Mirrors
/
diffusers
mirror of https://github.com/ShivamShrirao/diffusers.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

improve ddim comments

This commit is contained in:
Patrick von Platen 2022-06-09 16:30:56 +02:00
parent cbb19ee84e
commit f035fbfba7
1 changed files with 23 additions and 23 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

46
models/vision/ddpm/modeling_ddpm.py
View File

@ -30,43 +30,43 @@ class DDPM(DiffusionPipeline):
torch_device = "cuda" if torch.cuda.is_available() else "cpu" torch_device = "cuda" if torch.cuda.is_available() else "cpu"
self.unet.to(torch_device) self.unet.to(torch_device)
# 1. Sample gaussian noise
# Sample gaussian noise to begin loop
image = self.noise_scheduler.sample_noise( image = self.noise_scheduler.sample_noise(
(batch_size, self.unet.in_channels, self.unet.resolution, self.unet.resolution), (batch_size, self.unet.in_channels, self.unet.resolution, self.unet.resolution),
device=torch_device, device=torch_device,
generator=generator, generator=generator,
) )
for t in tqdm.tqdm(reversed(range(len(self.noise_scheduler))), total=len(self.noise_scheduler)):
# i) define coefficients for time step t
clipped_image_coeff = 1 / torch.sqrt(self.noise_scheduler.get_alpha_prod(t))
clipped_noise_coeff = torch.sqrt(1 / self.noise_scheduler.get_alpha_prod(t) - 1)
image_coeff = (
(1 - self.noise_scheduler.get_alpha_prod(t - 1))
* torch.sqrt(self.noise_scheduler.get_alpha(t))
/ (1 - self.noise_scheduler.get_alpha_prod(t))
)
clipped_coeff = (
torch.sqrt(self.noise_scheduler.get_alpha_prod(t - 1))
* self.noise_scheduler.get_beta(t)
/ (1 - self.noise_scheduler.get_alpha_prod(t))
)
# ii) predict noise residual for t in tqdm.tqdm(reversed(range(len(self.noise_scheduler))), total=len(self.noise_scheduler)):
# 1. predict noise residual
with torch.no_grad(): with torch.no_grad():
noise_residual = self.unet(image, t) noise_residual = self.unet(image, t)
# iii) compute predicted image from residual # 2. compute alphas, betas
# See 2nd formula at https://github.com/hojonathanho/diffusion/issues/5#issue-896554416 for comparison alpha_prod_t = self.noise_scheduler.get_alpha_prod(t)
pred_mean = clipped_image_coeff * image - clipped_noise_coeff * noise_residual alpha_prod_t_prev = self.noise_scheduler.get_alpha_prod(t - 1)
pred_mean = torch.clamp(pred_mean, -1, 1) beta_prod_t = 1 - alpha_prod_t
prev_image = clipped_coeff * pred_mean + image_coeff * image beta_prod_t_prev = 1 - alpha_prod_t_prev
# iv) sample variance # 3. compute predicted image from residual
# See 2nd formula at https://github.com/hojonathanho/diffusion/issues/5#issue-896554416 for comparison
# First: Compute inner formula
pred_mean = (1 / alpha_prod_t.sqrt()) * (image - beta_prod_t.sqrt() * noise_residual)
# Second: Clip
pred_mean = torch.clamp(pred_mean, -1, 1)
# Third: Compute outer coefficients
pred_mean_coeff = (alpha_prod_t_prev.sqrt() * self.noise_scheduler.get_beta(t)) / beta_prod_t
image_coeff = (beta_prod_t_prev * self.noise_scheduler.get_alpha(t).sqrt()) / beta_prod_t
# Fourth: Compute outer formula
prev_image = pred_mean_coeff * pred_mean + image_coeff * image
# 4. sample variance
prev_variance = self.noise_scheduler.sample_variance( prev_variance = self.noise_scheduler.sample_variance(
t, prev_image.shape, device=torch_device, generator=generator t, prev_image.shape, device=torch_device, generator=generator
) )
# v) sample x_{t-1} ~ N(prev_image, prev_variance) # 5. sample x_{t-1} ~ N(prev_image, prev_variance) = add variance to predicted image
sampled_prev_image = prev_image + prev_variance sampled_prev_image = prev_image + prev_variance
image = sampled_prev_image image = sampled_prev_image