 lukovnikov
|
cd8329cdd7
|
Revert "fix ddim sampling"
This reverts commit aec2a26984.
|
2022-08-31 15:25:51 +02:00 |
|
|
lukovnikov
|
aec2a26984
|
fix ddim sampling
|
2022-08-31 15:24:59 +02:00 |
|
|
TheDudeFromCI
|
e46a89e2bc
|
Update step before training checkpoint
|
2022-08-31 02:19:12 -07:00 |
|
|
Phil Wang
|
42158d6248
|
fix ddim, for issue https://github.com/lucidrains/denoising-diffusion-pytorch/issues/81
|
2022-08-30 20:30:29 -07:00 |
|
|
Phil Wang
|
d9275a744c
|
add weight standardization prior to groupnorm, lessen cosine sim attention scale to 10 for fp16
|
2022-08-17 11:42:54 -07:00 |
|
|
Phil Wang
|
f0d59acdfd
|
fix sampling ddpm tqdm
|
2022-08-10 12:02:56 -07:00 |
|
|
Phil Wang
|
689593a579
|
add the new self conditioning technique from hintons group from bit diffusion paper
0.27.0
|
2022-08-10 10:52:07 -07:00 |
|
|
Phil Wang
|
eba44498d1
|
higher epsilon for fp16 in layernorm
|
2022-07-29 13:29:52 -07:00 |
|
|
Phil Wang
|
12f95b33d8
|
rescale values to prevent linear attention from overflowing in fp16 setting
|
2022-07-27 12:25:40 -07:00 |
|
|
Phil Wang
|
6b504c4ae9
|
fix accelerator prepare bug for dataloader
|
2022-07-25 08:13:46 -07:00 |
|
|
Phil Wang
|
37334ae824
|
fix a bug with ddim and predict x0 objective
|
2022-07-18 19:04:57 -07:00 |
|
|
Phil Wang
|
555566c188
|
take a gamble on cosine sim attention
|
2022-07-18 11:29:20 -07:00 |
|
|
Phil Wang
|
2b742dd2cc
|
move accelerator backward outside of autocast context, also calculate total loss correctly across gradient accumulated steps
|
2022-07-11 21:02:06 -07:00 |
|
|
Phil Wang
|
1345a8a41d
|
do not noise at the last timestep for ddim
|
2022-07-09 18:36:45 -07:00 |
|
|
Phil Wang
|
931a5af2c3
|
bring in ddim sampling
|
2022-07-09 16:10:23 -07:00 |
|
|
Phil Wang
|
a0c3443eaa
|
optimizer should be saved and loaded
|
2022-07-08 17:43:41 -07:00 |
|
|
Phil Wang
|
662172851b
|
add convert_image_to keyword argument, for forcing images being loaded to be converted to some format, greyscale, rgb, rgba, whatever
|
2022-07-08 09:22:59 -07:00 |
|
|
Phil Wang
|
0248b5e4d3
|
also make sure grad scaler actually exists in the saved pt file
|
2022-07-06 11:48:04 -07:00 |
|
|
Phil Wang
|
6b56af08a2
|
support multi-gpu training using huggingface accelerate, addressing https://github.com/lucidrains/denoising-diffusion-pytorch/pull/54
|
2022-07-06 11:45:56 -07:00 |
|
|
Phil Wang
|
d4420248f1
|
tqdm auto instead
|
2022-06-29 19:48:49 -07:00 |
|
|
Arya Aftab
|
d97bc0278c
|
Solve issue #26
|
2022-06-29 11:42:54 +04:30 |
|
|
Phil Wang
|
d26acbcae6
|
more skip connections, as in guided diffusion
|
2022-06-27 13:23:32 -07:00 |
|
|
Phil Wang
|
9939a48139
|
make sure all versions of torch supported
|
2022-06-23 12:28:09 -07:00 |
|
|
Phil Wang
|
75ea49a7ef
|
pass parameter for Trainer to EMA properly
|
2022-06-21 07:37:43 -07:00 |
|
|
Phil Wang
|
8c3609a6e3
|
move EMA logic out of the repository for clarity
|
2022-06-20 13:17:51 -07:00 |
|
|
Phil Wang
|
1586d1a8a0
|
