File size: 46,239 Bytes
95456ed | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 | """
This code is adapted from Gong et al.'s 2023 DiffuSeq Model: https://github.com/Shark-NLP/DiffuSeq
"""
import math
import numpy as np
import torch as th
import sys
import os
import torch.nn
from .utils.nn import mean_flat
sys.path.append(".")
def get_named_beta_schedule(schedule_name, num_diffusion_timesteps):
"""
Get a pre-defined beta schedule for the given name.
The beta schedule library consists of beta schedules which remain similar
in the limit of num_diffusion_timesteps.
Beta schedules may be added, but should not be removed or changed once
they are committed to maintain backwards compatibility.
"""
if schedule_name == "linear":
# Linear schedule from Ho et al, extended to work for any number of
# diffusion steps.
scale = 1000 / num_diffusion_timesteps
beta_start = scale * 0.0001
beta_end = scale * 0.02
return np.linspace(beta_start, beta_end, num_diffusion_timesteps, dtype=np.float64)
elif schedule_name == "cosine":
return betas_for_alpha_bar(
num_diffusion_timesteps,
lambda t: math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2,
)
elif schedule_name == "sqrt":
return betas_for_alpha_bar(
num_diffusion_timesteps,
lambda t: 1 - np.sqrt(t + 0.0001),
)
elif schedule_name == "trunc_cos":
return betas_for_alpha_bar_left(
num_diffusion_timesteps,
lambda t: np.cos((t + 0.1) / 1.1 * np.pi / 2) ** 2,
)
elif schedule_name == "trunc_lin":
scale = 1000 / num_diffusion_timesteps
beta_start = scale * 0.0001 + 0.01
beta_end = scale * 0.02 + 0.01
return np.linspace(beta_start, beta_end, num_diffusion_timesteps, dtype=np.float64)
elif schedule_name == "pw_lin":
scale = 1000 / num_diffusion_timesteps
beta_start = scale * 0.0001 + 0.01
beta_mid = scale * 0.0001 # scale * 0.02
beta_end = scale * 0.02
first_part = np.linspace(beta_start, beta_mid, 10, dtype=np.float64)
second_part = np.linspace(
beta_mid, beta_end, num_diffusion_timesteps - 10, dtype=np.float64
)
return np.concatenate([first_part, second_part])
else:
raise NotImplementedError(f"unknown beta schedule: {schedule_name}")
def betas_for_alpha_bar_left(num_diffusion_timesteps, alpha_bar, max_beta=0.999):
"""
Create a beta schedule that discretizes the given alpha_t_bar function, but shifts towards left interval starting from 0
which defines the cumulative product of (1-beta) over time from t = [0,1].
:param num_diffusion_timesteps: the number of betas to produce.
:param alpha_bar: a lambda that takes an argument t from 0 to 1 and
produces the cumulative product of (1-beta) up to that
part of the diffusion process.
:param max_beta: the maximum beta to use; use values lower than 1 to
prevent singularities.
"""
betas = []
betas.append(min(1 - alpha_bar(0), max_beta))
for i in range(num_diffusion_timesteps - 1):
t1 = i / num_diffusion_timesteps
t2 = (i + 1) / num_diffusion_timesteps
betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
return np.array(betas)
def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999):
"""
Create a beta schedule that discretizes the given alpha_t_bar function,
which defines the cumulative product of (1-beta) over time from t = [0,1].
:param num_diffusion_timesteps: the number of betas to produce.
:param alpha_bar: a lambda that takes an argument t from 0 to 1 and
produces the cumulative product of (1-beta) up to that
part of the diffusion process.
:param max_beta: the maximum beta to use; use values lower than 1 to
prevent singularities.
"""
betas = []
for i in range(num_diffusion_timesteps):
t1 = i / num_diffusion_timesteps
t2 = (i + 1) / num_diffusion_timesteps
betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
return np.array(betas)
class GaussianDiffusion:
"""
Utilities for training and sampling diffusion models.
"""
def __init__(
self,
*,
betas,
predict_xstart,
rescale_learned_sigmas,
learn_sigmas,
sigma_small,
use_kl,
one_noise_step,
nll_in_loss,
mask_padding,
rescale_timesteps=False,
):
