Phil Wang
|
9878be760b
|
have researcher explicitly state upfront whether to condition with text encodings in cascading ddpm decoder, have DALLE-2 class take care of passing in text if feature turned on
|
2022-04-26 09:47:09 -07:00 |
|
Phil Wang
|
7ba6357c05
|
allow for training the Prior network with precomputed CLIP embeddings (or text encodings)
|
2022-04-26 09:29:51 -07:00 |
|
Phil Wang
|
76e063e8b7
|
refactor so that the causal transformer in the diffusion prior network can be conditioned without text encodings (for Laions parallel efforts, although it seems from the paper it is needed)
|
2022-04-26 09:00:11 -07:00 |
|
Phil Wang
|
4d25976f33
|
make sure non-latent diffusion still works
|
2022-04-26 08:36:00 -07:00 |
|
Phil Wang
|
0b28ee0d01
|
revert back to old upsampling, paper does not work
|
2022-04-26 07:39:04 -07:00 |
|
Phil Wang
|
f75d49c781
|
start a file for all attention-related modules, use attention-based upsampling in the unets in dalle-2
|
2022-04-25 18:59:10 -07:00 |
|
Phil Wang
|
8f2a0c7e00
|
better naming
|
2022-04-25 07:44:33 -07:00 |
|
Phil Wang
|
863f4ef243
|
just take care of the logic for setting all latent diffusion to predict x0, if needed
|
2022-04-24 10:06:42 -07:00 |
|
Phil Wang
|
fb8a66a2de
|
just in case latent diffusion performs better with prediction of x0 instead of epsilon, open up the research avenue
|
2022-04-24 10:04:22 -07:00 |
|
Phil Wang
|
579d4b42dd
|
does not seem right to clip for the prior diffusion part
|
2022-04-24 09:51:18 -07:00 |
|
Phil Wang
|
473808850a
|
some outlines to the eventual CLI endpoint
|
2022-04-24 09:27:15 -07:00 |
|
Phil Wang
|
05b74be69a
|
use null container pattern to cleanup some conditionals, save more cleanup for next week
|
2022-04-22 15:23:18 -07:00 |
|
Phil Wang
|
76b32f18b3
|
first pass at complete DALL-E2 + Latent Diffusion integration, latent diffusion on any layer(s) of the cascading ddpm in the decoder.
|
2022-04-22 13:53:13 -07:00 |
|
Phil Wang
|
46cef31c86
|
optional projection out for prior network causal transformer
|
2022-04-22 11:16:30 -07:00 |
|
Phil Wang
|
59b1a77d4d
|
be a bit more conservative and stick with layernorm (without bias) for now, given @borisdayma results https://twitter.com/borisdayma/status/1517227191477571585
|
2022-04-22 11:14:54 -07:00 |
|
Phil Wang
|
7f338319fd
|
makes more sense for blur augmentation to happen before the upsampling
|
2022-04-22 11:10:47 -07:00 |
|
Phil Wang
|
2c6c91829d
|
refactor blurring training augmentation to be taken care of by the decoder, with option to downsample to previous resolution before upsampling (cascading ddpm). this opens up the possibility of cascading latent ddpm
|
2022-04-22 11:09:17 -07:00 |
|
Phil Wang
|
ad17c69ab6
|
prepare for latent diffusion in the first DDPM of the cascade in the Decoder
|
2022-04-21 17:54:31 -07:00 |
|
Phil Wang
|
faebf4c8b8
|
from my vision transformer experience, dimension of attention head of 32 is sufficient for image feature maps
|
2022-04-20 11:40:32 -07:00 |
|
Phil Wang
|
f37c26e856
|
cleanup and DRY a little
|
2022-04-20 10:56:32 -07:00 |
|
Phil Wang
|
27a33e1b20
|
complete contextmanager method for keeping only one unet in GPU during training or inference
|
2022-04-20 10:46:13 -07:00 |
|
Phil Wang
|
6f941a219a
|
give time tokens a surface area of 2 tokens as default, make it so researcher can customize which unet actually is conditioned on image embeddings and/or text encodings
|
2022-04-20 10:04:47 -07:00 |
|
Phil Wang
|
ddde8ca1bf
|
fix cosine bbeta schedule, thanks to @Zhengxinyang
|
2022-04-19 20:54:28 -07:00 |
|
Phil Wang
|
a35c309b5f
|
add sparse attention layers in between convnext blocks in unet (grid like attention, used in mobilevit, maxvit [bytedance ai], as well as a growing number of attention-based GANs)
