forked from Hramchenko/diffusion_distiller
-
Notifications
You must be signed in to change notification settings - Fork 0
/
v_diffusion.py
185 lines (156 loc) · 7.15 KB
/
v_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
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
import math
import torch
import torch.nn.functional as F
def make_diffusion(model, n_timestep, time_scale, device):
betas = make_beta_schedule("cosine", cosine_s=8e-3, n_timestep=n_timestep).to(device)
return GaussianDiffusion(model, betas, time_scale=time_scale)
def make_beta_schedule(
schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3
):
if schedule == "cosine":
timesteps = (
torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s
)
alphas = timesteps / (1 + cosine_s) * math.pi / 2
alphas = torch.cos(alphas).pow(2)
alphas = alphas / alphas[0]
betas = 1 - alphas[1:] / alphas[:-1]
betas = betas.clamp(max=0.999)
else:
raise Exception()
return betas
def E_(input, t, shape):
out = torch.gather(input, 0, t)
reshape = [shape[0]] + [1] * (len(shape) - 1)
out = out.reshape(*reshape)
return out
def noise_like(shape, noise_fn, device, repeat=False):
if repeat:
resid = [1] * (len(shape) - 1)
shape_one = (1, *shape[1:])
return noise_fn(*shape_one, device=device).repeat(shape[0], *resid)
else:
return noise_fn(*shape, device=device)
class GaussianDiffusion:
def __init__(self, net, betas, time_scale=1, sampler="ddpm"):
super().__init__()
self.net_ = net
self.time_scale = time_scale
betas = betas.type(torch.float64)
self.num_timesteps = int(betas.shape[0])
alphas = 1 - betas
alphas_cumprod = torch.cumprod(alphas, 0)
alphas_cumprod_prev = torch.cat(
(torch.tensor([1], dtype=torch.float64, device=betas.device), alphas_cumprod[:-1]), 0
)
posterior_variance = betas * (1 - alphas_cumprod_prev) / (1 - alphas_cumprod)
self.betas = betas
self.alphas_cumprod = alphas_cumprod
self.posterior_variance = posterior_variance
self.sqrt_alphas_cumprod = torch.sqrt(alphas_cumprod)
self.sqrt_one_minus_alphas_cumprod = torch.sqrt(1 - alphas_cumprod)
self.posterior_log_variance_clipped = torch.log(posterior_variance.clamp(min=1e-20))
self.posterior_mean_coef1 = (betas * torch.sqrt(alphas_cumprod_prev) / (1 - alphas_cumprod))
self.posterior_mean_coef2 = (1 - alphas_cumprod_prev) * torch.sqrt(alphas) / (1 - alphas_cumprod)
if sampler == "ddpm":
self.p_sample = self.p_sample_ddpm
else:
self.p_sample = self.p_sample_clipped
def inference(self, x, t, extra_args):
return self.net_(x, t * self.time_scale, **extra_args)
def p_loss(self, x_0, t, extra_args, noise=None):
if noise is None:
noise = torch.randn_like(x_0)
alpha_t, sigma_t = self.get_alpha_sigma(x_0, t)
z = alpha_t * x_0 + sigma_t * noise
v_recon = self.inference(z.float(), t.float(), extra_args)
v = alpha_t * noise - sigma_t * x_0
return F.mse_loss(v_recon, v.float())
def q_posterior(self, x_0, x_t, t):
mean = E_(self.posterior_mean_coef1, t, x_t.shape) * x_0 \
+ E_(self.posterior_mean_coef2, t, x_t.shape) * x_t
var = E_(self.posterior_variance, t, x_t.shape)
log_var_clipped = E_(self.posterior_log_variance_clipped, t, x_t.shape)
return mean, var, log_var_clipped
