James Betker
44b09e5f20
Amplify dropout rate
2021-06-07 15:20:53 -06:00
James Betker
f0d4eb9182
Fixor
2021-06-07 11:58:36 -06:00
James Betker
c456a60466
Another go at fixing nan
2021-06-07 11:51:43 -06:00
James Betker
1c574c5bd1
Attempt to fix nan
2021-06-07 11:43:42 -06:00
James Betker
eda796985b
Try out dropout norm
2021-06-07 11:33:33 -06:00
James Betker
6c6e82406e
Pass a corruption factor through the dataset into the upsampling network
...
The intuition is this will help guide the network to make better informed decisions
about how it performs upsampling based on how it perceives the underlying content.
(I'm giving up on letting networks detect their own quality - I'm not convinced it is
actually feasible)
2021-06-07 09:13:54 -06:00
James Betker
061dbcd458
Another fix to anorm
2021-06-06 15:09:49 -06:00
James Betker
9a6991e461
Fix switch norm average
2021-06-06 15:04:28 -06:00
James Betker
57e1a6a0f2
cifar: add hard routing
...
Also mods switched_routing to support non-pixular inputs
2021-06-06 14:53:43 -06:00
James Betker
692e9c417b
Support diffusion unet
2021-06-06 13:57:22 -06:00
James Betker
a0158ebc69
Simplify cifar resnet further for faster training
2021-06-06 10:02:24 -06:00
James Betker
75567a9814
Only head norm removed
2021-06-05 23:29:11 -06:00
James Betker
65d0376b90
Re-add normalization at the tail of the RRDB
2021-06-05 23:04:05 -06:00
James Betker
184e887122
Remove rrdb normalization
2021-06-05 21:39:19 -06:00
James Betker
f5e75602b9
Add regular attention to cifar_resnet
2021-06-05 21:34:07 -06:00
James Betker
af52751d6b
Fix device error
2021-06-05 14:21:32 -06:00
James Betker
5f0cc65f3b
Register branched resnet properly
2021-06-05 14:19:03 -06:00
James Betker
fb405d9ef1
CIFAR stuff
...
- Extract coarse labels for the CIFAR dataset
- Add simple resnet that branches lower layers based on coarse labels
- Some other cleanup
2021-06-05 14:16:02 -06:00
James Betker
80d4404367
A few fixes:
...
- Output better prediction of xstart from eps
- Support LossAwareSampler
- Support AdamW
2021-06-05 13:40:32 -06:00
James Betker
7c251af7a8
Support cifar100 with resnet
2021-06-04 17:29:07 -06:00
James Betker
bf811f80c1
GD mods & fixes
...
- Report variational loss separately
- Report model prediction from injector
- Log these things
- Use respacing like guided diffusion
2021-06-04 17:13:16 -06:00
James Betker
6084915af8
Support gaussian diffusion models
...
Adds support for GD models, courtesy of some maths from openai.
Also:
- Fixes requirement for eval{} even when it isn't being used
- Adds support for denormalizing an imagenet norm
2021-06-02 21:47:32 -06:00
James Betker
f129eaa39e
Clean up byol a bit
...
- Remove option to aug in dataset (there's really no reason for this now that kornia works on GPU on windows)
- Other stufff
2021-05-24 21:35:46 -06:00
James Betker
1a2b9fa130
Get rid of old byol net wrapping
...
Simplifies and makes this usable with DLAS' multi-gpu trainer
2021-04-27 12:48:34 -06:00
James Betker
119f17c808
Add testing capabilities for segformer & contrastive feature
2021-04-27 09:59:50 -06:00
James Betker
9bbe6fc81e
Get segformer to a trainable state
2021-04-25 11:45:20 -06:00
James Betker
fc623d4b5a
Add segformer model. Start work on BYOL adaptation that will support training it.
2021-04-23 17:16:46 -06:00
James Betker
17555e7d07
misc adjustments for stylegan
2021-04-21 18:14:17 -06:00
James Betker
b687ef4cd0
Misc
2021-04-21 18:09:46 -06:00
James Betker
9fc3df3f5b
Switched conv: add conversion function with allowlist
2021-03-13 10:44:56 -07:00
James Betker
cf9a6da889
Fix some bugs, checkin work on vqvae3
2021-03-02 20:56:19 -07:00
James Betker
f89ea5f1c6
Mods to support lightweight_gan model
2021-03-02 20:51:48 -07:00
James Betker
39fd755baa
New benchmark numbers
2021-02-08 08:09:41 -07:00
James Betker
784b96c059
Misc options to add support for training stylegan2-rosinality models:
...