self.rescale_timesteps = rescale_timesteps
self.predict_xstart = predict_xstart
self.rescale_learned_sigmas = rescale_learned_sigmas
self.learn_sigmas = learn_sigmas
self.sigma_small = sigma_small
self.use_kl = use_kl
self.one_noise_step = one_noise_step
self.nll_in_loss = nll_in_loss
self.mask_padding = mask_padding
# Use float64 for accuracy.
betas = np.array(betas, dtype=np.float64) # shape [diffusion_steps]
self.betas = betas
assert len(betas.shape) == 1, "betas must be 1-D"
assert (betas > 0).all() and (betas <= 1).all()
self.num_timesteps = int(betas.shape[0])
alphas = 1.0 - betas
self.alphas_cumprod = np.cumprod(alphas, axis=0) # will approximate 0
self.alphas_cumprod_prev = np.append(
1.0, self.alphas_cumprod[:-1]
) # shifted one to the right
self.alphas_cumprod_next = np.append(
self.alphas_cumprod[1:], 0.0
) # shifted one to the left
assert self.alphas_cumprod_prev.shape == (self.num_timesteps,)
# calculations for diffusion q(x_t | x_{t-1}) and others
self.sqrt_alphas_cumprod = np.sqrt(self.alphas_cumprod)
self.sqrt_one_minus_alphas_cumprod = np.sqrt(1.0 - self.alphas_cumprod)
self.log_one_minus_alphas_cumprod = np.log(1.0 - self.alphas_cumprod)
self.sqrt_recip_alphas_cumprod = np.sqrt(1.0 / self.alphas_cumprod)
self.sqrt_recipm1_alphas_cumprod = np.sqrt(1.0 / self.alphas_cumprod - 1)
# calculations for posterior q(x_{t-1} | x_t, x_0)
self.posterior_variance = (
betas * (1.0 - self.alphas_cumprod_prev) / (1.0 - self.alphas_cumprod)
)
# log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain.
self.posterior_log_variance_clipped = np.log(
np.append(self.posterior_variance[1], self.posterior_variance[1:])
)
self.posterior_mean_coef1 = (
betas * np.sqrt(self.alphas_cumprod_prev) / (1.0 - self.alphas_cumprod)
)
self.posterior_mean_coef2 = (
(1.0 - self.alphas_cumprod_prev) * np.sqrt(alphas) / (1.0 - self.alphas_cumprod)
)
self.mapping_func = None # implement in train main()
self.add_mask_noise = False # TODO
def training_losses(self, model, *args, **kwargs):
self.model = model
return self.training_losses_seq2seq(model, *args, **kwargs)
def _predict_xstart_from_eps(self, x_t, t, eps):
assert x_t.shape == eps.shape
return (
_extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t
- _extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * eps
)
def _predict_eps_from_xstart(self, x_t, t, pred_xstart):
return (
_extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - pred_xstart
) / _extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
def _scale_timesteps(self, t):
if self.rescale_timesteps:
return t.float() * (1000.0 / self.num_timesteps)
return t
def q_mean_variance(self, x_start, t):
"""
Get the distribution q(x_t | x_0).
:param x_start: the [N x C x ...] tensor of noiseless inputs.
:param t: the number of diffusion steps (minus 1). Here, 0 means one step.
:return: A tuple (mean, variance, log_variance), all of x_start's shape.
"""
mean = _extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start
variance = _extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape)
log_variance = _extract_into_tensor(self.log_one_minus_alphas_cumprod, t, x_start.shape)
return mean, variance, log_variance
def q_sample(self, x_start, t, noise=None, mask=None):
"""
Diffuse the data for a given number of diffusion steps.
In other words, sample from q(x_t | x_0).
:param x_start: the initial data batch.
:param t: the number of diffusion steps (minus 1). Here, 0 means one step.
:param noise: if specified, the split-out normal noise.
:param mask: anchoring masked position
:return: A noisy version of x_start.
"""
if noise is None:
noise = th.randn_like(x_start)
assert noise.shape == x_start.shape
x_t = (
_extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start # mu * x_0
+ _extract_into_tensor(
self.sqrt_one_minus_alphas_cumprod, t, x_start.shape
) # sd * noise
* noise
)
if mask == None:
return x_t
else:
mask = th.broadcast_to(mask.unsqueeze(dim=-1), x_start.shape)
return th.where(mask == 0, x_start, x_t)
def q_posterior_mean_variance(self, x_start, x_t, t):
"""
Compute the mean and variance of the diffusion posterior:
q(x_{t-1} | x_t, x_0)
"""
assert x_start.shape == x_t.shape
posterior_mean = (
_extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start
+ _extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t
)
posterior_variance = _extract_into_tensor(self.posterior_variance, t, x_t.shape)
posterior_log_variance_clipped = _extract_into_tensor(
self.posterior_log_variance_clipped, t, x_t.shape
)
assert (
posterior_mean.shape[0]
== posterior_variance.shape[0]
== posterior_log_variance_clipped.shape[0]
== x_start.shape[0]
)
return posterior_mean, posterior_variance, posterior_log_variance_clipped
def p_mean_variance(
self,
model,
x,
sn_input_ids_emb,
pos_enc,
mask_sn_padding,
mask_transformer_att,
t,
clip_denoised=True,
denoised_fn=None,
model_kwargs=None,
subwords_list=None,
atten_vis=None,
atten_vis_fn=None,
atten_vis_path=None,
batch_idx=None,
rank=None,
atten_vis_sp=None,
):
"""
Apply the model to get p(x_{t-1} | x_t), as well as a prediction of
the initial x, x_0.
