quizz_machine.py

   1 #!/usr/bin/env python
   2
   3 # Any copyright is dedicated to the Public Domain.
   4 # https://creativecommons.org/publicdomain/zero/1.0/
   5
   6 # Written by Francois Fleuret <francois@fleuret.org>
   7
   8 import math, os, tqdm, warnings
   9
  10 import torch, torchvision
  11
  12 from torch import nn
  13 from torch.nn import functional as F
  14
  15 import mygpt
  16 from mygpt import BracketedSequence
  17
  18 ######################################################################
  19
  20 # ar_mask is a tensor with 0s and 1s, of same shape as input, with
  21 # 1s where tokens should be generated. The others are kept
  22 # unchanged.
  23
  24
  25 def one_batch_masked_inplace_autoregression(
  26     model,
  27     input,
  28     ar_mask,
  29     seq_logproba,
  30     temperature=1.0,
  31     deterministic_synthesis=False,
  32 ):
  33     to_generate = (ar_mask.sum(0) > 0).nonzero()
  34
  35     if to_generate.min() > 0:
  36         model(
  37             BracketedSequence(input, 0, to_generate.min())
  38         )  # Needed to initialize the model's cache
  39     for s in range(to_generate.min(), to_generate.max() + 1):
  40         output = model(BracketedSequence(input, s, 1)).x
  41
  42         logits = output[:, s]
  43
  44         logits = (logits / temperature).log_softmax(dim=-1)
  45
  46         if deterministic_synthesis:
  47             t_next = logits.argmax(-1)
  48         else:
  49             dist = torch.distributions.categorical.Categorical(logits=logits)
  50             t_next = dist.sample()
  51
  52         all_n = torch.arange(t_next.size(0))
  53         seq_logproba += logits[all_n, t_next].sum(dim=-1)
  54
  55         input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
  56
  57
  58 def masked_inplace_autoregression(
  59     model,
  60     batch_size,
  61     input,
  62     ar_mask,
  63     seq_logproba,
  64     temperature,
  65     deterministic_synthesis,
  66     forbidden_tokens=None,
  67     logit_biases=None,
  68     progress_bar_desc=None,
  69     device=torch.device("cpu"),
  70 ):
  71     assert input.size() == ar_mask.size()
  72
  73     batches = zip(
  74         input.split(batch_size),
  75         ar_mask.split(batch_size),
  76         seq_logproba.split(batch_size),
  77     )
  78
  79     if progress_bar_desc is not None:
  80         batches = tqdm.tqdm(
  81             batches,
  82             dynamic_ncols=True,
  83             desc=progress_bar_desc,
  84             total=(input.size(0) + batch_size - 1) // batch_size,
  85         )
  86
  87     with torch.autograd.no_grad():
  88         t = model.training
  89         model.eval()
  90
  91         for input, ar_mask, seq_logproba in batches:
  92             one_batch_masked_inplace_autoregression(
  93                 model=model,
  94                 input=input,
  95                 ar_mask=ar_mask,
  96                 seq_logproba=seq_logproba,
  97                 temperature=temperature,
  98                 deterministic_synthesis=deterministic_synthesis,
  99             )
 100
 101         model.train(t)
 102
 103
 104 ######################################################################
 105
 106
 107 class QuizzMachine:
 108     def indices_forward_and_backward(self, quizzes):
 109         i_forward = quizzes[:, 0] == self.token_forward
 110         j_forward = quizzes[:, 1 + self.prompt_len] == self.token_forward
 111         i_backward = quizzes[:, 0] == self.token_backward
 112         j_backward = quizzes[:, 1 + self.answer_len] == self.token_backward
 113         assert torch.logical_or(
 114             torch.logical_and(i_forward, j_forward),
 115             torch.logical_and(i_backward, j_backward),
 116         ).all()
 117         return i_forward, i_backward
 118
 119     def reverse_time(self, quizzes):
 120         i_forward, i_backward = self.indices_forward_and_backward(quizzes)
 121
 122         forward_to_backward = torch.cat(
 123             [
 124                 quizzes[:, 0:1],
 125                 quizzes[:, 2 + self.prompt_len :],
 126                 quizzes[:, 1 + self.prompt_len : 2 + self.prompt_len],
 127                 quizzes[:, 1 : 1 + self.prompt_len],
 128             ],
 129             dim=1,
 130         )
 131         forward_to_backward[:, 0] = self.token_backward
 132         forward_to_backward[:, 1 + self.answer_len] = self.token_backward
 133
 134         backward_to_forward = torch.cat(
 135             [
 136                 quizzes[:, 0:1],
 137                 quizzes[:, 2 + self.answer_len :],
 138                 quizzes[:, 1 + self.answer_len : 2 + self.answer_len],
 139                 quizzes[:, 1 : 1 + self.answer_len],
 140             ],
 141             dim=1,
 142         )
 143
 144         backward_to_forward[:, 0] = self.token_forward
