Phil Wang
fbba0f9aaf
bring in prediction of v objective, combining the findings from progressive distillation paper and imagen-video to the eventual extension of dalle2 to make-a-video
2022-10-28 18:21:07 -07:00
Romain Beaumont
9f37705d87
Add static graph param ( #226 )
...
* Add static graph param
* use static graph param
2022-10-25 19:31:29 +02:00
Phil Wang
c3df46e374
fix openclipadapter to be able to use latest open sourced sota model
2022-10-23 15:12:09 -07:00
Phil Wang
41fabf2922
fix a dtype conversion issue for the diffusion timesteps in the diffusion prior, thanks to @JiaHeng-DLUT
2022-10-19 09:26:06 -07:00
Heng Jia
5975e8222b
Fix assert message ( #253 )
2022-10-18 08:50:59 -07:00
Phil Wang
c18c080128
fix for use with larger openai clip models by extracting dimension of last layernorm in clip
2022-09-29 09:09:47 -07:00
Phil Wang
d0c11b30b0
handle open clip adapter image size being a tuple
2022-09-19 10:27:14 -07:00
Phil Wang
0d82dff9c5
in ddim, noise should be predicted after x0 is maybe clipped, thanks to @lukovnikov for pointing this out in another repository
2022-09-01 09:40:47 -07:00
Phil Wang
8bbc956ff1
fix bug with misnamed variable in diffusion prior network
2022-08-31 17:19:05 -07:00
Phil Wang
6fb7e91343
fix ddim to use alpha_cumprod
2022-08-31 07:40:46 -07:00
Phil Wang
ba58ae0bf2
add two asserts to diffusion prior to ensure matching image embedding dimensions for clip, diffusion prior network, and what was set on diffusion prior
2022-08-28 10:11:37 -07:00
Phil Wang
1cc5d0afa7
upgrade to best downsample
2022-08-25 10:37:02 -07:00
Phil Wang
59fa101c4d
fix classifier free guidance for diffusion prior, thanks to @jaykim9870 for spotting the issue
2022-08-23 08:29:01 -07:00
Aidan
cbaadb6931
Fixed issues with clip and deepspeed fp16
...
Also more more general compatibility fixes
2022-08-20 17:58:32 +00:00
Phil Wang
083508ff8e
cast attention matrix back to original dtype pre-softmax in attention
2022-08-20 10:56:01 -07:00
Phil Wang
7762edd0ff
make it work for @ethancohen123
2022-08-19 11:28:58 -07:00
Phil Wang
27f19ba7fa
make sure diffusion prior trainer can operate with no warmup
2022-08-15 14:27:40 -07:00
Phil Wang
8f38339c2b
give diffusion prior trainer cosine annealing lr too
2022-08-15 07:38:01 -07:00
Phil Wang
6b9b4b9e5e
add cosine annealing lr schedule
2022-08-15 07:29:56 -07:00
Phil Wang
44e09d5a4d
add weight standardization behind feature flag, which may potentially work well with group norm
2022-08-14 11:34:45 -07:00
Phil Wang
34806663e3
make it so diffusion prior p_sample_loop returns unnormalized image embeddings
2022-08-13 10:03:40 -07:00
Phil Wang
dc816b1b6e
dry up some code around handling unet outputs with learned variance
2022-08-12 15:25:03 -07:00
Phil Wang
05192ffac4
fix self conditioning shape in diffusion prior
2022-08-12 12:30:03 -07:00
Phil Wang
9440411954
make self conditioning technique work with diffusion prior
2022-08-12 12:20:51 -07:00
Phil Wang
981d407792
comment
2022-08-12 11:41:23 -07:00
Phil Wang
7c5477b26d
bet on the new self-conditioning technique out of geoffrey hintons group
2022-08-12 11:36:08 -07:00
Phil Wang
be3bb868bf
add gradient checkpointing for all resnet blocks
2022-08-02 19:21:44 -07:00
Phil Wang
451de34871
enforce clip anytorch version
2022-07-30 10:07:55 -07:00
Phil Wang
f22e8c8741
make open clip available for use with dalle2 pytorch
2022-07-30 09:02:31 -07:00
Phil Wang
87432e93ad
quick fix for linear attention
2022-07-29 13:17:12 -07:00
Phil Wang
d167378401
add cosine sim for self attention as well, as a setting
2022-07-29 12:48:20 -07:00
Phil Wang
2d67d5821e
change up epsilon in layernorm the case of using fp16, thanks to @Veldrovive for figuring out this stabilizes training
2022-07-29 12:41:02 -07:00
Phil Wang
748c7fe7af
allow for cosine sim cross attention, modify linear attention in attempt to resolve issue on fp16
2022-07-29 11:12:18 -07:00
Phil Wang
80046334ad
make sure entire readme runs without errors
2022-07-28 10:17:43 -07:00
Phil Wang
36fb46a95e
fix readme and a small bug in DALLE2 class
2022-07-28 08:33:51 -07:00
Phil Wang
07abfcf45b
rescale values in linear attention to mitigate overflows in fp16 setting
2022-07-27 12:27:38 -07:00
Phil Wang
406e75043f
add upsample combiner feature for the unets
2022-07-26 10:46:04 -07:00
Phil Wang
9646dfc0e6
fix path_or_state bug
2022-07-26 09:47:54 -07:00
Phil Wang
62043acb2f
fix repaint
2022-07-24 15:29:06 -07:00
Phil Wang
417ff808e6
1.0.3
2022-07-22 13:16:57 -07:00
Aidan Dempster
f3d7e226ba
Changed types to be generic instead of functions ( #215 )
...
This allows pylance to do proper type hinting and makes developing
extensions to the package much easier
2022-07-22 13:16:29 -07:00
Phil Wang
48a1302428
1.0.2
2022-07-20 23:01:51 -07:00
Aidan Dempster
ccaa46b81b
Re-introduced change that was accidentally rolled back ( #212 )
2022-07-20 23:01:19 -07:00
Phil Wang
76d08498cc
diffusion prior training updates from @nousr
2022-07-20 18:05:27 -07:00
zion
f9423d308b
Prior updates ( #211 )
...
* update configs for prior
add prior warmup to config
update example prior config
* update prior trainer & script
add deepspeed amp & warmup
adopt full accelerator support
reload at sample point
finish epoch resume code
* update tracker save method for prior
* helper functions for prior_loader
2022-07-20 18:04:26 -07:00
Phil Wang
06c65b60d2
1.0.0
2022-07-19 19:08:17 -07:00
Aidan Dempster
4145474bab
Improved upsampler training ( #181 )
...
Sampling is now possible without the first decoder unet
Non-training unets are deleted in the decoder trainer since they are never used and it is harder merge the models is they have keys in this state dict
Fixed a mistake where clip was not re-added after saving
2022-07-19 19:07:50 -07:00
Phil Wang
4b912a38c6
0.26.2
2022-07-19 17:50:36 -07:00
Aidan Dempster
f97e55ec6b
Quality of life improvements for tracker savers ( #210 )
...
The default save location is now none so if keys are not specified the
corresponding checkpoint type is not saved.
Models and checkpoints are now both saved with version number and the
config used to create them in order to simplify loading.
Documentation was fixed to be in line with current usage.
2022-07-19 17:50:18 -07:00
Phil Wang
291377bb9c
@jacobwjs reports dynamic thresholding works very well and 0.95 is a better value
2022-07-19 11:31:56 -07:00