:param model: the model, which takes a signal and a batch of timesteps
as input.
:param x: the [N x C x ...] tensor at time t; the noised input where the embedded condition/sn was not noised
and the target/sp was replaced with Gaussian noise; at time step t
:param sn_input_ids_emb: the BERT embeddings
:param pos_enc: the positional embeddings
:param mask_sn_padding:
:param mask_transformer_att: the attention mask for the transformer
:param t: a 1-D Tensor of timesteps.
:param clip_denoised: if True, clip the denoised signal into [-1, 1].
:param denoised_fn: if not None, a function which applies to the x_start prediction before it is used to sample.
Applies before clip_denoised.
:param model_kwargs: if not None, a dict of extra keyword arguments to pass to the model. This can be used for conditioning.
:param subwords_list: list of list containing the subwortokens of each instance (for attention visualization)
:param atten_vis: bool: if True, attention is visualized (heatmaps)
:param atten_vis_fn: the attention visualization function
:param atten_vis_path: the path where to save the heatmaps to
:param batch_idx: index of the current batch
:paran rank: if parallel processing, the GPU index
:param atten_vis_sp: bool: save attention scores of the last timestep
:return: a dict with the following keys:
- 'mean': the model mean output.
- 'variance': the model variance output.
- 'log_variance': the log of 'variance'.
- 'pred_xstart': the prediction for x_0.
"""
if model_kwargs is None:
model_kwargs = {}
B, C = x.size(0), x.size(-1)
assert t.shape == (B,)
if not atten_vis and not atten_vis_sp:
model_output = model(
x=x,
ts=self._scale_timesteps(t),
sn_input_ids_emb=sn_input_ids_emb,
pos_enc=pos_enc,
attention_mask=mask_transformer_att,
**model_kwargs,
)
else:
# visualise the attention (heatmaps)
if atten_vis:
model_output, attention_scores = model(
x=x,
ts=self._scale_timesteps(t),
sn_input_ids_emb=sn_input_ids_emb,
pos_enc=pos_enc,
attention_mask=mask_transformer_att,
atten_vis=atten_vis,
**model_kwargs,
)
# visualise attention for last denoising step
if t[0] % 200 == 0:
atten_vis_fn(
attention_scores=attention_scores,
subwords_list=subwords_list,
batch_idx=batch_idx,
path_to_dir=atten_vis_path,
denoising_step=t[0].item(),
aggregate=True,
rank=rank,
)
if atten_vis_sp:
model_output, attention_scores = model(
x=x,
ts=self._scale_timesteps(t),
sn_input_ids_emb=sn_input_ids_emb,
pos_enc=pos_enc,
attention_mask=mask_transformer_att,
atten_vis=True,
**model_kwargs,
)
if t[0] == 0:
out_path_heatmaps_sp = os.path.join(atten_vis_path, "heatmaps_sps")
if not os.path.exists(out_path_heatmaps_sp):
os.makedirs(out_path_heatmaps_sp)
filename = f"att_scores_rank{rank}_batch{batch_idx}.pt"
path_to_file = os.path.join(out_path_heatmaps_sp, filename)
torch.save(attention_scores, path_to_file)
model_variance = np.append(self.posterior_variance[1], self.betas[1:])
model_log_variance = np.log(np.append(self.posterior_variance[1], self.betas[1:]))
model_variance = _extract_into_tensor(model_variance, t, x.shape)
model_log_variance = _extract_into_tensor(model_log_variance, t, x.shape)
# The denoised_fn is applied to x_start (the model output) before it is used for sampling
def process_xstart(x):
"""here x is the model output"""
if denoised_fn is not None:
# print(denoised_fn)
x = denoised_fn(x, t)
if clip_denoised:
return x.clamp(-1, 1)
return x
if self.predict_xstart:
# the denoised fn is applied to the model output
pred_xstart = process_xstart(model_output)
else:
### model is used to predict eps
pred_xstart = process_xstart(
self._predict_xstart_from_eps(x_t=x, t=t, eps=model_output)
)
# this is the mean of the posterior distribution q(x_{t-1} | x_t, x_0), estimated from x_t, which is the noised
# input, and pred_xstart, which is what the model predicted to be x_0 from the noised input x_noised/x_t
model_mean, _, _ = self.q_posterior_mean_variance(x_start=pred_xstart, x_t=x, t=t)
assert model_mean.shape == model_log_variance.shape == pred_xstart.shape == x.shape
return {
"mean": model_mean,
"variance": model_variance,
"log_variance": model_log_variance,
"pred_xstart": pred_xstart,
}
def p_sample(
self,
model,
x,
sn_input_ids_emb,
pos_enc,
mask_sn_padding,
mask_transformer_att,
t,
clip_denoised=True,
denoised_fn=None,
model_kwargs=None,
top_p=None,
mask=None,
x_start=None,
subwords_list=None,
atten_vis=None,
atten_vis_fn=None,
atten_vis_path=None,
batch_idx=None,
rank=None,
atten_vis_sp=None,
):
"""
Sample x_{t-1} from the model at the given timestep.
