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 : 2 + self.prompt_len + self.answer_len],
 126                 quizzes[:, 1 + self.prompt_len : 1 + self.prompt_len + 1],
 127                 quizzes[:, 1 : 1 + self.prompt_len],
 128             ],
 129             dim=1,
 130         )
 131
 132         forward_to_backward[:, 0] = self.token_backward
 133         forward_to_backward[:, 1 + self.answer_len] = self.token_backward
 134
 135         backward_to_forward = torch.cat(
 136             [
 137                 quizzes[:, 0:1],
 138                 quizzes[:, 2 + self.answer_len :],
 139                 quizzes[:, 1 + self.answer_len : 2 + self.answer_len],
 140                 quizzes[:, 1 : 1 + self.answer_len],
 141             ],
 142             dim=1,
 143         )
 144
 145         backward_to_forward[:, 0] = self.token_forward
 146         backward_to_forward[:, 1 + self.prompt_len] = self.token_forward
 147
 148         m = i_forward.long()[:, None]
 149
 150         return m * forward_to_backward + (1 - m) * backward_to_forward
 151
 152     def make_ar_mask(self, quizzes, first=False):
 153         i_forward, i_backward = self.indices_forward_and_backward(quizzes)
 154
 155         t = torch.arange(quizzes.size(1), device=quizzes.device)
 156
 157         if first:
 158             m_forward = (t >= 1).long() * (t < 1 + self.prompt_len).long()
 159             m_backward = (t >= 1).long() * (t < 1 + self.answer_len).long()
 160         else:
 161             m_forward = (t >= 2 + self.prompt_len).long()
 162             m_backward = (t >= 2 + self.answer_len).long()
 163
 164         m = i_forward.long()[:, None]
 165
 166         return m * m_forward + (1 - m) * m_backward
 167
 168     def generate_token_sequences(self, nb):
 169         prompts, answers = self.problem.generate_prompts_and_answers(nb)
 170
 171         if self.prompt_len is None:
 172             self.prompt_len = prompts.size(1)
 173
 174         if self.answer_len is None:
 175             self.answer_len = answers.size(1)
 176
 177         assert prompts.size(1) == self.prompt_len and answers.size(1) == self.answer_len
 178
 179         result = []
 180
 181         for prompt, answer in zip(prompts, answers):
 182             if torch.rand(1) < 0.5:
 183                 a = [
 184                     torch.tensor([self.token_forward]),
 185                     prompt,
 186                     torch.tensor([self.token_forward]),
 187                     answer,
 188                 ]
 189             else:
 190                 a = [
 191                     torch.tensor([self.token_backward]),
 192                     answer,
 193                     torch.tensor([self.token_backward]),
 194                     prompt,
 195                 ]
 196
 197             result.append(torch.cat(a, dim=0)[None, :])
 198
 199         return torch.cat(result, dim=0)
 200
 201     def __init__(
 202         self,
 203         problem,
 204         nb_train_samples,
 205         nb_test_samples,
 206         back_accuracy,
 207         batch_size,
 208         result_dir,
 209         logger,
 210         device=torch.device("cpu"),
 211     ):
 212         super().__init__()
 213
 214         v = problem.nb_token_values()
 215         self.token_forward = v
 216         self.token_backward = v + 1
 217         self.nb_token_values = v + 2
 218
 219         self.problem = problem
 220         self.back_accuracy = back_accuracy
 221         self.batch_size = batch_size
 222         self.device = device
 223         self.logger = logger
 224         self.prompt_len = None
 225         self.answer_len = None
 226
 227         self.train_w_quizzes = self.generate_token_sequences(nb_train_samples).to(
 228             device
 229         )
 230
 231         self.test_w_quizzes = self.generate_token_sequences(nb_test_samples).to(device)
 232
 233         self.train_c_quizzes = []
 234         self.test_c_quizzes = []
 235
 236         if result_dir is not None:
 237             self.save_quizzes(
 238                 result_dir,
 239                 "culture_w_quizzes",
 240                 self.train_w_quizzes[:72],
 241                 show_to_be_predicted=True,
 242             )
 243
 244     def save_quizzes(
 245         self,
