-
Notifications
You must be signed in to change notification settings - Fork 33
/
reconstruction.py
executable file
·64 lines (52 loc) · 2.54 KB
/
reconstruction.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
from typing import Any
import torch
# from forward_process import *
import numpy as np
import os
os.environ['CUDA_VISIBLE_DEVICES'] = "0,1,2"
class Reconstruction:
'''
The reconstruction process
:param y: the target image
:param x: the input image
:param seq: the sequence of denoising steps
:param unet: the UNet model
:param x0_t: the prediction of x0 at time step t
'''
def __init__(self, unet, config) -> None:
self.unet = unet
self.config = config
def __call__(self, x, y0, w) -> Any:
def _compute_alpha(t):
betas = np.linspace(self.config.model.beta_start, self.config.model.beta_end, self.config.model.trajectory_steps, dtype=np.float64)
betas = torch.tensor(betas).type(torch.float).to(self.config.model.device)
beta = torch.cat([torch.zeros(1).to(self.config.model.device), betas], dim=0)
beta = beta.to(self.config.model.device)
a = (1 - beta).cumprod(dim=0).index_select(0, t + 1).view(-1, 1, 1, 1)
return a
test_trajectoy_steps = torch.Tensor([self.config.model.test_trajectoy_steps]).type(torch.int64).to(self.config.model.device).long()
at = _compute_alpha(test_trajectoy_steps)
xt = at.sqrt() * x + (1- at).sqrt() * torch.randn_like(x).to(self.config.model.device)
seq = range(0 , self.config.model.test_trajectoy_steps, self.config.model.skip)
with torch.no_grad():
n = x.size(0)
seq_next = [-1] + list(seq[:-1])
xs = [xt]
for index, (i, j) in enumerate(zip(reversed(seq), reversed(seq_next))):
t = (torch.ones(n) * i).to(self.config.model.device)
next_t = (torch.ones(n) * j).to(self.config.model.device)
at = _compute_alpha(t.long())
at_next = _compute_alpha(next_t.long())
xt = xs[-1].to(self.config.model.device)
self.unet = self.unet.to(self.config.model.device)
et = self.unet(xt, t)
yt = at.sqrt() * y0 + (1- at).sqrt() * et
et_hat = et - (1 - at).sqrt() * w * (yt-xt)
x0_t = (xt - et_hat * (1 - at).sqrt()) / at.sqrt()
c1 = (
self.config.model.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_hat
xs.append(xt_next)
return xs