Spaces:
Sleeping
Sleeping
File size: 16,929 Bytes
34d43ef bda45a0 34d43ef bda45a0 34d43ef bda45a0 34d43ef bda45a0 ab6a5ca bda45a0 34d43ef bda45a0 34d43ef bda45a0 34d43ef bda45a0 34d43ef |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 |
import torch
import numpy as np
from tqdm import tqdm
from functools import partial
from copy import deepcopy
from diffusers import AutoencoderKL, LMSDiscreteScheduler
from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like
import math
from ldm.models.diffusion.loss import caculate_loss_att_fixed_cnt, caculate_loss_self_att, caculate_loss_LoCo,caculate_loss_LAC, caculate_loss_LoCo_V2
class PLMSSampler(object):
def __init__(self, diffusion, model, schedule="linear", alpha_generator_func=None, set_alpha_scale=None):
super().__init__()
self.diffusion = diffusion
self.model = model
self.device = diffusion.betas.device
self.ddpm_num_timesteps = diffusion.num_timesteps
self.schedule = schedule
self.alpha_generator_func = alpha_generator_func
self.set_alpha_scale = set_alpha_scale
def register_buffer(self, name, attr):
if type(attr) == torch.Tensor:
attr = attr.to(self.device)
setattr(self, name, attr)
def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=False):
if ddim_eta != 0:
raise ValueError('ddim_eta must be 0 for PLMS')
self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
alphas_cumprod = self.diffusion.alphas_cumprod
assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.device)
self.register_buffer('betas', to_torch(self.diffusion.betas))
self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
self.register_buffer('alphas_cumprod_prev', to_torch(self.diffusion.alphas_cumprod_prev))
# calculations for diffusion q(x_t | x_{t-1}) and others
self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
# ddim sampling parameters
ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
ddim_timesteps=self.ddim_timesteps,
eta=ddim_eta,verbose=verbose)
self.register_buffer('ddim_sigmas', ddim_sigmas)
self.register_buffer('ddim_alphas', ddim_alphas)
self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
(1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
1 - self.alphas_cumprod / self.alphas_cumprod_prev))
self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
# @torch.no_grad()
def sample(self, S, shape, input, uc=None, guidance_scale=1, mask=None, x0=None, loss_type=None):
self.make_schedule(ddim_num_steps=S)
# import pdb; pdb.set_trace()
return self.plms_sampling(shape, input, uc, guidance_scale, mask=mask, x0=x0, loss_type=loss_type)
# @torch.no_grad()
def plms_sampling(self, shape, input, uc=None, guidance_scale=1, mask=None, x0=None, loss_type=None):
b = shape[0]
img = input["x"]
if img == None:
img = torch.randn(shape, device=self.device)
input["x"] = img
time_range = np.flip(self.ddim_timesteps)
total_steps = self.ddim_timesteps.shape[0]
old_eps = []
if self.alpha_generator_func != None:
alphas = self.alpha_generator_func(len(time_range))
# 新加的scheduler
noise_scheduler = LMSDiscreteScheduler(beta_start=0.00085, beta_end=0.012,
beta_schedule="scaled_linear", num_train_timesteps=1000)
noise_scheduler.set_timesteps(50)
for i, step in enumerate(time_range):
# set alpha and restore first conv layer
if self.alpha_generator_func != None:
self.set_alpha_scale(self.model, alphas[i])
if alphas[i] == 0:
self.model.restore_first_conv_from_SD()
# run
index = total_steps - i - 1
ts = torch.full((b,), step, device=self.device, dtype=torch.long)
ts_next = torch.full((b,), time_range[min(i + 1, len(time_range) - 1)], device=self.device, dtype=torch.long)
if mask is not None:
assert x0 is not None
img_orig = self.diffusion.q_sample(x0, ts)
img = img_orig * mask + (1. - mask) * img
input["x"] = img
# three loss types
if loss_type !=None and loss_type!='standard':
if input['object_position'] != []:
if loss_type=='SAR_CAR':
x = self.update_loss_self_cross( input,i, index, ts )
elif loss_type=='SAR':
x = self.update_only_self( input,i, index, ts )
elif loss_type=='CAR':
x = self.update_loss_only_cross( input,i, index, ts )
elif loss_type=='LoCo':
#print('Utilizing LoCo!!')
time_factor = noise_scheduler.sigmas[i] ** 2
x = self.update_loss_LoCo( input,i, index, ts, time_factor = time_factor)
elif loss_type=='LAC':
#print('Utilizing LoCo!!')
