:27 -06:00
James Betker
146ace0859 CSNLN changes (removed because it doesnt train well) 2020-09-08 08:04:16 -06:00
James Betker
f43df7f5f7 Make ExtensibleTrainer compatible with process_video 2020-09-08 08:03:41 -06:00
James Betker
a18ece62ee Add updated spsr net for test 2020-09-07 17:01:48 -06:00
James Betker
55475d2ac1 Clean up unused archs 2020-09-07 11:38:11 -06:00
James Betker
e8613041c0 Add novograd optimizer 2020-09-06 17:27:08 -06:00
James Betker
b1238d29cb Fix trainable not applying to discriminators 2020-09-05 20:31:26 -06:00
James Betker
21ae135f23 Allow Novograd to be used as an optimizer 2020-09-05 16:50:13 -06:00
James Betker
912a4d9fea Fix srg computer bug 2020-09-05 07:59:54 -06:00
James Betker
0dfd8eaf3b Support injectors that run in eval only 2020-09-05 07:59:45 -06:00
James Betker
44c75f7642 Undo SRG change 2020-09-04 17:32:16 -06:00
James Betker
6657a406ac Mods needed to support training a corruptor again: 
- Allow original SPSRNet to have a specifiable block increment
- Cleanup
- Bug fixes in code that hasnt been touched in awhile.
 2020-09-04 15:33:39 -06:00
James Betker
bfdfaab911 Checkpoint RRDB 
Greatly reduces memory consumption with a low performance penalty
 2020-09-04 15:32:00 -06:00
James Betker
696242064c Use tensor checkpointing to drastically reduce memory usage 
This comes at the expense of computation, but since we can use much larger
batches, it results in a net speedup.
 2020-09-03 11:33:36 -06:00
James Betker
365813bde3 Add InterpolateInjector 2020-09-03 11:32:47 -06:00
James Betker
d90c96e55e Fix greyscale injector 2020-09-02 10:29:40 -06:00
James Betker
8b52d46847 Interpreted feature loss to extensibletrainer 2020-09-02 10:08:24 -06:00
James Betker
886d59d5df Misc fixes & adjustments 2020-09-01 07:58:11 -06:00
James Betker
0a9b85f239 Fix vgg_gn input_img_factor 2020-08-31 09:50:30 -06:00
James Betker
4b4d08bdec Enable testing in ExtensibleTrainer, fix it in SRGAN_model 
Also compute fea loss for this.
 2020-08-31 09:41:48 -06:00
James Betker
b2091cb698 feamod fix 2020-08-30 08:08:49 -06:00
James Betker
a56e906f9c train HR feature trainer 2020-08-29 09:27:48 -06:00
James Betker
0e859a8082 4x spsr ref (not workin) 2020-08-29 09:27:18 -06:00
James Betker
25832930db Update loss with lr crossgan 2020-08-26 17:57:22 -06:00
James Betker
cbd5e7a986 Support old school crossgan in extensibletrainer 2020-08-26 17:52:35 -06:00
James Betker
8a6a2e6e2e Rev3 of the full image ref arch 2020-08-26 17:11:01 -06:00
James Betker
f35b3ad28f Fix val behavior for ExtensibleTrainer 2020-08-26 08:44:22 -06:00
James Betker
434ed70a9a Wrap vgg disc 2020-08-25 18:14:45 -06:00
James Betker
83f2f8d239 more debugging 2020-08-25 18:12:12 -06:00
James Betker
3f60281da7 Print when wrapping 2020-08-25 18:08:46 -06:00
James Betker
bae18c05e6 wrap disc grad 2020-08-25 17:58:20 -06:00
James Betker
f85f1e21db Turns out, can't do that 2020-08-25 17:18:52 -06:00
James Betker
935a735327 More dohs 2020-08-25 17:05:16 -06:00
James Betker
53e67bdb9c Distribute get_grad_no_padding 2020-08-25 17:03:18 -06:00
James Betker
2f706b7d93 I an inept. 2020-08-25 16:42:59 -06:00
James Betker
8bae0de769 ffffffffffffffffff 2020-08-25 16:41:01 -06:00
James Betker
1fe16f71dd Fix bug reporting spsr gan weight 2020-08-25 16:37:45 -06:00
James Betker
96586d6592 Fix distributed d_grad 2020-08-25 16:06:27 -06:00
James Betker
09a9079e17 Check rank before doing image logging. 2020-08-25 16:00:49 -06:00
James Betker
a1800f45ef Fix for referencingmultiplexer 2020-08-25 15:43:12 -06:00
James Betker
a65b07607c Reference network 2020-08-25 11:56:59 -06:00
James Betker
f9276007a8 More fixes to corrupt_fea 2020-08-23 17:52:18 -06:00
James Betker
0005c56cd4 dbg 2020-08-23 17:43:03 -06:00
James Betker
4bb5b3c981 corfea debugging 2020-08-23 17:39:02 -06:00
James Betker
7713cb8df5 Corrupted features in srgan 2020-08-23 17:32:03 -06:00
James Betker
dffc15184d More ExtensibleTrainer work 
It runs now, just need to debug it to reach performance parity with SRGAN. Sweet.
 2020-08-23 17:22:45 -06:00
James Betker
afdd93fbe9 Grey feature 2020-08-22 13:41:38 -06:00
James Betker
e59e712e39 More ExtensibleTrainer work 2020-08-22 13:08:33 -06:00
James Betker
f40545f235 ExtensibleTrainer work 2020-08-22 08:24:34 -06:00
James Betker
a498d7b1b3 Report l_g_gan_grad before weight multiplication 2020-08-20 11:57:53 -06:00
James Betker
9d77a4db2e Allow initial temperature to be specified to SPSR net for inference 2020-08-20 11:57:34 -06:00
James Betker
24bdcc1181 Let SwitchedSpsr transform count be specified 2020-08-18 09:10:25 -06:00
James Betker
74cdaa2226 Some work on extensible trainer 2020-08-18 08:49:32 -06:00
James Betker
868d0aa442 Undo early dim reduction on grad branch for SPSR_arch 2020-08-14 16:23:42 -06:00
James Betker
2d205f52ac Unite spsr_arch switched gens 
Found a pretty good basis model.
