LYL1015's picture
test
eea83e8
Raw
History Blame Contribute Delete
3.73 kB
import torch
from torchvision.transforms.functional import crop
# This script is adapted from the following repository: https://github.com/ermongroup/ddim
def compute_alpha(beta, t):
beta = torch.cat([torch.zeros(1).to(beta.device), beta], dim=0)
a = (1 - beta).cumprod(dim=0).index_select(0, t + 1).view(-1, 1, 1, 1)
return a
def data_transform(X):
return 2 * X - 1.0
def inverse_data_transform(X):
return torch.clamp((X + 1.0) / 2.0, 0.0, 1.0)
def generalized_steps(x, x_cond, seq, model, b, eta=0.):
with torch.no_grad():
n = x.size(0)
seq_next = [-1] + list(seq[:-1])
x0_preds = []
xs = [x]
for i, j in zip(reversed(seq), reversed(seq_next)):
t = (torch.ones(n) * i).to(x.device)
next_t = (torch.ones(n) * j).to(x.device)
at = compute_alpha(b, t.long())
at_next = compute_alpha(b, next_t.long())
xt = xs[-1].to('cuda')
et = model(torch.cat([x_cond, xt], dim=1), t)
x0_t = (xt - et * (1 - at).sqrt()) / at.sqrt()
x0_preds.append(x0_t.to('cpu'))
c1 = eta * ((1 - at / at_next) * (1 - at_next) / (1 - at)).sqrt()
c2 = ((1 - at_next) - c1 ** 2).sqrt()
xt_next = at_next.sqrt() * x0_t + c1 * torch.randn_like(x) + c2 * et
xs.append(xt_next.to('cpu'))
return xs, x0_preds
def generalized_steps_overlapping(x, x_cond, seq, model, b, eta=0., corners=None, p_size=None, manual_batching=True):
with torch.no_grad():
n = x.size(0)
seq_next = [-1] + list(seq[:-1])
x0_preds = []
xs = [x]
x_grid_mask = torch.zeros_like(x_cond, device=x.device)
for (hi, wi) in corners:
x_grid_mask[:, :, hi:hi + p_size, wi:wi + p_size] += 1
for i, j in zip(reversed(seq), reversed(seq_next)):
t = (torch.ones(n) * i).to(x.device)
next_t = (torch.ones(n) * j).to(x.device)
at = compute_alpha(b, t.long())
at_next = compute_alpha(b, next_t.long())
xt = xs[-1].to('cuda')
et_output = torch.zeros_like(x_cond, device=x.device)
if manual_batching:
manual_batching_size = 64
xt_patch = torch.cat([crop(xt, hi, wi, p_size, p_size) for (hi, wi) in corners], dim=0)
x_cond_patch = torch.cat([data_transform(crop(x_cond, hi, wi, p_size, p_size)) for (hi, wi) in corners], dim=0)
for i in range(0, len(corners), manual_batching_size):
outputs = model(torch.cat([x_cond_patch[i:i+manual_batching_size],
xt_patch[i:i+manual_batching_size]], dim=1), t)
for idx, (hi, wi) in enumerate(corners[i:i+manual_batching_size]):
et_output[0, :, hi:hi + p_size, wi:wi + p_size] += outputs[idx]
else:
for (hi, wi) in corners:
xt_patch = crop(xt, hi, wi, p_size, p_size)
x_cond_patch = crop(x_cond, hi, wi, p_size, p_size)
x_cond_patch = data_transform(x_cond_patch)
et_output[:, :, hi:hi + p_size, wi:wi + p_size] += model(torch.cat([x_cond_patch, xt_patch], dim=1), t)
et = torch.div(et_output, x_grid_mask)
x0_t = (xt - et * (1 - at).sqrt()) / at.sqrt()
x0_preds.append(x0_t.to('cpu'))
c1 = eta * ((1 - at / at_next) * (1 - at_next) / (1 - at)).sqrt()
c2 = ((1 - at_next) - c1 ** 2).sqrt()
xt_next = at_next.sqrt() * x0_t + c1 * torch.randn_like(x) + c2 * et
xs.append(xt_next.to('cpu'))
return xs, x0_preds