x = self.update_loss_LAC( input,i, index, ts )
input["x"] = x
img, pred_x0, e_t = self.p_sample_plms(input, ts, index=index, uc=uc, guidance_scale=guidance_scale, old_eps=old_eps, t_next=ts_next)
input["x"] = img
old_eps.append(e_t)
if len(old_eps) >= 4:
old_eps.pop(0)
return img
def update_loss_self_cross(self, input,index1, index, ts,type_loss='self_accross' ):
if index1 < 10:
loss_scale = 3
max_iter = 5
elif index1 < 20:
loss_scale = 2
max_iter = 3
else:
loss_scale = 1
max_iter = 1
loss_threshold = 0.1
max_index = 30
x = deepcopy(input["x"])
iteration = 0
loss = torch.tensor(10000)
input["timesteps"] = ts
print("optimize", index1)
while loss.item() > loss_threshold and iteration < max_iter and (index1 < max_index) :
print('iter', iteration)
x = x.requires_grad_(True)
input['x'] = x
e_t, att_first, att_second, att_third, self_first, self_second, self_third = self.model(input)
bboxes = input['boxes']
object_positions = input['object_position']
loss1 = caculate_loss_self_att(self_first, self_second, self_third, bboxes=bboxes,
object_positions=object_positions, t = index1)*loss_scale
loss2 = caculate_loss_att_fixed_cnt(att_second,att_first,att_third, bboxes=bboxes,
object_positions=object_positions, t = index1)*loss_scale
loss = loss1 + loss2
print('AR loss:', loss, 'SAR:', loss1, 'CAR:', loss2)
hh = torch.autograd.backward(loss)
grad_cond = x.grad
x = x - grad_cond
x = x.detach()
iteration += 1
torch.cuda.empty_cache()
return x
def update_loss_only_cross(self, input,index1, index, ts,type_loss='self_accross'):
if index1 < 10:
loss_scale = 3
max_iter = 5
elif index1 < 20:
loss_scale = 2
max_iter = 5
else:
loss_scale = 1
max_iter = 1
loss_threshold = 0.1
max_index = 30
x = deepcopy(input["x"])
iteration = 0
loss = torch.tensor(10000)
input["timesteps"] = ts
print("optimize", index1)
while loss.item() > loss_threshold and iteration < max_iter and (index1 < max_index) :
print('iter', iteration)
x = x.requires_grad_(True)
print('x shape', x.shape)
input['x'] = x
e_t, att_first, att_second, att_third, self_first, self_second, self_third = self.model(input)
bboxes = input['boxes']
object_positions = input['object_position']
loss2 = caculate_loss_att_fixed_cnt(att_second,att_first,att_third, bboxes=bboxes,
object_positions=object_positions, t = index1)*loss_scale
loss = loss2
print('loss', loss)
hh = torch.autograd.backward(loss, retain_graph=True)
grad_cond = x.grad
x = x - grad_cond
x = x.detach()
iteration += 1
torch.cuda.empty_cache()
return x
def update_loss_LoCo(self, input,index1, index, ts, time_factor, type_loss='self_accross'):
# loss_scale = 30
# max_iter = 5
#print('time_factor is: ', time_factor)
if index1 < 10:
loss_scale = 8
max_iter = 5
elif index1 < 20:
loss_scale = 5
max_iter = 5
else:
loss_scale = 1
max_iter = 1
loss_threshold = 0.1
max_index = 30
x = deepcopy(input["x"])
iteration = 0
loss = torch.tensor(10000)
input["timesteps"] = ts
# print("optimize", index1)
while loss.item() > loss_threshold and iteration < max_iter and (index1 < max_index) :
# print('iter', iteration)
x = x.requires_grad_(True)
# print('x shape', x.shape)
input['x'] = x
e_t, att_first, att_second, att_third, self_first, self_second, self_third = self.model(input)
bboxes = input['boxes']
object_positions = input['object_position']
loss2 = caculate_loss_LoCo_V2(att_second,att_first,att_third, bboxes=bboxes,
object_positions=object_positions, t = index1)*loss_scale
# loss = loss2
# loss.requires_grad_(True)
#print('LoCo loss', loss)
hh = torch.autograd.backward(loss2, retain_graph=True)
grad_cond = x.grad
x = x - grad_cond
x = x.detach()
iteration += 1
torch.cuda.empty_cache()
return x
def update_loss_LAC(self, input,index1, index, ts,type_loss='self_accross'):
# loss_scale = 30
# max_iter = 5
if index1 < 10:
loss_scale = 6
max_iter = 5
elif index1 < 20:
loss_scale = 4
max_iter = 3
else:
loss_scale = 1
max_iter = 1
loss_threshold = 0.002