:param model: the model to sample from; the transformer model that learned the denoising
:param x: the current tensor at x_{t-1}.
:param sn_input_ids_emb: the BERT embeddings
:param pos_enc: the positional embeddings
:param mask_sn_padding:
:param mask_transformer_att: the attention mask for the transformer
:param t: the value of t, starting at 0 for the first diffusion step.
:param clip_denoised: if True, clip the x_start prediction to [-1, 1].
:param denoised_fn: if not None, a function which applies to the x_start prediction before it is used to sample.
:param model_kwargs: if not None, a dict of extra keyword arguments to pass to the model. This can be used for conditioning.
:param top_p:
:param mask: anchoring masked position to x_start
:param x_start:
:param subwords_list: list of list containing the subwortokens of each instance (for attention visualization)
:param atten_vis: bool: if True, attention is visualized (heatmaps)
:param atten_vis_fn: the attention visualization function
:param atten_vis_path: the path where to save the heatmaps to
:param batch_idx: index of the current batch
:paran rank: if parallel processing, the GPU index
:param atten_vis_sp: bool: save attention scores of the last timestep
:return: a dict containing the following keys:
- 'sample': a random sample from the model.
- 'pred_xstart': a prediction of x_0.
"""
out = self.p_mean_variance(
model=model,
x=x,
sn_input_ids_emb=sn_input_ids_emb,
pos_enc=pos_enc,
mask_sn_padding=mask_sn_padding,
mask_transformer_att=mask_transformer_att,
t=t,
clip_denoised=clip_denoised,
denoised_fn=denoised_fn,
model_kwargs=model_kwargs,
subwords_list=subwords_list,
atten_vis=atten_vis,
atten_vis_fn=atten_vis_fn,
atten_vis_path=atten_vis_path,
batch_idx=batch_idx,
rank=rank,
atten_vis_sp=atten_vis_sp,
)
if top_p is not None and top_p > 0:
# print('top_p sampling')
noise = th.randn_like(x)
replace_mask = th.abs(noise) > top_p
while replace_mask.any():
noise[replace_mask] = th.randn_like(noise[replace_mask])
replace_mask = th.abs(noise) > top_p
assert (th.abs(noise) <= top_p).all()
else:
noise = th.randn_like(x)
nonzero_mask = (
(t != 0).float().view(-1, *([1] * (len(x.shape) - 1)))
) # no noise when t == 0
sample = out["mean"] + nonzero_mask * th.exp(0.5 * out["log_variance"]) * noise
if mask == None:
pass
else:
# the original embedding for the sn, and the predicted sample for the sp
sample = th.where(mask == 0, x_start, sample)
return {
"sample": sample,
"pred_xstart": out["pred_xstart"],
"greedy_mean": out["mean"],
"out": out,
}
def p_sample_loop(
self,
model,
shape,
noise=None,
sn_input_ids_emb=None,
pos_enc=None,
mask_sn_padding=None,
mask_transformer_att=None,
clip_denoised=True,
denoised_fn=None,
model_kwargs=None,
device=None,
progress=False,
top_p=None,
clamp_step=None,
clamp_first=None,
mask=None,
x_start=None,
subwords_list=None,
atten_vis=None,
atten_vis_fn=None,
atten_vis_path=None,
batch_idx=None,
gap=1,
rank=None,
atten_vis_sp=None,
):
"""
Generate samples from the model.
:param model: the transformer model that was trained to learn the denoising
:param shape: the shape of the samples, (N, C, H, W).
:param noise: the Gaussian noise that should be denoised at inference (the replaced word ID emb)
:param sn_input_ids_emb: the BERT embeddings
:param pos_enc: the positional embeddings
:param mask_sn_padding:
:param mask_transformer_att: the attention mask for the transformer
:param clip_denoised: if True, clip x_start predictions to [-1, 1].
:param denoised_fn: if not None, a function which applies to the x_start prediction before it is used to sample.
:param model_kwargs: if not None, a dict of extra keyword arguments to pass to the model. This can be used for conditioning.
:param device: if specified, the device to create the samples on. If not specified, use a model parameter's device.
:param progress: if True, show a tqdm progress bar.
:param top_p:
:param clamp_step: in clamp_first mode, choose end clamp step, otherwise starting clamp step
:param clamp_first: bool, clamp_first mode
:param mask: anchoring masked position to x_start
:param x_start: the word ID embedding before replaced by noise
:param subwords_list: list of list containing the subwortokens of each instance (for attention visualization)
:param atten_vis: bool: if True, attention is visualized (heatmaps)
:param atten_vis_fn: the attention visualization function
:param atten_vis_path: the path where to save the heatmaps to
:param batch_idx: index of the current batch
:param gap:
:paran rank: if parallel processing, the GPU index
:param atten_vis_sp: bool: save attention scores of the last timestep
:return: a non-differentiable batch of samples.
