Spaces:
Running
on
Zero
Running
on
Zero
File size: 7,487 Bytes
90a9dd3 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 |
import torch
from tqdm import tqdm
import torchvision.utils as tvu
import torchvision
import os
class_num = 951
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 inverse_data_transform(x):
x = (x + 1.0) / 2.0
return torch.clamp(x, 0.0, 1.0)
def ddnm_diffusion(x, model, b, eta, A_funcs, y, cls_fn=None, classes=None, config=None):
with torch.no_grad():
# setup iteration variables
skip = config.diffusion.num_diffusion_timesteps//config.time_travel.T_sampling
n = x.size(0)
x0_preds = []
xs = [x]
# generate time schedule
times = get_schedule_jump(config.time_travel.T_sampling,
config.time_travel.travel_length,
config.time_travel.travel_repeat,
)
time_pairs = list(zip(times[:-1], times[1:]))
# reverse diffusion sampling
for i, j in tqdm(time_pairs):
i, j = i*skip, j*skip
if j<0: j=-1
if j < i: # normal sampling
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')
if cls_fn == None:
et = model(xt, t)
else:
classes = torch.ones(xt.size(0), dtype=torch.long, device=torch.device("cuda"))*class_num
et = model(xt, t, classes)
et = et[:, :3]
et = et - (1 - at).sqrt()[0, 0, 0, 0] * cls_fn(x, t, classes)
if et.size(1) == 6:
et = et[:, :3]
x0_t = (xt - et * (1 - at).sqrt()) / at.sqrt()
x0_t_hat = x0_t - A_funcs.A_pinv(
A_funcs.A(x0_t.reshape(x0_t.size(0), -1)) - y.reshape(y.size(0), -1)
).reshape(*x0_t.size())
c1 = (1 - at_next).sqrt() * eta
c2 = (1 - at_next).sqrt() * ((1 - eta ** 2) ** 0.5)
xt_next = at_next.sqrt() * x0_t_hat + c1 * torch.randn_like(x0_t) + c2 * et
x0_preds.append(x0_t.to('cpu'))
xs.append(xt_next.to('cpu'))
else: # time-travel back
next_t = (torch.ones(n) * j).to(x.device)
at_next = compute_alpha(b, next_t.long())
x0_t = x0_preds[-1].to('cuda')
xt_next = at_next.sqrt() * x0_t + torch.randn_like(x0_t) * (1 - at_next).sqrt()
xs.append(xt_next.to('cpu'))
return [xs[-1]], [x0_preds[-1]]
def ddnm_plus_diffusion(x, model, b, eta, A_funcs, y, sigma_y, cls_fn=None, classes=None, config=None):
with torch.no_grad():
# setup iteration variables
skip = config.diffusion.num_diffusion_timesteps//config.time_travel.T_sampling
n = x.size(0)
x0_preds = []
xs = [x]
# generate time schedule
times = get_schedule_jump(config.time_travel.T_sampling,
config.time_travel.travel_length,
config.time_travel.travel_repeat,
)
time_pairs = list(zip(times[:-1], times[1:]))
# reverse diffusion sampling
for i, j in tqdm(time_pairs):
i, j = i*skip, j*skip
if j<0: j=-1
if j < i: # normal sampling
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')
if cls_fn == None:
et = model(xt, t)
else:
classes = torch.ones(xt.size(0), dtype=torch.long, device=torch.device("cuda"))*class_num
et = model(xt, t, classes)
et = et[:, :3]
et = et - (1 - at).sqrt()[0, 0, 0, 0] * cls_fn(x, t, classes)
if et.size(1) == 6:
et = et[:, :3]
# Eq. 12
x0_t = (xt - et * (1 - at).sqrt()) / at.sqrt()
sigma_t = (1 - at_next).sqrt()[0, 0, 0, 0]
# Eq. 17
x0_t_hat = x0_t - A_funcs.Lambda(A_funcs.A_pinv(
A_funcs.A(x0_t.reshape(x0_t.size(0), -1)) - y.reshape(y.size(0), -1)
).reshape(x0_t.size(0), -1), at_next.sqrt()[0, 0, 0, 0], sigma_y, sigma_t, eta).reshape(*x0_t.size())
# Eq. 51
xt_next = at_next.sqrt() * x0_t_hat + A_funcs.Lambda_noise(
torch.randn_like(x0_t).reshape(x0_t.size(0), -1),
at_next.sqrt()[0, 0, 0, 0], sigma_y, sigma_t, eta, et.reshape(et.size(0), -1)).reshape(*x0_t.size())
x0_preds.append(x0_t.to('cpu'))
xs.append(xt_next.to('cpu'))
else: # time-travel back
next_t = (torch.ones(n) * j).to(x.device)
at_next = compute_alpha(b, next_t.long())
x0_t = x0_preds[-1].to('cuda')
xt_next = at_next.sqrt() * x0_t + torch.randn_like(x0_t) * (1 - at_next).sqrt()
xs.append(xt_next.to('cpu'))
# #ablation
# if i%50==0:
# os.makedirs('/userhome/wyh/ddnm/debug/x0t', exist_ok=True)
# tvu.save_image(
# inverse_data_transform(x0_t[0]),
# os.path.join('/userhome/wyh/ddnm/debug/x0t', f"x0_t_{i}.png")
# )
# os.makedirs('/userhome/wyh/ddnm/debug/x0_t_hat', exist_ok=True)
# tvu.save_image(
# inverse_data_transform(x0_t_hat[0]),
# os.path.join('/userhome/wyh/ddnm/debug/x0_t_hat', f"x0_t_hat_{i}.png")
# )
# os.makedirs('/userhome/wyh/ddnm/debug/xt_next', exist_ok=True)
# tvu.save_image(
# inverse_data_transform(xt_next[0]),
# os.path.join('/userhome/wyh/ddnm/debug/xt_next', f"xt_next_{i}.png")
# )
return [xs[-1]], [x0_preds[-1]]
# form RePaint
def get_schedule_jump(T_sampling, travel_length, travel_repeat):
jumps = {}
for j in range(0, T_sampling - travel_length, travel_length):
jumps[j] = travel_repeat - 1
t = T_sampling
ts = []
while t >= 1:
t = t-1
ts.append(t)
if jumps.get(t, 0) > 0:
jumps[t] = jumps[t] - 1
for _ in range(travel_length):
t = t + 1
ts.append(t)
ts.append(-1)
_check_times(ts, -1, T_sampling)
return ts
def _check_times(times, t_0, T_sampling):
# Check end
assert times[0] > times[1], (times[0], times[1])
# Check beginning
assert times[-1] == -1, times[-1]
# Steplength = 1
for t_last, t_cur in zip(times[:-1], times[1:]):
assert abs(t_last - t_cur) == 1, (t_last, t_cur)
# Value range
for t in times:
assert t >= t_0, (t, t_0)
assert t <= T_sampling, (t, T_sampling)
|