aljaz/DALLE2-pytorch
1
1
Fork 0
mirror of https://github.com/lucidrains/DALLE2-pytorch.git synced 2025-12-19 17:54:20 +01:00
Code Issues Packages Projects Releases Wiki Activity

turn off classifier free guidance if predicting x_start for diffusion prior

Browse Source
This commit is contained in:
Phil Wang
2022-05-07 09:38:17 -07:00
parent c87b84a259
commit 4010aec033
2 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
dalle2_pytorch/dalle2_pytorch.py
View File

@@ -800,7 +800,7 @@ class DiffusionPrior(BaseGaussianDiffusion):
image_size = None, image_size = None,
image_channels = 3, image_channels = 3,
timesteps = 1000, timesteps = 1000,
cond_drop_prob = 0.2, cond_drop_prob = 0.,
loss_type = "l1", loss_type = "l1",
predict_x_start = True, predict_x_start = True,
beta_schedule = "cosine", beta_schedule = "cosine",
@@ -834,7 +834,7 @@ class DiffusionPrior(BaseGaussianDiffusion):
self.image_embed_dim = default(image_embed_dim, lambda: clip.dim_latent) self.image_embed_dim = default(image_embed_dim, lambda: clip.dim_latent)
self.channels = default(image_channels, lambda: clip.image_channels) self.channels = default(image_channels, lambda: clip.image_channels)
self.cond_drop_prob = cond_drop_prob self.cond_drop_prob = cond_drop_prob if not predict_x_start else 0.
self.condition_on_text_encodings = condition_on_text_encodings self.condition_on_text_encodings = condition_on_text_encodings
# in paper, they do not predict the noise, but predict x0 directly for image embedding, claiming empirically better results. I'll just offer both. # in paper, they do not predict the noise, but predict x0 directly for image embedding, claiming empirically better results. I'll just offer both.

2
setup.py
View File

@@ -10,7 +10,7 @@ setup(
'dream = dalle2_pytorch.cli:dream' 'dream = dalle2_pytorch.cli:dream'
], ],
}, },
version = '0.1.8', version = '0.1.9',
license='MIT', license='MIT',
description = 'DALL-E 2', description = 'DALL-E 2',
author = 'Phil Wang', author = 'Phil Wang',