"""
final = []
for sample in self.p_sample_loop_progressive(
model,
shape,
noise=noise,
sn_input_ids_emb=sn_input_ids_emb,
pos_enc=pos_enc,
mask_sn_padding=mask_sn_padding,
mask_transformer_att=mask_transformer_att,
clip_denoised=clip_denoised,
denoised_fn=denoised_fn,
model_kwargs=model_kwargs,
device=device,
progress=progress,
top_p=top_p,
clamp_step=clamp_step,
clamp_first=clamp_first,
mask=mask,
x_start=x_start,
subwords_list=subwords_list,
atten_vis=atten_vis,
atten_vis_fn=atten_vis_fn,
atten_vis_path=atten_vis_path,
batch_idx=batch_idx,
rank=rank,
atten_vis_sp=atten_vis_sp,
):
final.append(sample["sample"])
return final
def p_sample_loop_progressive(
self,
model,
shape,
noise=None,
sn_input_ids_emb=None,
pos_enc=None,
mask_sn_padding=None,
mask_transformer_att=None,
clip_denoised=True,
denoised_fn=None,
model_kwargs=None,
device=None,
progress=False,
top_p=None,
clamp_step=None,
clamp_first=None,
mask=None,
x_start=None,
subwords_list=None,
atten_vis=None,
atten_vis_fn=None,
atten_vis_path=None,
batch_idx=None,
rank=None,
atten_vis_sp=None,
):
"""
Generate samples from the model and yield intermediate samples from
each timestep of diffusion.
Arguments are the same as p_sample_loop().
Returns a generator over dicts, where each dict is the return value of
p_sample().
"""
if device is None:
device = next(model.parameters()).device
assert isinstance(shape, (tuple, list))
# noise/sample_x is the input that was noised: the concatenated sn-sp embedding where the sp was completely
# replaced with Gaussian noise from the standard normal distribution
if noise is not None:
sample_x = noise
else:
sample_x = th.randn(*shape, device=device)
# the number of diffusion steps in reverse order
indices = list(range(self.num_timesteps))[::-1]
if progress:
# Lazy import so that we don't depend on tqdm.
from tqdm.auto import tqdm
indices = tqdm(indices)
# denoising from the number of diffusion steps T to t=0
for i in indices: # from T to 0
t = th.tensor([i] * shape[0], device=device)
if not clamp_first:
if i > clamp_step:
denoised_fn_cur = None
else:
denoised_fn_cur = denoised_fn
else:
if i >= clamp_step:
denoised_fn_cur = denoised_fn
else:
denoised_fn_cur = None
with th.no_grad():
out = self.p_sample(
model=model,
x=sample_x,
sn_input_ids_emb=sn_input_ids_emb,
pos_enc=pos_enc,
mask_sn_padding=mask_sn_padding,
mask_transformer_att=mask_transformer_att,
t=t,
clip_denoised=clip_denoised,
denoised_fn=denoised_fn_cur,
model_kwargs=model_kwargs,
top_p=top_p,
mask=mask,
subwords_list=subwords_list,
x_start=x_start,
atten_vis=atten_vis,
atten_vis_fn=atten_vis_fn,
atten_vis_path=atten_vis_path,
batch_idx=batch_idx,
rank=rank,
atten_vis_sp=atten_vis_sp,
)
yield out
sample_x = out["sample"]
def _get_x_start(self, x_start_mean, std):
"""
Word embedding projection from {Emb(w)} to {x_0}
:param x_start_mean: word embedding
:return: x_0
"""
noise = th.randn_like(x_start_mean)
assert noise.shape == x_start_mean.shape
# print(x_start_mean.device, noise.device)
return x_start_mean + std * noise
def _token_discrete_loss(self, x_t, get_logits, input_ids, mask=None, truncate=False, t=None):
"""
the loss of -log p(w|z_0)
:param x_start_mean: word embedding
:return: x_0
"""
reshaped_x_t = x_t
logits = get_logits(reshaped_x_t) # shape [microbatch size, seq_len, vocabulary]
# print(logits.shape)
loss_fct = th.nn.CrossEntropyLoss(reduction="none")
decoder_nll = loss_fct(logits.view(-1, logits.size(-1)), input_ids.view(-1)).view(
input_ids.shape
)
if mask != None:
decoder_nll *= mask
# print(decoder_nll.shape)
if mask != None:
decoder_nll = decoder_nll.sum(dim=-1) / mask.sum(dim=-1)
else:
decoder_nll = decoder_nll.mean(dim=-1)
return decoder_nll
def _x0_helper(self, model_output, x, t):
if self.predict_xstart:
pred_xstart = model_output
pred_prev, _, _ = self.q_posterior_mean_variance(x_start=pred_xstart, x_t=x, t=t)
else: # predict eps
pred_xstart = self._predict_xstart_from_eps(x_t=x, t=t, eps=model_output)
pred_prev, _, _ = self.q_posterior_mean_variance(x_start=pred_xstart, x_t=x, t=t)
return {"pred_xprev": pred_prev, "pred_xstart": pred_xstart}
def training_losses_seq2seq(
self,
model, # the transformer model
t, # the number of noise adding steps for each instance in the microbatch
sn_sp_repr,
mask,
sn_input_ids,
indices_pos_enc,
mask_sn_padding,
mask_transformer_att,
noise=None,
):
"""
Compute training losses for a single timestep.