 145         backward_to_forward[:, 1 + self.prompt_len] = self.token_forward
 146
 147         m = i_forward.long()[:, None]
 148
 149         return m * forward_to_backward + (1 - m) * backward_to_forward
 150
 151     def make_ar_mask(self, quizzes, first=False):
 152         i_forward, i_backward = self.indices_forward_and_backward(quizzes)
 153
 154         t = torch.arange(quizzes.size(1), device=quizzes.device)
 155
 156         if first:
 157             m_forward = (t >= 1).long() * (t < 1 + self.prompt_len).long()
 158             m_backward = (t >= 1).long() * (t < 1 + self.answer_len).long()
 159         else:
 160             m_forward = (t >= 2 + self.prompt_len).long()
 161             m_backward = (t >= 2 + self.answer_len).long()
 162
 163         m = i_forward.long()[:, None]
 164
 165         return m * m_forward + (1 - m) * m_backward
 166
 167     def generate_token_sequences(self, nb):
 168         prompts, answers = self.problem.generate_prompts_and_answers(nb)
 169
 170         if self.prompt_len is None:
 171             self.prompt_len = prompts.size(1)
 172
 173         if self.answer_len is None:
 174             self.answer_len = answers.size(1)
 175
 176         assert prompts.size(1) == self.prompt_len and answers.size(1) == self.answer_len
 177
 178         result = []
 179
 180         for prompt, answer in zip(prompts, answers):
 181             if torch.rand(1) < 0.5:
 182                 a = [
 183                     torch.tensor([self.token_forward]),
 184                     prompt,
 185                     torch.tensor([self.token_forward]),
 186                     answer,
 187                 ]
 188             else:
 189                 a = [
 190                     torch.tensor([self.token_backward]),
 191                     answer,
 192                     torch.tensor([self.token_backward]),
 193                     prompt,
 194                 ]
 195
 196             result.append(torch.cat(a, dim=0)[None, :])
 197
 198         return torch.cat(result, dim=0)
 199
 200     def __init__(
 201         self,
 202         problem,
 203         nb_train_samples,
 204         nb_test_samples,
 205         back_accuracy,
 206         batch_size,
 207         result_dir,
 208         logger,
 209         device=torch.device("cpu"),
 210     ):
 211         super().__init__()
 212
 213         v = problem.nb_token_values()
 214         self.token_forward = v
 215         self.token_backward = v + 1
 216         self.nb_token_values = v + 2
 217
 218         self.problem = problem
 219         self.back_accuracy = back_accuracy
 220         self.batch_size = batch_size
 221         self.device = device
 222         self.logger = logger
 223         self.prompt_len = None
 224         self.answer_len = None
 225
 226         self.train_w_quizzes = self.generate_token_sequences(nb_train_samples).to(
 227             device
 228         )
 229
 230         self.test_w_quizzes = self.generate_token_sequences(nb_test_samples).to(device)
 231
 232         self.train_c_quizzes = []
 233         self.test_c_quizzes = []
 234
 235         if result_dir is not None:
 236             self.save_quizzes(
 237                 result_dir, "culture_w_quizzes", self.train_w_quizzes[:72]
 238             )
 239
 240             # toto = self.reverse_time(self.train_w_quizzes[:72])
 241             # self.save_quizzes(result_dir, "toto", toto)
 242             # exit(0)
 243
 244     def save_quizzes(self, result_dir, filename_prefix, quizzes, prediction=False):
 245         forward = quizzes[quizzes[:, 0] == self.token_forward]
 246         ib = quizzes[:, 0] == self.token_backward
 247         backward = quizzes[ib]
 248         assert forward.size(0) + backward.size(0) == quizzes.size(0)
 249         quizzes[ib] = self.reverse_time(quizzes[ib])
 250
 251         if prediction:
 252             predicted_prompts = ib
 253             predicted_answers = torch.logical_not(ib)
 254         else:
 255             predicted_prompts = None
 256             predicted_answers = None
 257
 258         self.problem.save_quizzes(
 259             result_dir,
 260             filename_prefix,
 261             quizzes[:, 1 : 1 + self.prompt_len],
 262             quizzes[:, 2 + self.prompt_len :],
 263             predicted_prompts,
 264             predicted_answers,
 265         )
 266
 267     def batches(self, split="train", desc=None):
 268         assert split in {"train", "test"}
 269         if split == "train":
 270             w_quizzes = self.train_w_quizzes
 271             c_quizzes = self.train_c_quizzes
 272         else:
 273             w_quizzes = self.test_w_quizzes
 274             c_quizzes = self.test_c_quizzes
 275
 276         if len(c_quizzes) > 0:
 277             c_quizzes = torch.cat(c_quizzes, dim=0)