- Allow image_folder_dataset to normalize inbound images
- ExtensibleTrainer can denormalize images on the output path
- Support .webp - an output from LSUN
- Support logistic GAN divergence loss
- Support stylegan2 TF weight extraction for discriminator
- New injector that produces latent noise (with separated paths)
- Modify FID evaluator to be operable with rosinality-style GANs
2021-02-08 08:09:21 -07:00
James Betker
e7be4bdff3
Revert
2021-02-05 08:43:07 -07:00
James Betker
6dec1f5968
Back to groupnorm
2021-02-05 08:42:11 -07:00
James Betker
336f807c8e
lambda2
2021-02-05 00:00:24 -07:00
James Betker
025a5867c4
Use syncbatchnorm instead
2021-02-04 22:26:36 -07:00
James Betker
bb79fafb89
Fix groupnorm specification
2021-02-04 22:15:38 -07:00
James Betker
43da1f9c4b
Convert lambda coupler to use groupnorm instead of batchnorm
2021-02-04 21:59:44 -07:00
James Betker
7070142805
Make vqvae3_hard more configurable
2021-02-04 09:03:22 -07:00
James Betker
b980028ca8
Add get_debug_values for vqvae_3_hardswitch
2021-02-03 14:12:24 -07:00
James Betker
1405ff06b8
Fix SwitchedConvHardRoutingFunction for current cuda router
2021-02-03 14:11:55 -07:00
James Betker
d7bec392dd
...
2021-02-02 23:50:25 -07:00
James Betker
b0a8fa00bc
Visual dbg in vqvae3hs
2021-02-02 23:50:01 -07:00
James Betker
f5f91850fd
hardswitch variant of vqvae3
2021-02-02 21:00:04 -07:00
James Betker
320edbaa3c
Move switched_conv logic around a bit
2021-02-02 20:41:24 -07:00
James Betker
0dca36946f
Hard Routing mods
...
- Turns out my custom convolution was RIDDLED with backwards bugs, which is
why the existing implementation wasn't working so well.
- Implements the switch logic from both Mixture of Experts and Switch Transformers
for testing purposes.
2021-02-02 20:35:58 -07:00
James Betker
29c1c3bede
Register vqvae3
2021-01-29 15:26:28 -07:00
James Betker
bc20b4739e
vqvae3
...
Changes VQVAE as so:
- Reverts back to smaller codebook
- Adds an additional conv layer at the highest resolution for both the encoder & decoder
- Uses LeakyReLU on trunk
2021-01-29 15:24:26 -07:00
James Betker
96bc80313c
Add switch norm, up dropout rate, detach selector
2021-01-26 09:31:53 -07:00
James Betker
2cdac6bd09
Add PWCNet for human optical flow
2021-01-25 08:25:44 -07:00
James Betker
51b63b2aa6
Add switched_conv with hard routing and make vqvae use it.
2021-01-25 08:25:29 -07:00
James Betker
ae4ff4a1e7
Enable lambda visualization
2021-01-23 15:53:27 -07:00
James Betker
10ec6bda1d
lambda nets in switched_conv and a vqvae to use it
2021-01-23 14:57:57 -07:00
James Betker
b374dcdd46
update vqvae to double codebook size for bottom quantizer
2021-01-23 13:47:07 -07:00
James Betker
1b8a26db93
New switched_conv
2021-01-23 13:46:30 -07:00
James Betker
d919ae7148
Add VQVAE with no Conv2dTranspose
2021-01-18 08:49:59 -07:00
James Betker
587a4f4050
resnet_unet_3
...
I'm being really lazy here - these nets are not really different from each other
except at which layer they terminate. This one terminates at 2x downsampling,
which is simply indicative of a direction I want to go for testing these pixpro networks.
2021-01-15 14:51:03 -07:00
James Betker
038b8654b6
Pixpro: unwrap losses
2021-01-13 11:54:25 -07:00
James Betker
8990801a3f
Fix pixpro stochastic sampling bugs
2021-01-13 11:34:24 -07:00
James Betker
19475a072f
Pixpro: Rather than using a latent square for pixpro, use an entirely stochastic sampling of the pixels
2021-01-13 11:26:51 -07:00
James Betker
d1007ccfe7
Adjustments to pixpro to allow training against networks with arbitrarily large structural latents
...
- The pixpro latent now rescales the latent space instead of using a "coordinate vector", which
**might** have performance implications.
- The latent against which the pixel loss is computed can now be a small, randomly sampled patch
out of the entire latent, allowing further memory/computational discounts. Since the loss
computation does not have a receptive field, this should not alter the loss.
- The instance projection size can now be separate from the pixel projection size.
- PixContrast removed entirely.
- ResUnet with full resolution added.
2021-01-12 09:17:45 -07:00
James Betker
34f8c8641f
Support training imagenet classifier
2021-01-11 20:09:16 -07:00
James Betker
f3db381fa1
Allow uresnet to use pretrained resnet50
2021-01-10 12:57:31 -07:00
James Betker
07168ecfb4
Enable vqvae to use a switched_conv variant
2021-01-09 20:53:14 -07:00
James Betker
5a8156026a
Did anyone ask for k-means clustering?
...
This is so cool...