:param model: the transformer model
:param t: a batch of timestep indices.
:param sn_sp_repr: the word IDs of sn and sp
:param mask: masking the sn
:param sn_input_ids: the tokenizer input IDs
:param indices_pos_enc: the indices for pos enc
:param_mask_sn_padding:
:param mask_transformer_att: the transformer att
:param model_kwargs: if not None, a dict of extra keyword arguments to pass to the model. This can be used for conditioning.
:param noise: if specified, the specific Gaussian noise to try to remove.
:return: a dict with the key "loss" containing a tensor of shape [N].
Some mean or variance settings may also have other keys.
"""
microbatch_size, seq_len = sn_sp_repr.shape
# get the word ID embedding, BERT embedding, positional embedding
sn_sp_emb, pos_enc, sn_input_ids_emb = model.model.module.get_embeds(
sn_sp_repr=sn_sp_repr,
sn_input_ids=sn_input_ids,
indices_pos_enc=indices_pos_enc,
)
# get the standard deviation, shape [microbatch, args.seq_len, hidden_size=768]
std = _extract_into_tensor(
self.sqrt_one_minus_alphas_cumprod, th.tensor([0]).to(sn_sp_emb.device), sn_sp_emb.shape
)
# map sn_sp_emb to x_start, which is a one-step noised sn_sp_emb (in paper it's z_0)
if (
self.one_noise_step
): # this should always be true actually (without it performance is bad)
x_start = self._get_x_start(sn_sp_emb, std)
else:
x_start = sn_sp_emb
# sample noise in the same shape as our input
if noise is None:
noise = th.randn_like(x_start)
# get the noised sample x_t, which is still of shape [microbatch, args.seq_len, hidden_size=768]
# the condition/sn is not noised (hence the input mask)
# each instance in the microbatch receives a different amount of noise (t noising steps, as given in vector t)
x_t = self.q_sample(
x_start=x_start,
t=t,
noise=noise,
mask=mask,
)
terms = {}
target = x_start
# model_output is of shape [microbatch, args.seq_len, emb_dim=768]
model_output = model(
x=x_t,
ts=self._scale_timesteps(t),
sn_input_ids_emb=sn_input_ids_emb,
pos_enc=pos_enc,
attention_mask=mask_transformer_att,
)
assert model_output.shape == target.shape == x_start.shape
# Loss 1: Mean Squared Error (MSE) (L_{VLB})
terms["mse"] = mean_flat((target - model_output) ** 2)
model_out_x_start = self._x0_helper(model_output, x_t, t)["pred_xstart"]
t0_mask = (
t == 0
) # mask that says true for every instance where no noise was received, i.e. t=0
# MSE between the model output and the embedded input before the one noise step
t0_loss = mean_flat((sn_sp_emb - model_out_x_start) ** 2)
# update the MSE between the model output and the one-step noised input embeddings with the MSE between the
# model output and the embeddings before the one noise step wherever there was no noise received in the noising
# process (i.e., wherever t was 0)
terms["mse"] = th.where(t0_mask, t0_loss, terms["mse"])
# Loss 2: L_{round}
out_mean, _, _ = self.q_mean_variance(
x_start, th.LongTensor([self.num_timesteps - 1]).to(x_start.device)
)
tT_loss = mean_flat(out_mean**2)
# for the NLL losses, we need to convert the model output into logits
get_logits = model.model.module.get_logits
# Loss 3: L_{EMB}
# compute the NLL between the one-noised embeddings and the initial representation (word IDs)
# embedding regularisation
decoder_nll = self._token_discrete_loss(x_start, get_logits, sn_sp_repr)
# unused Loss
terms["nll"] = self._token_discrete_loss(model_output, get_logits, sn_sp_repr, mask=mask)
# combined loss
if self.nll_in_loss: # should be False; model performance drops if nll included
terms["loss"] = terms["mse"] + tT_loss + decoder_nll + terms["nll"]
else:
terms["loss"] = terms["mse"] + tT_loss + decoder_nll
return terms
def ddim_sample(
self,
model,
x,
t,
clip_denoised=True,
denoised_fn=None,
model_kwargs=None,
eta=0.0,
langevin_fn=None,
mask=None,
x_start=None,
):
"""
Sample x_{t-1} from the model using DDIM.