 278             if c_quizzes.size(0) > w_quizzes.size(0) // 2:
 279                 i = torch.randperm(c_quizzes.size(0))[: w_quizzes.size(0) // 2]
 280                 c_quizzes = c_quizzes[i]
 281
 282             i = torch.randperm(w_quizzes.size(0))[
 283                 : w_quizzes.size(0) - c_quizzes.size(0)
 284             ]
 285             w_quizzes = w_quizzes[i]
 286
 287             self.nb_batch_w_quizzes = w_quizzes.size(0)
 288             self.nb_batch_c_quizzes = c_quizzes.size(0)
 289
 290             input = torch.cat([w_quizzes, c_quizzes], dim=0)
 291         else:
 292             input = w_quizzes
 293             self.nb_batch_w_quizzes = w_quizzes.size(0)
 294             self.nb_batch_c_quizzes = 0
 295
 296         # Shuffle
 297         input = input[torch.randperm(input.size(0))]
 298
 299         if desc is None:
 300             desc = f"epoch-{split}"
 301         for batch in tqdm.tqdm(
 302             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 303         ):
 304             yield batch
 305
 306     def vocabulary_size(self):
 307         return self.nb_token_values
 308
 309     def produce_results(
 310         self, n_epoch, model, result_dir, deterministic_synthesis, nmax=1000
 311     ):
 312         def compute_accuracy(input):
 313             ar_mask = self.make_ar_mask(input)
 314             result = input.clone() * (1 - ar_mask)
 315             seq_logproba = torch.empty(input.size(0), device=self.device)
 316
 317             masked_inplace_autoregression(
 318                 model=model,
 319                 batch_size=self.batch_size,
 320                 input=result,
 321                 ar_mask=ar_mask,
 322                 seq_logproba=seq_logproba,
 323                 temperature=1.0,
 324                 deterministic_synthesis=deterministic_synthesis,
 325                 progress_bar_desc=None,
 326                 device=self.device,
 327             )
 328
 329             if self.back_accuracy:
 330                 n_forward = input[:, 0] == self.token_forward
 331                 nb_total = input[n_forward].size(0)
 332                 nb_correct = (
 333                     (input[n_forward] == result[n_forward])
 334                     .long()
 335                     .min(dim=1)
 336                     .values.sum()
 337                 )
 338
 339                 n_backward = input[:, 0] == self.token_backward
 340                 back_input = self.reverse_time(result[n_backward])
 341                 if back_input.size(0) > 0:
 342                     back_input[:, 2 + self.prompt_len :] = input[
 343                         n_backward, 2 + self.prompt_len :
 344                     ]
 345                     back_nb_total, back_nb_correct = compute_accuracy(back_input)
 346                     nb_total += back_nb_total
 347                     nb_correct += back_nb_correct
 348             else:
 349                 nb_total = input.size(0)
 350                 nb_correct = (input == result).long().min(dim=1).values.sum()
 351
 352             return nb_total, nb_correct
 353
 354         train_nb_total, train_nb_correct = compute_accuracy(self.train_w_quizzes[:nmax])
 355
 356         self.logger(
 357             f"accuracy_train {n_epoch} nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 358         )
 359
 360         test_nb_total, test_nb_correct = compute_accuracy(self.test_w_quizzes[:nmax])
 361
 362         self.logger(
 363             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 364         )
 365
 366         main_test_accuracy = test_nb_correct / test_nb_total
 367         self.logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
 368
 369         ##############################
 370
 371         input = self.test_w_quizzes[:96]
 372         ar_mask = self.make_ar_mask(input)
 373         result = input.clone() * (1 - ar_mask)
 374         seq_logproba = torch.empty(input.size(0), device=self.device)
 375
 376         masked_inplace_autoregression(
 377             model=model,
 378             batch_size=self.batch_size,
 379             input=result,
 380             ar_mask=ar_mask,
 381             seq_logproba=seq_logproba,
 382             temperature=1.0,
 383             deterministic_synthesis=deterministic_synthesis,
 384             progress_bar_desc=None,
 385             device=self.device,
 386         )
 387
 388         self.save_quizzes(
 389             result_dir,
 390             f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
 391             quizzes=result[:72],
 392             prediction=True,
 393         )
 394
 395         return main_test_accuracy
 396
 397     def renew_w_quizzes(self, nb, for_train=True):
 398         input = self.train_w_quizzes if for_train else self.test_w_quizzes
 399         nb = min(nb, input.size(0))