 246         result_dir,
 247         filename_prefix,
 248         quizzes,
 249         show_to_be_predicted=False,
 250         mistakes=None,
 251     ):
 252         quizzes = quizzes.clone()
 253         forward = quizzes[quizzes[:, 0] == self.token_forward]
 254         ib = quizzes[:, 0] == self.token_backward
 255         backward = quizzes[ib]
 256         assert forward.size(0) + backward.size(0) == quizzes.size(0)
 257         quizzes[ib] = self.reverse_time(quizzes[ib])
 258
 259         if show_to_be_predicted:
 260             predicted_prompts = ib.long()
 261             predicted_answers = 1 - predicted_prompts
 262             if mistakes is not None:
 263                 # 0/-1/+1 ~ not-to-predict / predicted wrong / predicted correct
 264                 predicted_prompts *= mistakes
 265                 predicted_answers *= mistakes
 266             else:
 267                 # 0/2 ~ not-to-predict / to predict
 268                 predicted_prompts *= 2
 269                 predicted_answers *= 2
 270         else:
 271             predicted_prompts = None
 272             predicted_answers = None
 273
 274         self.problem.save_quizzes(
 275             result_dir,
 276             filename_prefix,
 277             quizzes[:, 1 : 1 + self.prompt_len],
 278             quizzes[:, 2 + self.prompt_len :],
 279             predicted_prompts,
 280             predicted_answers,
 281         )
 282
 283     def batches(self, split="train", desc=None):
 284         assert split in {"train", "test"}
 285         if split == "train":
 286             w_quizzes = self.train_w_quizzes
 287             c_quizzes = self.train_c_quizzes
 288         else:
 289             w_quizzes = self.test_w_quizzes
 290             c_quizzes = self.test_c_quizzes
 291
 292         if len(c_quizzes) > 0:
 293             c_quizzes = torch.cat(c_quizzes, dim=0)
 294             if c_quizzes.size(0) > w_quizzes.size(0) // 2:
 295                 i = torch.randperm(c_quizzes.size(0))[: w_quizzes.size(0) // 2]
 296                 c_quizzes = c_quizzes[i]
 297
 298             i = torch.randperm(w_quizzes.size(0))[
 299                 : w_quizzes.size(0) - c_quizzes.size(0)
 300             ]
 301             w_quizzes = w_quizzes[i]
 302
 303             self.nb_batch_w_quizzes = w_quizzes.size(0)
 304             self.nb_batch_c_quizzes = c_quizzes.size(0)
 305
 306             input = torch.cat([w_quizzes, c_quizzes], dim=0)
 307         else:
 308             input = w_quizzes
 309             self.nb_batch_w_quizzes = w_quizzes.size(0)
 310             self.nb_batch_c_quizzes = 0
 311
 312         # Shuffle
 313         input = input[torch.randperm(input.size(0))]
 314
 315         if desc is None:
 316             desc = f"epoch-{split}"
 317         for batch in tqdm.tqdm(
 318             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 319         ):
 320             yield batch
 321
 322     def vocabulary_size(self):
 323         return self.nb_token_values
 324
 325     def produce_results(
 326         self, n_epoch, model, result_dir, deterministic_synthesis, nmax=1000
 327     ):
 328         def compute_accuracy(input):
 329             ar_mask = self.make_ar_mask(input)
 330             result = input.clone() * (1 - ar_mask)
 331             seq_logproba = torch.empty(input.size(0), device=self.device)
 332
 333             masked_inplace_autoregression(
 334                 model=model,
 335                 batch_size=self.batch_size,
 336                 input=result,
 337                 ar_mask=ar_mask,
 338                 seq_logproba=seq_logproba,
 339                 temperature=1.0,
 340                 deterministic_synthesis=deterministic_synthesis,
 341                 progress_bar_desc=None,
 342                 device=self.device,
 343             )
 344
 345             if self.back_accuracy:
 346                 # If back_accuracy is True, we compute the accuracy on
 347                 # the backward quizzes not by counting how many time
 348                 # the real prompt A is equal to the reconstructed
 349                 # prompt A*, but how many time the answers B* computed
 350                 # from A* is equal to the correct answer. So we look