Same usage as p_sample().
"""
out = self.p_mean_variance(
model,
x,
t,
clip_denoised=clip_denoised,
denoised_fn=denoised_fn,
model_kwargs=model_kwargs,
)
# Usually our model outputs epsilon, but we re-derive it
# in case we used x_start or x_prev prediction.
eps = self._predict_eps_from_xstart(x, t, out["pred_xstart"])
alpha_bar = _extract_into_tensor(self.alphas_cumprod, t, x.shape)
alpha_bar_prev = _extract_into_tensor(self.alphas_cumprod_prev, t, x.shape)
sigma = (
eta
* th.sqrt((1 - alpha_bar_prev) / (1 - alpha_bar))
* th.sqrt(1 - alpha_bar / alpha_bar_prev)
)
# Equation 12.
noise = th.randn_like(x)
mean_pred = (
out["pred_xstart"] * th.sqrt(alpha_bar_prev)
+ th.sqrt(1 - alpha_bar_prev - sigma**2) * eps
)
nonzero_mask = (
(t != 0).float().view(-1, *([1] * (len(x.shape) - 1)))
) # no noise when t == 0
# print(sigma.mean())
sample = mean_pred + nonzero_mask * sigma * noise
if langevin_fn:
print(t.shape)
sample = langevin_fn(sample, mean_pred, sigma, self.alphas_cumprod_prev[t[0]], t, x)
if mask == None:
pass
else:
sample = th.where(mask == 0, x_start, sample)
return {"sample": sample, "pred_xstart": out["pred_xstart"]}
def ddim_reverse_sample(
self,
model,
x,
t,
clip_denoised=True,
denoised_fn=None,
model_kwargs=None,
eta=0.0,
):
"""
Sample x_{t+1} from the model using DDIM reverse ODE.
"""
assert eta == 0.0, "Reverse ODE only for deterministic path"
out = self.p_mean_variance(
model,
x,
t,
clip_denoised=clip_denoised,
denoised_fn=denoised_fn,
model_kwargs=model_kwargs,
)
# Usually our model outputs epsilon, but we re-derive it
# in case we used x_start or x_prev prediction.
eps = (
_extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x.shape) * x
- out["pred_xstart"]
) / _extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x.shape)
alpha_bar_next = _extract_into_tensor(self.alphas_cumprod_next, t, x.shape)
# Equation 12. reversed
mean_pred = out["pred_xstart"] * th.sqrt(alpha_bar_next) + th.sqrt(1 - alpha_bar_next) * eps
return {"sample": mean_pred, "pred_xstart": out["pred_xstart"]}
def ddim_sample_loop(
self,
model,
shape,
noise=None,
clip_denoised=True,
denoised_fn=None,
model_kwargs=None,
device=None,
progress=False,
top_p=None,
clamp_step=None,
clamp_first=None,
mask=None,
x_start=None,
gap=1,
):
"""
Generate samples from the model using DDIM.
:param gap: compute ddim sampling for each {gap} step
Same usage as p_sample_loop().
"""
final = []
for sample in self.ddim_sample_loop_progressive(
model,
shape,
noise=noise,
clip_denoised=clip_denoised,
denoised_fn=denoised_fn,
model_kwargs=model_kwargs,
device=device,
progress=progress,
mask=mask,
x_start=x_start,
gap=gap,
):
final.append(sample["sample"])
return final
def ddim_sample_loop_progressive(
self,
model,
shape,
noise=None,
clip_denoised=True,
denoised_fn=None,
model_kwargs=None,
device=None,
progress=False,
eta=0.0,
langevin_fn=None,
mask=None,
x_start=None,
gap=1,
):
"""
Use DDIM to sample from the model and yield intermediate samples from
each timestep of DDIM.
Same usage as p_sample_loop_progressive().
"""
if device is None:
device = next(model.parameters()).device
assert isinstance(shape, (tuple, list))
if noise is not None:
sample_x = noise
else:
sample_x = th.randn(*shape, device=device)
indices = list(range(self.num_timesteps))[::-1][::gap]
if progress:
# Lazy import so that we don't depend on tqdm.
from tqdm.auto import tqdm
indices = tqdm(indices)
for i in indices:
t = th.tensor([i] * shape[0], device=device)
with th.no_grad():
out = self.ddim_sample(
model,
sample_x,
t,
clip_denoised=clip_denoised,
denoised_fn=denoised_fn,
model_kwargs=model_kwargs,
mask=mask,
x_start=x_start,
)
yield out
sample_x = out["sample"]
def _extract_into_tensor(arr, timesteps, broadcast_shape):
"""
Extract values from a 1-D numpy array for a batch of indices.