just pluck the image size off the gaussian diffusion class
|
2022-06-17 13:54:41 -07:00 |
|
|
Phil Wang
|
b4fb8804d2
|
conditioning on final resnet block
|
2022-06-17 10:38:17 -07:00 |
|
|
Phil Wang
|
9fd05f1b1f
|
switch to learned sinsuoidal pos emb for the continuous case
|
2022-06-17 09:24:51 -07:00 |
|
|
Phil Wang
|
ec2397f0ba
|
add one more residual
|
2022-06-16 11:08:42 -07:00 |
|
|
Phil Wang
|
844e557dfb
|
fix a missing residual needed at the top most resolution in the unet
|
2022-06-15 19:10:05 -07:00 |
|
|
Phil Wang
|
8b30be8042
|
add p2 loss reweighting for default ddpm as an option
|
2022-06-14 10:49:13 -07:00 |
|
|
Phil Wang
|
ecc6f30901
|
for https://github.com/lucidrains/denoising-diffusion-pytorch/issues/36
|
2022-06-11 10:51:13 -07:00 |
|
|
Phil Wang
|
f900f40f14
|
allow for turning off horizontal flip augmentation
|
2022-06-09 20:59:59 -07:00 |
|
|
Phil Wang
|
3bf5e768c2
|
use a non-sinusoidal embedded condition for continuous time gaussian diffusion conditioned on log(snr)
|
2022-06-07 21:15:27 -07:00 |
|
|
Phil Wang
|
3bbb6ebf16
|
get working version of gaussian diffusion with continuous time (only beta linear schedule for now, but will eventually contain alpha cosine schedule as well as parameterized, learned monotonic MLP)
|
2022-06-07 15:59:29 -07:00 |
|
|
Phil Wang
|
a291da5098
|
bring back linear noise schedule, but default to cosine
|
2022-05-27 19:13:05 -07:00 |
|
|
Phil Wang
|
e5a18bb25c
|
switch over to film like conditioning, used by both openai and google at this point
|
2022-05-24 23:47:34 -07:00 |
|
|
Phil Wang
|
fc8e4547aa
|
higher default learning rate
|
2022-05-16 13:39:55 -07:00 |
|
|
Phil Wang
|
cae9f4a71f
|
whoops
|
2022-05-14 13:59:21 -07:00 |
|
|
Phil Wang
|
91f03fb88b
|
optimize for simplicity and clarity - researcher does not need to worry about normalizing and unnormalizing now
|
2022-05-14 11:38:43 -07:00 |
|
|
Phil Wang
|
60128257c5
|
use tqdm pbar during training
|
2022-05-13 20:25:49 -07:00 |
|
|
Phil Wang
|
84ebb9ad13
|
offer predict_x0 objective
|
2022-05-13 10:15:54 -07:00 |
|
|
Phil Wang
|
55c658b967
|
cleanup unused
|
2022-05-12 11:52:06 -07:00 |
|
|
Phil Wang
|
62e8490385
|
complete the gaussian diffusion with hybrid loss (learned variance) as in the improved ddpm paper
|
2022-05-12 08:54:47 -07:00 |
|
|
Phil Wang
|
d412d8816b
|
first pass at ddpm with learned variance
|
2022-05-11 17:38:29 -07:00 |
|
|
Phil Wang
|
402b7c26df
|
calculate noise schedule with float64 for numerical accuracy
|
2022-05-10 15:23:34 -07:00 |
|
|
Phil Wang
|
989f0fcb8e
|
remove convnext blocks, they do not work well, validated in video diffusion repository
|
2022-05-05 07:03:55 -07:00 |
|
|
Phil Wang
|
84731bb03d
|
groupnorm groups should be actually configurable
|
2022-05-04 10:38:29 -07:00 |
|
|
Phil Wang
|
c6ecca555b
|
allow for configuring expansion factor in convnext
|
2022-05-04 10:33:23 -07:00 |
|
|
Phil Wang
|
1f5c233072
|
bring back resnet blocks, make convnext blocks an experimental option
|
2022-05-04 10:30:09 -07:00 |
|