aljaz/DALLE2-pytorch
1
1
Fork 0
mirror of https://github.com/lucidrains/DALLE2-pytorch.git synced 2026-02-21 03:04:20 +01:00
Code Issues Packages Projects Releases Wiki Activity

schedule to Prior

Browse Source
This commit is contained in:
Kashif Rasul
2022-04-17 15:21:47 +02:00
parent 51361c2d15
commit b0f2fbaa95
1 changed files with 17 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
dalle2_pytorch/dalle2_pytorch.py
View File

@@ -450,10 +450,11 @@ class DiffusionPrior(nn.Module):
net, net,
*, *,
clip, clip,
timesteps = 1000, timesteps=1000,
cond_drop_prob = 0.2, cond_drop_prob=0.2,
loss_type = 'l1', loss_type="l1",
predict_x0 = True predict_x0=True,
beta_schedule="cosine",
): ):
super().__init__() super().__init__()
assert isinstance(clip, CLIP) assert isinstance(clip, CLIP)
@@ -469,7 +470,18 @@ class DiffusionPrior(nn.Module):
self.predict_x0 = predict_x0 self.predict_x0 = predict_x0
# 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.
betas = cosine_beta_schedule(timesteps) if beta_schedule == "cosine":
betas = cosine_beta_schedule(timesteps)
elif beta_schedule == "linear":
betas = linear_beta_schedule(timesteps)
elif beta_schedule == "quadratic":
betas = quadratic_beta_schedule(timesteps)
elif beta_schedule == "jsd":
betas = 1.0 / torch.linspace(timesteps, 1, timesteps)
elif beta_schedule == "sigmoid":
betas = sigmoid_beta_schedule(timesteps)
else:
raise NotImplementedError()
alphas = 1. - betas alphas = 1. - betas
alphas_cumprod = torch.cumprod(alphas, axis=0) alphas_cumprod = torch.cumprod(alphas, axis=0)