def p_mean_variance(self, x, t, extra_args, clip_denoised):
v = self.inference(x.float(), t.float(), extra_args).double()
alpha_t, sigma_t = self.get_alpha_sigma(x, t)
x_recon = alpha_t * x - sigma_t * v
if clip_denoised:
x_recon = x_recon.clamp(min=-1, max=1)
mean, var, log_var = self.q_posterior(x_recon, x, t)
return mean, var, log_var
def p_sample_ddpm(self, x, t, extra_args, clip_denoised=True, **kwargs):
mean, _, log_var = self.p_mean_variance(x, t, extra_args, clip_denoised)
noise = torch.randn_like(x)
shape = [x.shape[0]] + [1] * (x.ndim - 1)
nonzero_mask = (1 - (t == 0).type(torch.float32)).view(*shape)
return mean + nonzero_mask * torch.exp(0.5 * log_var) * noise
def p_sample_clipped(self, x, t, extra_args, eta=0, clip_denoised=True, clip_value=3):
v = self.inference(x.float(), t, extra_args)
alpha, sigma = self.get_alpha_sigma(x, t)
# if clip_denoised:
# x = x.clip(-1, 1)
pred = (x * alpha - v * sigma)
if clip_denoised:
pred = pred.clip(-clip_value, clip_value)
eps = (x - alpha * pred) / sigma
if clip_denoised:
eps = eps.clip(-clip_value, clip_value)
t_mask = (t > 0)
if t_mask.any().item():
if not t_mask.all().item():
raise Exception()
alpha_, sigma_ = self.get_alpha_sigma(x, (t - 1).clip(min=0))
ddim_sigma = eta * (sigma_ ** 2 / sigma ** 2).sqrt() * \
(1 - alpha ** 2 / alpha_ ** 2).sqrt()
adjusted_sigma = (sigma_ ** 2 - ddim_sigma ** 2).sqrt()
pred = pred * alpha_ + eps * adjusted_sigma
if eta:
pred += torch.randn_like(pred) * ddim_sigma
return pred
@torch.no_grad()
def p_sample_loop(self, x, extra_args, eta=0):
mode = self.net_.training
self.net_.eval()
for i in reversed(range(self.num_timesteps)):
x = self.p_sample(
x,
torch.full((x.shape[0],), i, dtype=torch.int64).to(x.device),
extra_args,
eta=eta,
)
self.net_.train(mode)
return x
def get_alpha_sigma(self, x, t):
alpha = E_(self.sqrt_alphas_cumprod, t, x.shape)
sigma = E_(self.sqrt_one_minus_alphas_cumprod, t, x.shape)
return alpha, sigma
class GaussianDiffusionDefault(GaussianDiffusion):
def __init__(self, net, betas, time_scale=1, gamma=0.3):
super.__init__(net, betas, time_scale)
self.gamma = gamma
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
def distill_loss(self, student_diffusion, x, t, extra_args, eps=None, student_device=None):
if eps is None:
eps = torch.randn_like(x)
with torch.no_grad():
alpha, sigma = self.get_alpha_sigma(x, t + 1)
z = alpha * x + sigma * eps
alpha_s, sigma_s = student_diffusion.get_alpha_sigma(x, t // 2)
alpha_1, sigma_1 = self.get_alpha_sigma(x, t)
v = self.inference(z.float(), t.float() + 1, extra_args).double()
rec = (alpha * z - sigma * v).clip(-1, 1)
z_1 = alpha_1 * rec + (sigma_1 / sigma) * (z - alpha * rec)
v_1 = self.inference(z_1.float(), t.float(), extra_args).double()
x_2 = (alpha_1 * z_1 - sigma_1 * v_1).clip(-1, 1)
eps_2 = (z - alpha_s * x_2) / sigma_s
v_2 = alpha_s * eps_2 - sigma_s * x_2
if self.gamma == 0:
w = 1
else:
w = torch.pow(1 + alpha_s / sigma_s, self.gamma)
v = student_diffusion.net_(z.float(), t.float() * self.time_scale, **extra_args)
my_rec = (alpha_s * z - sigma_s * v).clip(-1, 1)
return F.mse_loss(w * v.float(), w * v_2.float())