 Phil Wang
|
4284c8840d
|
clipping for continuous time diffusion not working
|
2022-06-08 16:26:18 -07:00 |
|
|
Phil Wang
|
c44d3ea01d
|
learned noise schedule seems to be working, allow for one to make the monotonic net learn a bit more slowly than the unet
|
2022-06-08 12:34:07 -07:00 |
|
|
Phil Wang
|
c4991f576f
|
allow for configuring the hidden dimension of the monotonic mlp parameterizing the noise schedule
|
2022-06-08 11:18:54 -07:00 |
|
|
Phil Wang
|
a19331aa59
|
fix learned noise schedule
|
2022-06-08 10:19:02 -07:00 |
|
|
Phil Wang
|
94eabaca1a
|
complete learned noise schedule for variational ddpm paper, still need to finish cosine alpha schedule in log(snr) form
|
2022-06-08 09:47:09 -07:00 |
|
|
Phil Wang
|
eaf9d9fdc4
|
unet needs to be conditioned on log(snr) in p_mean_variance for continuous time gaussian diffusion
|
2022-06-08 00:41:41 -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
|
532178a6a3
|
assume when sampling all batch samples are at the same time, and do not noise for the last time step
|
2022-06-07 16:12:44 -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
|
cf6db71985
|
add gaussian diffusion where model predicts both noise and x_start, with a learned weighting between the two (experimental)
|
2022-05-13 13:56:32 -07:00 |
|
|
Phil Wang
|
84ebb9ad13
|
offer predict_x0 objective
|
2022-05-13 10:15:54 -07:00 |
|
|
Phil Wang
|
caa5af170d
|
final cleanup
|
2022-05-12 13:58:16 -07:00 |
|
|
Phil Wang
|
e0f26677d6
|
make sure predicted mean is actually detached for all of the kl loss calculations
|
2022-05-12 11:12:34 -07:00 |
|
|
Phil Wang
|
e147839d74
|
make sure to clip when sampling from gaussian diffusion with learned variance
|
2022-05-12 10:08:53 -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
|
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 |
|
|
Phil Wang
|
e274fb305a
|
give an initial conv
|
2022-05-01 08:49:38 -07:00 |
|
|
Phil Wang
|
f39b3b1d3f
|
make sure time embedding dimension is kept at 4 x dimension (thanks @borisdayma)
|
2022-04-29 14:55:12 -07:00 |
|
|
Phil Wang
|
782c904d3b
|
fix cosine beta schedule, thanks to @Zhengxinyang
|
2022-04-19 20:51:50 -07:00 |
|
|
Phil Wang
|
71953ebd22
|
fix bug, thanks to @jihoonerd
|
2022-04-15 06:37:31 -07:00 |
|
|
Phil Wang
|
0b8cdb4c8b
|
remove outdated apex in favor of native pytorch AMP
|
2022-04-13 08:59:18 -07:00 |
|
|
Phil Wang
|
bd1e3b676e
|
get rid of numpy
|
2022-04-12 11:58:46 -07:00 |
|
|
Phil Wang
|
f4615599bc
|
use full attention at the center of the unet
|
2022-04-04 09:03:41 -07:00 |
|
|
Phil Wang
|
eb6e1b508e
|
greater kernel size in convnext blocks
|
2022-01-31 17:13:27 -08:00 |
|
|
Phil Wang
|
91cff45939
|
replace resnets with convnext blocks
|
2022-01-25 09:02:45 -08:00 |
|
|
Phil Wang
|
7b51e30da7
|
fix layernorm
|
2021-08-24 14:28:15 -07:00 |
|
|
Phil Wang
|
dadbf20154
|
remove stray print
|
2021-07-16 15:18:20 -07:00 |
|
|
Phil Wang
|
7706bdfc6f
|
use pre-layernorm with linear attention, and also allow for turning off time embedding
|
2021-06-25 10:48:11 -07:00 |
|
|
Phil Wang
|
183e5f3cc5
|
move all constants into configurable class init parameters
|
2021-06-25 10:37:36 -07:00 |
|
|
Phil Wang
|
16c9ae7bb3
|
fix data not being normalized to range of -1 to 1
|
2021-06-21 18:51:36 -07:00 |
|
|
Phil Wang
|
f5916111f8
|
0.6.3
|
2021-06-21 17:41:16 -07:00 |
|
|
Phil Wang
|
ad9e303ff3
|
fix channels
|
2021-06-11 15:28:36 -07:00 |
|
|
Phil Wang
|
ae42f48f6a
|
prepare so that unet can work with a channel of one, and also make it so image size is hard coded in diffusion class. preparing for training on protein distograms
|
2021-06-11 14:06:39 -07:00 |
|
|
Phil Wang
|
d4ce9f6c38
|
save samples and models to ./results path
|
2020-10-09 21:50:23 -07:00 |
|
|
Phil Wang
|
ff451f697e
|
update with new and improved cosine noise scheduler
|
2020-10-09 21:21:02 -07:00 |
|
|
Phil Wang
|
ef2ca0b625
|
new paper suggests image linear attention is more effective without query normalization
|
2020-10-04 21:53:53 -07:00 |
|
|
Phil Wang
|
9f95a03c07
|
fix bug with rezero and linear attention
|
2020-09-21 20:11:34 -07:00 |
|
|
Phil Wang
|
a4c68d3569
|
fix bug
|
2020-09-15 15:57:39 -07:00 |
|
|
Phil Wang
|
b33a48e342
|
make sure when sampling, batch does not exceed training batch size
|
2020-09-15 15:15:10 -07:00 |
|
|
Phil Wang
|
4bf28914bc
|
allow for mixed precision training with fp16 flag
|
2020-09-08 17:26:23 -07:00 |
|