max_index = 30
x = deepcopy(input["x"])
iteration = 0
loss = torch.tensor(10000)
input["timesteps"] = ts
print("optimize", index1)
while loss.item() > loss_threshold and iteration < max_iter and (index1 < max_index) :
print('iter', iteration)
x = x.requires_grad_(True)
# print('x shape', x.shape)
input['x'] = x
e_t, att_first, att_second, att_third, self_first, self_second, self_third = self.model(input)
bboxes = input['boxes']
object_positions = input['object_position']
loss2 = caculate_loss_LAC(att_second,att_first,att_third, bboxes=bboxes,
object_positions=object_positions, t = index1)*loss_scale
loss = loss2
print('LoCo loss', loss)
hh = torch.autograd.backward(loss, retain_graph=True)
grad_cond = x.grad
x = x - grad_cond
x = x.detach()
iteration += 1
torch.cuda.empty_cache()
return x
def update_only_self(self, input,index1, index, ts,type_loss='self_accross' ):
if index1 < 10:
loss_scale = 4
max_iter = 5
elif index1 < 20:
loss_scale = 3
max_iter = 5
else:
loss_scale = 1
max_iter = 1
loss_threshold = 0.1
max_index = 30
x = deepcopy(input["x"])
iteration = 0
loss = torch.tensor(10000)
input["timesteps"] = ts
print("optimize", index1)
while loss.item() > loss_threshold and iteration < max_iter and (index1 < max_index) :
print('iter', iteration)
x = x.requires_grad_(True)
input['x'] = x
e_t, att_first, att_second, att_third, self_first, self_second, self_third = self.model(input)
bboxes = input['boxes']
object_positions = input['object_position']
loss = caculate_loss_self_att(self_first, self_second, self_third, bboxes=bboxes,
object_positions=object_positions, t = index1)*loss_scale
print('loss', loss)
hh = torch.autograd.backward(loss)
grad_cond = x.grad
x = x - grad_cond
x = x.detach()
iteration += 1
torch.cuda.empty_cache()
return x
@torch.no_grad()
def p_sample_plms(self, input, t, index, guidance_scale=1., uc=None, old_eps=None, t_next=None):
x = deepcopy(input["x"])
b = x.shape[0]
def get_model_output(input):
e_t, first, second, third,_,_,_ = self.model(input)
if uc is not None and guidance_scale != 1:
unconditional_input = dict(x=input["x"], timesteps=input["timesteps"], context=uc, inpainting_extra_input=input["inpainting_extra_input"], grounding_extra_input=input['grounding_extra_input'])
e_t_uncond, _, _, _, _, _, _ = self.model( unconditional_input )
e_t = e_t_uncond + guidance_scale * (e_t - e_t_uncond)
return e_t
def get_x_prev_and_pred_x0(e_t, index):
# select parameters corresponding to the currently considered timestep
a_t = torch.full((b, 1, 1, 1), self.ddim_alphas[index], device=self.device)
a_prev = torch.full((b, 1, 1, 1), self.ddim_alphas_prev[index], device=self.device)
sigma_t = torch.full((b, 1, 1, 1), self.ddim_sigmas[index], device=self.device)
sqrt_one_minus_at = torch.full((b, 1, 1, 1), self.ddim_sqrt_one_minus_alphas[index],device=self.device)
# current prediction for x_0
pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
# direction pointing to x_t
dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
noise = sigma_t * torch.randn_like(x)
x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
return x_prev, pred_x0
input["timesteps"] = t
e_t = get_model_output(input)
if len(old_eps) == 0:
# Pseudo Improved Euler (2nd order)
x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index)
input["x"] = x_prev
input["timesteps"] = t_next
e_t_next = get_model_output(input)
e_t_prime = (e_t + e_t_next) / 2
elif len(old_eps) == 1:
# 2nd order Pseudo Linear Multistep (Adams-Bashforth)
e_t_prime = (3 * e_t - old_eps[-1]) / 2
elif len(old_eps) == 2:
# 3nd order Pseudo Linear Multistep (Adams-Bashforth)
e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12
elif len(old_eps) >= 3:
# 4nd order Pseudo Linear Multistep (Adams-Bashforth)
e_t_prime = (55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]) / 24
x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index)
return x_prev, pred_x0, e_t
|