mrq/DL-Art-School
1
3
Fork 5
Code Issues Pull Requests Packages Projects Releases Wiki Activity
442 Commits 1 Branch 0 Tags 5.6 MiB
f5cd23e2d5
Commit Graph

62 Commits

Author SHA1 Message Date
James Betker
f5cd23e2d5 Further patch size adjustments 2020-09-16 16:50:35 -06:00
James Betker
0b047e5f80 Increase scale of the patch selector random distribution 
This will cause larger slices of an image to appear more frequently,
increasing the difficulty of the generator.
 2020-09-16 08:27:42 -06:00
James Betker
94deab2792 Fix error serving gt_fullsize_ref in full_image_dataset 2020-09-14 15:05:44 -06:00
James Betker
5189b11dac Add combined dataset for training across multiple datasets 2020-09-11 08:44:06 -06:00
James Betker
3027e6e27d Enable amp to be disabled 2020-09-09 10:45:59 -06:00
James Betker
a5c2388368 Use lower LQ image size when it is being fed in 2020-09-06 17:26:32 -06:00
James Betker
17aa205e96 New dataset that reads from lmdb 2020-09-04 17:32:57 -06:00
James Betker
8580490a85 Reduce usage of resize operations when not needed in dataloaders. 2020-09-04 15:31:24 -06:00
James Betker
6226b52130 Pin memory in dataloaders by default 2020-09-04 15:30:46 -06:00
James Betker
03eb29a4d9 Fix LQGT dataset 2020-08-25 11:57:25 -06:00
James Betker
a65b07607c Reference network 2020-08-25 11:56:59 -06:00
James Betker
f224907603 Fix LQGT_dataset, add full_image_dataset 2020-08-24 17:12:43 -06:00
James Betker
5ec04aedc8 Let noise be configurable 
LQ noise is not currently configurable for some reason..
 2020-08-24 15:00:14 -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
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
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
0892d5fe99 LQGT_dataset gan debug 2020-07-26 22:48:35 -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
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
8a0a1569f3 Enable force_multiple in LQ_dataset 2020-07-22 20:51:16 -06:00
James Betker
ef9f1307eb Sometimes don't use compression artifacts 2020-07-10 22:25:53 -06:00
James Betker
ed6a15e768 Add feature to dataset which allows it to force images to be a certain size. 2020-07-03 15:19:16 -06:00
James Betker
296135ec18 Add doResizeLoss to dataset 
doResizeLoss has a 50% chance to resize the LQ image to 50% size,
then back to original size. This is useful to training a generator to
recover these lost pixel values while also being able to do
repairs on higher resolution images during training.
 2020-06-08 11:27:06 -06:00
James Betker
318a604405 Allow weighting of input data 
This essentially allows you to give some datasets more
importance than others for the purposes of reaching a more
refined network.
 2020-06-04 10:05:21 -06:00
James Betker
90125f5bed Allow blurring to be specified 2020-06-02 08:40:52 -06:00
James Betker
1eb9c5a059 Fix grayscale 2020-05-29 22:04:50 -06:00
James Betker
74b313aaa9 Add grayscale downsampling option 2020-05-29 20:34:00 -06:00
James Betker
6962ccb306 Adjust motion blur 
0 is invalid.
 2020-05-27 13:09:46 -06:00
James Betker
e27a49454e Enable vertical splitting on inference images to support very large resolutions. 2020-05-27 08:04:35 -06:00
James Betker
96ac26a8b7 Allow injection of random low-amplitude noise & motion blur into generator 2020-05-27 08:04:11 -06:00
James Betker
69cbfa2f0c Adjust dataset mutations a bit 
Adjusts the compression artfacts to be more aggressive, and blurring to be less aggressive.
 2020-05-26 13:48:34 -06:00
James Betker
2931142458 Allow multiple gt image dirs 2020-05-25 19:21:09 -06:00
James Betker
4e44b8a1aa Clean up video stuff 2020-05-25 19:20:49 -06:00
James Betker
8464cae168 HQ blurring doesnt actually work right - hq images arent the right size when they are blurred 
Just revert it and blur the lq images..
 2020-05-24 22:32:54 -06:00
James Betker
5fd8749cf2 More updates - need more blurring 2020-05-24 22:13:27 -06:00
James Betker
9627cc2c49 Update HR gaussian blur params 2020-05-24 18:00:31 -06:00
James Betker
2f8b0250b9 Blur HR image before downsizing, when available 2020-05-24 17:18:44 -06:00
James Betker
cc4571eb8d Randomize blur effect 2020-05-24 12:35:41 -06:00
James Betker
27a548c019 Enable blurring via settings 2020-05-24 11:56:39 -06:00
James Betker
90073fc761 Update LQ_dataset to support inference on split image videos 2020-05-23 21:05:49 -06:00
James Betker
74bb0fad33 Allow dataset classes to add noise internally 2020-05-23 21:04:24 -06:00
James Betker
67139602f5 Test modifications 
Allows bifurcating large images put into the test pipeline

This code is fixed and not dynamic. Needs some fixes.
 2020-05-19 09:37:58 -06:00
James Betker
d72e154442 Allow more LQ than GT images in corrupt mode 2020-05-14 20:46:20 -06:00
James Betker
585b05e66b Cap test workers at 10 2020-05-13 09:20:45 -06:00
James Betker
f994466289 Initialize test dataloader with a worker count proportional to the batch size. 2020-05-10 10:49:37 -06:00
James Betker
8969a3ce70 Add capability to start at arbitrary frames 2020-05-10 10:48:05 -06:00
James Betker
dbca0d328c Fix multi-lq bug 2020-05-06 23:16:35 -06:00
James Betker
5c1832e124 Add support for multiple LQ paths 
I want to be able to specify many different transformations onto
the target data; the model should handle them all. Do this by
allowing multiple LQ paths to be selected and the dataset class
selects one at random.
 2020-05-06 17:24:17 -06:00
James Betker
44b89330c2 Support inference across batches, support inference on cpu, checkpoint 
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.
 2020-05-04 08:48:25 -06:00
James Betker
61d3040cf5 Add doCrop into LQGT 2020-05-02 17:46:30 -06:00