2021-01-07 22:37:41 -07:00
James Betker
de10c7246a
Add injected noise into bypass maps
2021-01-07 16:31:12 -07:00
James Betker
61a86a3c1e
VQVAE
2021-01-07 10:20:15 -07:00
James Betker
01a589e712
Adjustments to pixpro & resnet-unet
...
I'm not really satisfied with what I got out of these networks on round 1.
Lets try again..
2021-01-06 15:00:46 -07:00
James Betker
2f2f87bbea
Styled SR fixes
2021-01-05 20:14:39 -07:00
James Betker
9fed90393f
Add lucidrains pixpro trainer
2021-01-05 20:14:22 -07:00
James Betker
ade2732c82
Transfer learning for styleSR
...
This is a concept from "Lifelong Learning GAN", although I'm skeptical of it's novelty -
basically you scale and shift the weights for the generator and discriminator of a pretrained
GAN to "shift" into new modalities, e.g. faces->birds or whatever. There are some interesting
applications of this that I would like to try out.
2021-01-04 20:10:48 -07:00
James Betker
2c65b6b28e
More mods to support styledsr
2021-01-04 11:32:28 -07:00
James Betker
2225fe6ac2
Undo lucidrains changes for new discriminator
...
This "new" code will live in the styledsr directory from now on.
2021-01-04 10:57:09 -07:00
James Betker
40ec71da81
Move styled_sr into its own folder
2021-01-04 10:54:34 -07:00
James Betker
5916f5f7d4
Misc fixes
2021-01-04 10:53:53 -07:00
James Betker
4d8064c32c
Modifications to allow partially trained stylegan discriminators to be used
2021-01-03 16:37:18 -07:00
James Betker
bdbab65082
Allow optimizers to train separate param groups, add higher dimensional VGG discriminator
...
Did this to support training 512x512px networks off of a pretrained 256x256 network.
2021-01-02 15:10:06 -07:00
James Betker
193cdc6636
Move discriminators to the create_model paradigm
...
Also cleans up a lot of old discriminator models that I have no intention
of using again.
2021-01-01 15:56:09 -07:00
James Betker
f39179e85a
styled_sr: fix bug when using initial_stride
2021-01-01 12:13:21 -07:00
James Betker
913fc3b75e
Need init to pick up styled_sr
2021-01-01 12:10:32 -07:00
James Betker
e992e18767
Add initial_stride term to style_sr
...
Also fix fid and a networks.py issue.
2021-01-01 11:59:36 -07:00
James Betker
e214e6ce33
Styled SR model
2020-12-31 20:54:18 -07:00
James Betker
b1fb82476b
Add gp debug (fix)
2020-12-30 15:26:54 -07:00
James Betker
63cf3d3126
Injector auto-registration
...
I love it!
2020-12-29 20:58:02 -07:00
James Betker
a777c1e4f9
Misc script fixes
2020-12-29 20:25:09 -07:00
James Betker
ba543d1152
Glean mods
...
- Fixes fixed upscale factor issues
- Refines a few ops to decrease computation & parameterization
2020-12-27 12:25:06 -07:00
James Betker
f9be049adb
GLEAN mod to support custom initial strides
2020-12-26 13:51:14 -07:00
James Betker
3fd627fc62
Mods to support image classification & filtering
2020-12-26 13:49:27 -07:00
James Betker
10fdfa1563
Migrate generators to dynamic model registration
2020-12-24 23:02:10 -07:00
James Betker
29db7c7a02
Further mods to BYOL
2020-12-24 09:28:41 -07:00
James Betker
036684893e
Add LARS optimizer & support for BYOL idiosyncrasies
...
- Added LARS and SGD optimizer variants that support turning off certain
features for BN and bias layers
- Added a variant of pytorch's resnet model that supports gradient checkpointing.
- Modify the trainer infrastructure to support above
- Fix bug with BYOL (should have been nonfunctional)
2020-12-23 20:33:43 -07:00
James Betker
1bbcb96ee8
Implement a few changes to support training BYOL networks
2020-12-23 10:50:23 -07:00
James Betker
ae666dc520
Fix bugs with srflow after refactor
2020-12-19 10:28:23 -07:00
James Betker
4328c2f713
Change default ReLU slope to .2 BREAKS COMPATIBILITY
...
This conforms my ConvGnLelu implementation with the generally accepted negative_slope=.2. I have no idea where I got .1. This will break backwards compatibility with some older models but will likely improve their performance when freshly trained. I did some auditing to find what these models might be, and I am not actively using any of them, so probably OK.
2020-12-19 08:28:03 -07:00
James Betker
9377d34ac3
glean mods
2020-12-19 08:26:07 -07:00
James Betker
92f9a129f7
GLEAN!
2020-12-18 16:04:19 -07:00
James Betker
c717765bcb
Notes for lucidrains converter.
2020-12-18 09:55:38 -07:00
James Betker
b4720ea377
Move stylegan to new location
2020-12-18 09:52:36 -07:00