:param arr: the 1-D numpy array.
:param timesteps: a tensor of indices into the array to extract.
:param broadcast_shape: a larger shape of K dimensions with the batch
dimension equal to the length of timesteps.
:return: a tensor of shape [batch_size, 1, ...] where the shape has K dims.
"""
res = th.from_numpy(arr).to(device=timesteps.device)[timesteps].float()
while len(res.shape) < len(broadcast_shape):
res = res[..., None]
return res.expand(broadcast_shape)
def space_timesteps(num_timesteps, section_counts):
"""
Create a list of timesteps to use from an original diffusion process,
given the number of timesteps we want to take from equally-sized portions
of the original process.
For example, if there's 300 timesteps and the section counts are [10,15,20]
then the first 100 timesteps are strided to be 10 timesteps, the second 100
are strided to be 15 timesteps, and the final 100 are strided to be 20.
If the stride is a string starting with "ddim", then the fixed striding
from the DDIM paper is used, and only one section is allowed.
:param num_timesteps: the number of diffusion steps in the original
process to divide up.
:param section_counts: either a list of numbers, or a string containing
comma-separated numbers, indicating the step count
per section. As a special case, use "ddimN" where N
is a number of steps to use the striding from the
DDIM paper.
:return: a set of diffusion steps from the original process to use.
"""
if isinstance(section_counts, str):
if section_counts.startswith("ddim"):
desired_count = int(section_counts[len("ddim") :])
for i in range(1, num_timesteps):
if len(range(0, num_timesteps, i)) == desired_count:
return set(range(0, num_timesteps, i))
raise ValueError(f"cannot create exactly {num_timesteps} steps with an integer stride")
section_counts = [int(x) for x in section_counts.split(",")]
size_per = num_timesteps // len(section_counts)
extra = num_timesteps % len(section_counts)
start_idx = 0
all_steps = []
for i, section_count in enumerate(section_counts):
size = size_per + (1 if i < extra else 0)
if size < section_count:
raise ValueError(f"cannot divide section of {size} steps into {section_count}")
if section_count <= 1:
frac_stride = 1
else:
frac_stride = (size - 1) / (section_count - 1)
cur_idx = 0.0
taken_steps = []
for _ in range(section_count):
taken_steps.append(start_idx + round(cur_idx))
cur_idx += frac_stride
all_steps += taken_steps
start_idx += size
return set(all_steps)
class SpacedDiffusion(GaussianDiffusion):
"""
A diffusion process which can skip steps in a base diffusion process.
:param use_timesteps: a collection (sequence or set) of timesteps from the
original diffusion process to retain.
:param kwargs: the kwargs to create the base diffusion process.
"""
def __init__(self, use_timesteps, **kwargs):
self.use_timesteps = set(use_timesteps)
self.timestep_map = []
self.original_num_steps = len(kwargs["betas"])
# print(kwargs.keys())
base_diffusion = GaussianDiffusion(**kwargs) # pylint: disable=missing-kwoa
last_alpha_cumprod = 1.0
new_betas = []
for i, alpha_cumprod in enumerate(base_diffusion.alphas_cumprod):
if i in self.use_timesteps:
new_betas.append(1 - alpha_cumprod / last_alpha_cumprod)
last_alpha_cumprod = alpha_cumprod
self.timestep_map.append(i)
kwargs["betas"] = np.array(new_betas)
super().__init__(**kwargs)
def p_mean_variance(self, model, *args, **kwargs): # pylint: disable=signature-differs
# print('called p_mean_var')
return super().p_mean_variance(self._wrap_model(model), *args, **kwargs)
def training_losses(self, model, *args, **kwargs): # pylint: disable=signature-differs
# print('called training_losses')
return super().training_losses(self._wrap_model(model), *args, **kwargs)
def _wrap_model(self, model):
if isinstance(model, _WrappedModel):
return model
return _WrappedModel(
model, self.timestep_map, self.rescale_timesteps, self.original_num_steps
)
def _scale_timesteps(self, t):
# Scaling is done by the wrapped model.
return t
class _WrappedModel:
def __init__(self, model, timestep_map, rescale_timesteps, original_num_steps):
self.model = model
self.timestep_map = timestep_map
self.rescale_timesteps = rescale_timesteps
self.original_num_steps = original_num_steps
def __call__(self, x, ts, **kwargs):
# print(ts)
map_tensor = th.tensor(self.timestep_map, device=ts.device, dtype=ts.dtype)
new_ts = map_tensor[ts]
# print(new_ts)
if self.rescale_timesteps:
new_ts = new_ts.float() * (1000.0 / self.original_num_steps)
# temp = self.model(x, new_ts, **kwargs)
# print(temp.shape)
# return temp
# print(new_ts)
return self.model(x, new_ts, **kwargs)
|