|
|
import torch |
|
|
import torchvision |
|
|
from torchvision.utils import save_image |
|
|
import os |
|
|
from config import Config |
|
|
|
|
|
def simple_sample(model, noise_scheduler, device, epoch=None, writer=None, n_samples=4): |
|
|
"""Standard DDPM sampling - this should actually work""" |
|
|
config = Config() |
|
|
model.eval() |
|
|
|
|
|
with torch.no_grad(): |
|
|
|
|
|
x = torch.randn(n_samples, 3, config.image_size, config.image_size, device=device) |
|
|
|
|
|
print(f"Starting reverse diffusion for {n_samples} samples...") |
|
|
|
|
|
|
|
|
alphas = noise_scheduler.alphas.to(device) |
|
|
alpha_bars = noise_scheduler.alpha_bars.to(device) |
|
|
betas = noise_scheduler.betas.to(device) |
|
|
|
|
|
|
|
|
for step, t in enumerate(reversed(range(config.T))): |
|
|
if step % 100 == 0: |
|
|
print(f"Step {step}/{config.T}, t={t}") |
|
|
|
|
|
t_tensor = torch.full((n_samples,), t, device=device, dtype=torch.long) |
|
|
|
|
|
|
|
|
pred_noise = model(x, t_tensor) |
|
|
|
|
|
|
|
|
alpha_t = alphas[t] |
|
|
alpha_bar_t = alpha_bars[t] |
|
|
beta_t = betas[t] |
|
|
|
|
|
|
|
|
if t > 0: |
|
|
alpha_bar_prev = alpha_bars[t-1] |
|
|
|
|
|
|
|
|
pred_x0 = (x - torch.sqrt(1 - alpha_bar_t) * pred_noise) / torch.sqrt(alpha_bar_t) |
|
|
|
|
|
|
|
|
mean = (torch.sqrt(alpha_bar_prev) * beta_t / (1 - alpha_bar_t)) * pred_x0 + \ |
|
|
(torch.sqrt(alpha_t) * (1 - alpha_bar_prev) / (1 - alpha_bar_t)) * x |
|
|
|
|
|
|
|
|
noise = torch.randn_like(x) |
|
|
variance = (1 - alpha_bar_prev) / (1 - alpha_bar_t) * beta_t |
|
|
x = mean + torch.sqrt(variance) * noise |
|
|
else: |
|
|
|
|
|
x = (x - torch.sqrt(1 - alpha_bar_t) * pred_noise) / torch.sqrt(alpha_bar_t) |
|
|
|
|
|
|
|
|
x = torch.clamp(x, -1, 1) |
|
|
|
|
|
|
|
|
if epoch is not None and epoch % 10 == 0: |
|
|
print(f"Sample stats at epoch {epoch}: range [{x.min().item():.3f}, {x.max().item():.3f}], mean {x.mean().item():.3f}") |
|
|
|
|
|
grid = torchvision.utils.make_grid(x, nrow=2, normalize=True) |
|
|
|
|
|
if writer: |
|
|
writer.add_image('Samples', grid, epoch) |
|
|
|
|
|
if epoch is not None: |
|
|
os.makedirs("samples", exist_ok=True) |
|
|
save_image(grid, f"samples/epoch_{epoch}.png") |
|
|
|
|
|
return x, grid |
|
|
|
|
|
|
|
|
def sample(model, noise_scheduler, device, epoch=None, writer=None, n_samples=4): |
|
|
return simple_sample(model, noise_scheduler, device, epoch, writer, n_samples) |
|
|
|