 400         input[:-nb] = input[nb:].clone()
 401         input[-nb:] = self.generate_token_sequences(nb).to(self.device)
 402
 403     def store_c_quizzes(self, new_c_quizzes, for_train=True):
 404         if for_train:
 405             self.train_c_quizzes.append(new_c_quizzes)
 406         else:
 407             self.test_c_quizzes.append(new_c_quizzes)
 408
 409     def compute_correctness(
 410         self,
 411         c_quizzes,
 412         models_for_validation,
 413         bidirectional_validation=False,
 414         deterministic_validation=True,
 415     ):
 416         if bidirectional_validation:
 417             backward_c_quizzes = self.forward_to_backward(c_quizzes)
 418
 419         seq_logproba = torch.zeros(
 420             c_quizzes.size(0),
 421             max([m.id for m in models_for_validation]) + 1,
 422             device=self.device,
 423         )
 424
 425         nb_correct = 0
 426
 427         for model in models_for_validation:
 428             result = c_quizzes.clone()
 429
 430             seq_logproba[...] = 0.0
 431
 432             ar_mask = self.make_ar_mask(result)
 433
 434             masked_inplace_autoregression(
 435                 model=model,
 436                 batch_size=self.batch_size,
 437                 input=result,
 438                 ar_mask=ar_mask,
 439                 seq_logproba=seq_logproba[:, model.id],
 440                 temperature=1.0,
 441                 deterministic_synthesis=deterministic_validation,
 442                 # progress_bar_desc="solving c_quizzes",
 443                 device=self.device,
 444             )
 445
 446             correct = (c_quizzes == result).long().min(dim=-1).values
 447
 448             if bidirectional_validation:
 449                 backward_result = backward_c_quizzes.clone()
 450
 451                 ar_mask = self.make_ar_mask(backward_result)
 452
 453                 masked_inplace_autoregression(
 454                     model=model,
 455                     batch_size=self.batch_size,
 456                     input=backward_result,
 457                     ar_mask=ar_mask,
 458                     seq_logproba=seq_logproba[:, model.id],
 459                     temperature=1.0,
 460                     deterministic_synthesis=deterministic_validation,
 461                     # progress_bar_desc="solving backward c_quizzes",
 462                     device=self.device,
 463                 )
 464
 465                 backward_correct = (
 466                     (backward_c_quizzes == backward_result).long().min(dim=-1).values
 467                 )
 468
 469                 correct *= backward_correct
 470
 471             # endif
 472
 473             nb_correct += correct
 474
 475         return nb_correct, seq_logproba
 476
 477     ###############################################################
 478
 479     def generate_quizzes(self, nb, model_for_generation, temperature=1.0):
 480         c_quizzes = torch.empty(
 481             nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
 482         )
 483
 484         seq_logproba = torch.zeros(nb, device=self.device)
 485
 486         # First, we generate the answer at high temperature
 487
 488         c_quizzes[:, 0] = self.token_backward
 489         c_quizzes[:, 1 + self.answer_len] = self.token_backward
 490
 491         masked_inplace_autoregression(
 492             model=model_for_generation,
 493             batch_size=self.batch_size,
 494             input=c_quizzes,
 495             ar_mask=self.make_ar_mask(c_quizzes, first=True),
 496             seq_logproba=seq_logproba,
 497             temperature=temperature,
 498             deterministic_synthesis=False,
 499             device=self.device,
 500         )
 501
 502         # Then, we generate the prompt at low temperature
 503
 504         masked_inplace_autoregression(
 505             model=model_for_generation,
 506             batch_size=self.batch_size,
 507             input=c_quizzes,
 508             ar_mask=self.make_ar_mask(c_quizzes),
 509             seq_logproba=seq_logproba,
 510             temperature=1 / temperature,
 511             deterministic_synthesis=False,
 512             device=self.device,
 513         )
 514
 515         # Then we return the quizz, and re-generate the response, now
 516         # at low temperature
 517
 518         c_quizzes = self.reverse_time(c_quizzes)
 519
 520         masked_inplace_autoregression(
 521             model=model_for_generation,
 522             batch_size=self.batch_size,
 523             input=c_quizzes,
 524             ar_mask=self.make_ar_mask(c_quizzes),
 525             seq_logproba=seq_logproba,
 526             temperature=1 / temperature,
 527             deterministic_synthesis=False,
 528             device=self.device,
 529         )
 530
 531         return c_quizzes