 2020-08-12 17:04:45 -06:00
James Betker
3d0ece804b SPSR LR2 2020-08-12 08:45:49 -06:00
James Betker
ab04ca1778 Extensible trainer (in progress) 2020-08-12 08:45:23 -06:00
James Betker
cb316fabc7 Use LR data for image gradient prediction when HR data is disjoint 2020-08-10 15:00:28 -06:00
James Betker
f0e2816239 Denoise attention maps 2020-08-10 14:59:58 -06:00
James Betker
59aba1daa7 LR switched SPSR arch 
This variant doesn't do conv processing at HR, which should save
a ton of memory in inference. Lets see how it works.
 2020-08-10 13:03:36 -06:00
James Betker
4e972144ae More attention fixes for switched_spsr 2020-08-07 21:11:50 -06:00
James Betker
d02509ef97 spsr_switched missing import 2020-08-07 21:05:29 -06:00
James Betker
887806ffa0 Finish up spsr_switched 2020-08-07 21:03:48 -06:00
James Betker
1d5f4f6102 Crossgan 2020-08-07 21:03:39 -06:00
James Betker
fd7b6ca0a9 Comptue gan_grad_branch.... 2020-08-06 12:11:40 -06:00
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
James Betker
726e946e79 Turn BN off in SRG1 
This wont work well but just testing if GAN performance comes back
 2020-07-04 14:51:27 -06:00
James Betker
0ee39d419b OrderedDict not needed 2020-07-04 14:09:27 -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
510b2f887d Remove RDB from srg2 
Doesnt seem to work so great.
 2020-07-03 22:31:20 -06:00
James Betker
da4335c25e Add a feature-based validation test 2020-07-03 15:18:57 -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
3ed7a2b9ab Move ConvBnRelu/Lelu to arch_util 2020-07-03 12:06:38 -06:00
James Betker
e9ee67ff10 Integrate RDB into SRG 
The last RDB for each cluster is switched.
 2020-07-01 17:19:55 -06:00
James Betker
6ac6c95177 Fix scaling bug 2020-07-01 16:42:27 -06:00
James Betker
30653181ba Experiment: get rid of post_switch_conv 2020-07-01 16:30:40 -06:00
James Betker
17191de836 Experiment: bring initialize_weights back again 
Something really strange going on here..
 2020-07-01 15:58:13 -06:00
James Betker
d1d573de07 Experiment: new init and post-switch-conv 2020-07-01 15:25:54 -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
e2398ac83c Experiment: revert initialization changes 2020-07-01 12:08:09 -06:00
James Betker
78276afcaa Experiment: Back to lelu 2020-07-01 11:43:25 -06:00
James Betker
b945021c90 SRG v2 - Move to Relu, rely on Module-based initialization 2020-07-01 11:33:32 -06:00
James Betker
604763be68 NSG r7 
Converts the switching trunk to a VGG-style network to make it more comparable
to SRG architectures.
 2020-07-01 09:54:29 -06:00
James Betker
87f1e9c56f Invert ResGen2 to operate in LR space 2020-06-30 20:57:40 -06:00
James Betker
e07d8abafb NSG rev 6 
- Disable style passthrough
- Process multiplexers starting at base resolution
 2020-06-30 20:47:26 -06:00
James Betker
3ce1a1878d NSG improvements (r5) 
- 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.
 2020-06-30 16:59:57 -06:00
James Betker
75f148022d Even more NSG improvements (r4) 2020-06-30 13:52:47 -06:00
James Betker
773753073f More NSG improvements (v3) 
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.
 2020-06-29 20:26:51 -06:00
James Betker
4b82d0815d NSG improvements 
- 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.
 2020-06-29 10:09:51 -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
407224eba1 Re-work SwitchedResgen2 
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.
 2020-06-25 18:17:05 -06:00
James Betker
42a10b34ce Re-enable batch norm on switch processing blocks 
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.
 2020-06-24 21:15:17 -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
0584c3b587 Add negative_transforms switch to resgen 2020-06-23 09:41:12 -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
030648f2bc Remove batchnorms from resgen 2020-06-22 17:23:36 -06:00
James Betker
68bcab03ae Add growth channel to switch_growths for flat networks 2020-06-22 10:40:16 -06:00
James Betker
3b81712c49 Remove BN from transforms 2020-06-19 16:52:56 -06:00
James Betker
61364ec7d0 Fix inverse temperature curve logic and add upsample factor 2020-06-19 09:18:30 -06:00
James Betker
0551139b8d Fix resgen temperature curve below 1 
It needs to be inverted to maintain a true linear curve
 2020-06-18 16:08:07 -06:00
James Betker
efc80f041c Save & load amp state 2020-06-18 11:38:48 -06:00
James Betker
778e7b6931 Add a double-step to attention temperature 2020-06-18 11:29:31 -06:00
James Betker
d2d5e097d5 Add profiling to SRGAN for testing timings 2020-06-18 11:29:10 -06:00
James Betker
59b0533b06 Fix attimage step size 2020-06-17 18:45:24 -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
6f8406fbdc Fixed ConfigurableSwitchedGenerator bug 2020-06-16 16:53:57 -06:00
James Betker
7d541642aa Get rid of SwitchedResidualGenerator 
Just use the configurable one instead..
 2020-06-16 16:23:29 -06:00