 351                 # for the accuracy of A->B*=B for the forward, but for
 352                 # the backward we look at B->A*->B*=B instead of B->A*=A
 353
 354                 n_forward = input[:, 0] == self.token_forward
 355                 nb_total = input[n_forward].size(0)
 356                 nb_correct = (
 357                     (input[n_forward] == result[n_forward])
 358                     .long()
 359                     .min(dim=1)
 360                     .values.sum()
 361                     .item()
 362                 )
 363
 364                 n_backward = input[:, 0] == self.token_backward
 365                 back_input = self.reverse_time(result[n_backward])
 366
 367                 if back_input.size(0) > 0:
 368                     back_input[:, 2 + self.prompt_len :] = input[
 369                         n_backward, 1 : 1 + self.answer_len
 370                     ]
 371                     back_nb_total, back_nb_correct = compute_accuracy(back_input)
 372
 373                     self.logger(
 374                         f"accuracy {n_epoch=} {model.id=} {nb_correct} / {nb_total}"
 375                     )
 376                     self.logger(
 377                         f"back_accuracy {n_epoch=} {model.id=} {back_nb_correct} / {back_nb_total}"
 378                     )
 379
 380                     nb_total += back_nb_total
 381                     nb_correct += back_nb_correct
 382                 else:
 383                     self.logger(
 384                         f"accuracy {n_epoch=} {model.id=} {nb_correct} / {nb_total}"
 385                     )
 386
 387             else:
 388                 nb_total = input.size(0)
 389                 nb_correct = (input == result).long().min(dim=1).values.sum()
 390
 391             return nb_total, nb_correct
 392
 393         train_nb_total, train_nb_correct = compute_accuracy(self.train_w_quizzes[:nmax])
 394
 395         self.logger(
 396             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}%"
 397         )
 398
 399         test_nb_total, test_nb_correct = compute_accuracy(self.test_w_quizzes[:nmax])
 400
 401         self.logger(
 402             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}%"
 403         )
 404
 405         main_test_accuracy = test_nb_correct / test_nb_total
 406         self.logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
 407
 408         ##############################
 409
 410         input = self.test_w_quizzes[:96]
 411         ar_mask = self.make_ar_mask(input)
 412         result = input.clone() * (1 - ar_mask)
 413         seq_logproba = torch.empty(input.size(0), device=self.device)
 414
 415         masked_inplace_autoregression(
 416             model=model,
 417             batch_size=self.batch_size,
 418             input=result,
 419             ar_mask=ar_mask,
 420             seq_logproba=seq_logproba,
 421             temperature=1.0,
 422             deterministic_synthesis=deterministic_synthesis,
 423             progress_bar_desc=None,
 424             device=self.device,
 425         )
 426
 427         mistakes = (input == result).flatten(1).long().min(dim=1).values * 2 - 1
 428
 429         self.save_quizzes(
 430             result_dir,
 431             f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
 432             quizzes=result[:72],
 433             show_to_be_predicted=True,
 434             mistakes=mistakes[:72],
 435         )
 436
 437         return main_test_accuracy
 438
 439     def renew_w_quizzes(self, nb, for_train=True):
 440         input = self.train_w_quizzes if for_train else self.test_w_quizzes
 441         nb = min(nb, input.size(0))
 442         input[:-nb] = input[nb:].clone()
 443         input[-nb:] = self.generate_token_sequences(nb).to(self.device)
 444
 445     def store_c_quizzes(self, new_c_quizzes, for_train=True):
 446         if for_train:
 447             self.train_c_quizzes.append(new_c_quizzes)
 448         else:
 449             self.test_c_quizzes.append(new_c_quizzes)
 450
 451     def compute_correctness(
 452         self,
 453         c_quizzes,
 454         models_for_validation,
 455         bidirectional_validation=False,
 456         deterministic_validation=True,
 457     ):
 458         if bidirectional_validation:
 459             backward_c_quizzes = self.forward_to_backward(c_quizzes)
 460
 461         seq_logproba = torch.zeros(
 462             c_quizzes.size(0),
 463             max([m.id for m in models_for_validation]) + 1,
 464             device=self.device,
 465         )
 466
 467         nb_correct = 0
 468
 469         for model in models_for_validation:
 470             result = c_quizzes.clone()
 471
 472             seq_logproba[...] = 0.0
 473
 474             ar_mask = self.make_ar_mask(result)
 475
 476             masked_inplace_autoregression(
 477                 model=model,
 478                 batch_size=self.batch_size,
 479                 input=result,
 480                 ar_mask=ar_mask,
 481                 seq_logproba=seq_logproba[:, model.id],
 482                 temperature=1.0,
 483                 deterministic_synthesis=deterministic_validation,
 484                 # progress_bar_desc="solving c_quizzes",
 485                 device=self.device,
 486             )
 487
 488             correct = (c_quizzes == result).long().min(dim=-1).values
 489
 490             if bidirectional_validation:
 491                 backward_result = backward_c_quizzes.clone()
 492
 493                 ar_mask = self.make_ar_mask(backward_result)
 494
 495                 masked_inplace_autoregression(
 496                     model=model,
 497                     batch_size=self.batch_size,
 498                     input=backward_result,
 499                     ar_mask=ar_mask,
 500                     seq_logproba=seq_logproba[:, model.id],
 501                     temperature=1.0,
 502                     deterministic_synthesis=deterministic_validation,
 503                     # progress_bar_desc="solving backward c_quizzes",
 504                     device=self.device,
 505                 )
 506
 507                 backward_correct = (
 508                     (backward_c_quizzes == backward_result).long().min(dim=-1).values
 509                 )
 510
 511                 correct *= backward_correct
 512
 513             # endif
 514
 515             nb_correct += correct
 516
 517         return nb_correct, seq_logproba
 518
 519     ###############################################################
 520
 521     def generate_quizzes(self, nb, model_for_generation, temperature=1.0):
 522         c_quizzes = torch.empty(
 523             nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
 524         )
 525
 526         seq_logproba = torch.zeros(nb, device=self.device)
 527
 528         # First, we generate the answer at high temperature
 529
 530         c_quizzes[:, 0] = self.token_backward
 531         c_quizzes[:, 1 + self.answer_len] = self.token_backward
 532
 533         masked_inplace_autoregression(
 534             model=model_for_generation,
 535             batch_size=self.batch_size,
 536             input=c_quizzes,
 537             ar_mask=self.make_ar_mask(c_quizzes, first=True),
 538             seq_logproba=seq_logproba,
 539             temperature=temperature,
 540             deterministic_synthesis=False,
 541             device=self.device,
 542         )
 543
 544         # Then, we generate the prompt at low temperature
 545
 546         masked_inplace_autoregression(
 547             model=model_for_generation,
 548             batch_size=self.batch_size,
 549             input=c_quizzes,
 550             ar_mask=self.make_ar_mask(c_quizzes),
 551             seq_logproba=seq_logproba,
 552             temperature=1 / temperature,
 553             deterministic_synthesis=False,
 554             device=self.device,
 555         )
 556
 557         # Then we return the quizz, and re-generate the response, now
 558         # at low temperature
 559
 560         c_quizzes = self.reverse_time(c_quizzes)
 561
 562         masked_inplace_autoregression(
 563             model=model_for_generation,
 564             batch_size=self.batch_size,
 565             input=c_quizzes,
 566             ar_mask=self.make_ar_mask(c_quizzes),
 567             seq_logproba=seq_logproba,
 568             temperature=1 / temperature,
 569             deterministic_synthesis=False,
 570             device=self.device,
 571         )
 572
 573         return c_quizzes