-
Notifications
You must be signed in to change notification settings - Fork 34
/
Copy pathdiffusion.py
145 lines (117 loc) · 5.92 KB
/
diffusion.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
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
import numpy as np
import torch
import torch.nn as nn
from tqdm import trange
from torchvision.transforms import Compose
class Diffusion(nn.Module):
def __init__(
self, nn_backbone, device, n_timesteps=1000, in_channels=3, image_size=128, out_channels=6, motion_transforms=None):
super(Diffusion, self).__init__()
self.nn_backbone = nn_backbone
self.n_timesteps = n_timesteps
self.in_channels = in_channels
self.out_channels = out_channels
self.x_shape = (image_size, image_size)
self.device = device
self.motion_transforms = motion_transforms if motion_transforms else Compose([])
self.timesteps = torch.arange(n_timesteps)
self.beta = self.get_beta_schedule()
self.set_params()
self.device = device
def sample(self, x_cond, audio_emb, n_audio_motion_embs=2, n_motion_frames=2, motion_channels=3):
with torch.no_grad():
n_frames = audio_emb.shape[1]
xT = torch.randn(x_cond.shape[0], n_frames, self.in_channels, self.x_shape[0], self.x_shape[1]).to(x_cond.device)
audio_ids = [0] * n_audio_motion_embs
for i in range(n_audio_motion_embs + 1):
audio_ids += [i]
motion_frames = [self.motion_transforms(x_cond) for _ in range(n_motion_frames)]
motion_frames = torch.cat(motion_frames, dim=1)
samples = []
for i in trange(n_frames, desc=f'Sampling'):
sample_frame = self.sample_loop(xT[:, i].to(x_cond.device), x_cond, motion_frames, audio_emb[:, audio_ids])
samples.append(sample_frame.unsqueeze(1))
motion_frames = torch.cat([motion_frames[:, motion_channels:, :], self.motion_transforms(sample_frame)], dim=1)
audio_ids = audio_ids[1:] + [min(i + n_audio_motion_embs + 1, n_frames - 1)]
return torch.cat(samples, dim=1)
def sample_loop(self, xT, x_cond, motion_frames, audio_emb):
xt = xT
for i, t in reversed(list(enumerate(self.timesteps))):
timesteps = torch.tensor([t] * xT.shape[0]).to(xT.device)
timesteps_ids = torch.tensor([i] * xT.shape[0]).to(xT.device)
nn_out = self.nn_backbone(xt, timesteps, x_cond, motion_frames=motion_frames, audio_emb=audio_emb)
mean, logvar = self.get_p_params(xt, timesteps_ids, nn_out)
noise = torch.randn_like(xt) if t > 0 else torch.zeros_like(xt)
xt = mean + noise * torch.exp(logvar / 2)
return xt
def get_p_params(self, xt, timesteps, nn_out):
if self.in_channels == self.out_channels:
eps_pred = nn_out
p_logvar = self.expand(torch.log(self.beta[timesteps]))
else:
eps_pred, nu = nn_out.chunk(2, 1)
nu = (nu + 1) / 2
p_logvar = nu * self.expand(torch.log(self.beta[timesteps])) + (1 - nu) * self.expand(self.log_beta_tilde_clipped[timesteps])
p_mean, _ = self.get_q_params(xt, timesteps, eps_pred=eps_pred)
return p_mean, p_logvar
def get_q_params(self, xt, timesteps, eps_pred=None, x0=None):
if x0 is None:
# predict x0 from xt and eps_pred
coef1_x0 = self.expand(self.coef1_x0[timesteps])
coef2_x0 = self.expand(self.coef2_x0[timesteps])
x0 = coef1_x0 * xt - coef2_x0 * eps_pred
x0 = x0.clamp(-1, 1)
# q(x_{t-1} | x_t, x_0)
coef1_q = self.expand(self.coef1_q[timesteps])
coef2_q = self.expand(self.coef2_q[timesteps])
q_mean = coef1_q * x0 + coef2_q * xt
q_logvar = self.expand(self.log_beta_tilde_clipped[timesteps])
return q_mean, q_logvar
def get_beta_schedule(self, max_beta=0.999):
alpha_bar = lambda t: np.cos((t + 0.008) / 1.008 * np.pi / 2) ** 2
betas = []
for i in range(self.n_timesteps):
t1 = i / self.n_timesteps
t2 = (i + 1) / self.n_timesteps
betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
return torch.tensor(betas).float()
def set_params(self):
self.alpha = 1 - self.beta
self.alpha_bar = torch.cumprod(self.alpha, dim=0)
self.alpha_bar_prev = torch.cat([torch.ones(1,), self.alpha_bar[:-1]])
self.beta_tilde = self.beta * (1.0 - self.alpha_bar_prev) / (1.0 - self.alpha_bar)
self.log_beta_tilde_clipped = torch.log(torch.cat([self.beta_tilde[1, None], self.beta_tilde[1:]]))
# to caluclate x0 from eps_pred
self.coef1_x0 = torch.sqrt(1.0 / self.alpha_bar)
self.coef2_x0 = torch.sqrt(1.0 / self.alpha_bar - 1)
# for q(x_{t-1} | x_t, x_0)
self.coef1_q = self.beta * torch.sqrt(self.alpha_bar_prev) / (1.0 - self.alpha_bar)
self.coef2_q = (1.0 - self.alpha_bar_prev) * torch.sqrt(self.alpha) / (1.0 - self.alpha_bar)
def space(self, n_timesteps_new):
# change parameters for spaced timesteps during sampling
self.timesteps = self.space_timesteps(self.n_timesteps, n_timesteps_new)
self.n_timesteps = n_timesteps_new
self.beta = self.get_spaced_beta()
self.set_params()
def space_timesteps(self, n_timesteps, target_timesteps):
all_steps = []
frac_stride = (n_timesteps - 1) / (target_timesteps - 1)
cur_idx = 0.0
taken_steps = []
for _ in range(target_timesteps):
taken_steps.append(round(cur_idx))
cur_idx += frac_stride
all_steps += taken_steps
return all_steps
def get_spaced_beta(self):
last_alpha_cumprod = 1.0
new_beta = []
for i, alpha_cumprod in enumerate(self.alpha_bar):
if i in self.timesteps:
new_beta.append(1 - alpha_cumprod / last_alpha_cumprod)
last_alpha_cumprod = alpha_cumprod
return torch.tensor(new_beta)
def expand(self, arr, dim=4):
while arr.dim() < dim:
arr = arr[:, None]
return arr.to(self.device)