|
2022-04-19 09:49:03 -07:00 |
|
Phil Wang
|
82328f16cd
|
same for text encodings for decoder ddpm training
|
2022-04-18 14:41:02 -07:00 |
|
Phil Wang
|
6fee4fce6e
|
also allow for image embedding to be passed into the diffusion model, in the case one wants to generate image embedding once and then train multiple unets in one iteration
|
2022-04-18 14:00:38 -07:00 |
|
Phil Wang
|
960a79857b
|
use some magic just this once to remove the need for researchers to think
|
2022-04-18 12:40:43 -07:00 |
|
Phil Wang
|
00ae50999b
|
make kernel size and sigma for gaussian blur for cascading DDPM overridable at forward. also make sure unets are wrapped in a modulelist so that at sample time, blurring does not happen
|
2022-04-18 12:04:31 -07:00 |
|
Phil Wang
|
0332eaa6ff
|
complete first pass at full cascading DDPM setup in Decoder, flexible enough to support one unet for testing
|
2022-04-18 11:44:56 -07:00 |
|
Kashif Rasul
|
b0f2fbaa95
|
schedule to Prior
|
2022-04-17 15:21:47 +02:00 |
|
Kashif Rasul
|
51361c2d15
|
added beta_schedule argument
|
2022-04-17 15:19:33 +02:00 |
|
Kashif Rasul
|
42d6e47387
|
added huber loss and other schedulers
|
2022-04-17 15:14:05 +02:00 |
|
Phil Wang
|
c400d8758c
|
prepare for cascading diffusion in unet, save the full progressive upsampling architecture to be built next week
|
2022-04-15 07:03:28 -07:00 |
|
Phil Wang
|
bece206699
|
fix bug thanks to @jihoonerd
|
2022-04-15 06:44:40 -07:00 |
|
Phil Wang
|
6e27f617f1
|
use t5 relative positional bias in prior network causal transformer, since it makes more sense than rotary embeddings
|
2022-04-14 12:01:09 -07:00 |
|
Phil Wang
|
9f55c24db6
|
allow for decoder conditioning with the text encodings from CLIP, if it is passed in. use lazy linear to avoid researchers having to worry about text encoding dimensions, but remove later if it does not work well
|
2022-04-14 11:46:45 -07:00 |
|
Phil Wang
|
23c401a5d5
|
use the eval decorator
|
2022-04-14 10:13:43 -07:00 |
|
Phil Wang
|
68e9883f59
|
use cross attention for conditioning unet based on image embedding tokens (which opens up the door on conditioning on text encodings as well
|
2022-04-14 10:10:04 -07:00 |
|
Phil Wang
|
95b018374a
|
start using swish glu everywhere, given success of PaLM
|
2022-04-14 09:34:32 -07:00 |
|
Phil Wang
|
8b5c2385b0
|
better naming
|
2022-04-14 09:24:31 -07:00 |
|
Phil Wang
|
f2c52d8239
|
fix bug with classifier free guidance for prior network, even though it seems it may not be used
|
2022-04-14 09:21:51 -07:00 |
|
Phil Wang
|
97e951221b
|
bring in blur, as it will be used somewhere in the cascading DDPM in the decoder eventually, once i figure it out
|
2022-04-14 09:16:09 -07:00 |
|
Phil Wang
|
82464d7bd3
|
per-fect
|
2022-04-14 08:30:07 -07:00 |
|
Phil Wang
|
7fb3f695d5
|
offer continuously parameterized time embedding for diffusion prior network, remove a hyperparameter that may trip up people, if not set correctly
|
2022-04-14 08:28:11 -07:00 |
|
Phil Wang
|
7e93b9d3c8
|
make sure classifier free guidance condition scaling is exposed on DALLE2 forward function
|
2022-04-13 20:14:28 -07:00 |
|
Phil Wang
|
14ddbc159c
|
cleanup
|
2022-04-13 18:24:32 -07:00 |
|
Phil Wang
|
5e06cde4cb
|
always work in the l2normed space for image and text embeddings
|
2022-04-13 18:08:42 -07:00 |
|
Phil Wang
|
a1a8a78f21
|
fix everything and make sure it runs end to end, document everything in readme for public
|
2022-04-13 18:05:25 -07:00 |
|
Phil Wang
|
e5e415297c
|
prepare non-causal attention, for use in the unet in the decoder
|
2022-04-13 12:04:09 -07:00 |
|
Phil Wang
|
c9377efc93
|
go for the multi-headed queries, one-headed key/values, proven out in AlphaCode as well as PaLM by now
|
2022-04-13 12:01:43 -07:00 |
|