quiz_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,
  31     deterministic_synthesis,
  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
  54         seq_logproba += logits[all_n, t_next]
  55
  56         input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
  57
  58
  59 def masked_inplace_autoregression(
  60     model,
  61     batch_size,
  62     input,
  63     ar_mask,
  64     seq_logproba,
  65     temperature,
  66     deterministic_synthesis,
  67     forbidden_tokens=None,
  68     logit_biases=None,
  69     progress_bar_desc=None,
  70     device=torch.device("cpu"),
  71 ):
  72     assert input.size() == ar_mask.size()
  73
  74     batches = zip(
  75         input.split(batch_size),
  76         ar_mask.split(batch_size),
  77         seq_logproba.split(batch_size),
  78     )
  79
  80     if progress_bar_desc is not None:
  81         batches = tqdm.tqdm(
  82             batches,
  83             dynamic_ncols=True,
  84             desc=progress_bar_desc,
  85             total=(input.size(0) + batch_size - 1) // batch_size,
  86         )
  87
  88     with torch.autograd.no_grad():
  89         t = model.training
  90         model.eval()
  91
  92         for input, ar_mask, seq_logproba in batches:
  93             one_batch_masked_inplace_autoregression(
  94                 model=model,
  95                 input=input,
  96                 ar_mask=ar_mask,
  97                 seq_logproba=seq_logproba,
  98                 temperature=temperature,
  99                 deterministic_synthesis=deterministic_synthesis,
 100             )
 101
 102         model.train(t)
 103
 104
 105 ######################################################################
 106
 107
 108 class QuizMachine:
 109     def indices_forward_and_backward(self, quizzes):
 110         i_forward = quizzes[:, 0] == self.token_forward
 111         j_forward = quizzes[:, 1 + self.prompt_len] == self.token_forward
 112         i_backward = quizzes[:, 0] == self.token_backward
 113         j_backward = quizzes[:, 1 + self.answer_len] == self.token_backward
 114         assert torch.logical_or(
 115             torch.logical_and(i_forward, j_forward),
 116             torch.logical_and(i_backward, j_backward),
 117         ).all()
 118         return i_forward, i_backward
 119
 120     def non_trivial(self, quizzes):
 121         quizzes = quizzes.clone()
 122         n_forward = quizzes[quizzes[:, 0] == self.token_forward]
 123         n_backward = quizzes[:, 0] == self.token_backward
 124         backward = quizzes[n_backward]
 125         quizzes[n_backward] = self.reverse_time(quizzes[n_backward])
 126         return torch.logical_not(
 127             self.problem.trivial_prompts_and_answers(
 128                 quizzes[:, 1 : 1 + self.prompt_len],
 129                 quizzes[:, 2 + self.prompt_len :],
 130             )
 131         )
 132
 133     def reverse_time(self, quizzes):
 134         i_forward, i_backward = self.indices_forward_and_backward(quizzes)
 135
 136         forward_to_backward = torch.cat(
 137             [
 138                 quizzes[:, 0:1],
 139                 quizzes[:, 2 + self.prompt_len : 2 + self.prompt_len + self.answer_len],
 140                 quizzes[:, 1 + self.prompt_len : 1 + self.prompt_len + 1],
 141                 quizzes[:, 1 : 1 + self.prompt_len],
 142             ],
 143             dim=1,
 144         )
 145
 146         forward_to_backward[:, 0] = self.token_backward
 147         forward_to_backward[:, 1 + self.answer_len] = self.token_backward
 148
 149         backward_to_forward = torch.cat(
 150             [
 151                 quizzes[:, 0:1],
 152                 quizzes[:, 2 + self.answer_len :],
 153                 quizzes[:, 1 + self.answer_len : 2 + self.answer_len],
 154                 quizzes[:, 1 : 1 + self.answer_len],
 155             ],
 156             dim=1,
 157         )
 158
 159         backward_to_forward[:, 0] = self.token_forward
 160         backward_to_forward[:, 1 + self.prompt_len] = self.token_forward
 161
 162         m = i_forward.long()[:, None]
 163
 164         return m * forward_to_backward + (1 - m) * backward_to_forward
 165
 166     def reverse_random_half_in_place(self, quizzes):
 167         i = torch.rand(quizzes.size(0)) < 0.5
 168         if i.any():
 169             quizzes[i] = self.reverse_time(quizzes[i])
 170
 171     def make_ar_mask(self, quizzes, first=False):
 172         i_forward, i_backward = self.indices_forward_and_backward(quizzes)
 173
 174         t = torch.arange(quizzes.size(1), device=quizzes.device)
 175
 176         if first:
 177             m_forward = (t >= 1).long() * (t < 1 + self.prompt_len).long()
 178             m_backward = (t >= 1).long() * (t < 1 + self.answer_len).long()
 179         else:
 180             m_forward = (t >= 2 + self.prompt_len).long()
 181             m_backward = (t >= 2 + self.answer_len).long()
 182
 183         m = i_forward.long()[:, None]
 184
 185         return m * m_forward + (1 - m) * m_backward
 186
 187     def generate_token_sequences(self, nb):
 188         prompts, answers = self.problem.generate_prompts_and_answers(nb)
 189
 190         if self.prompt_len is None:
 191             self.prompt_len = prompts.size(1)
 192
 193         if self.answer_len is None:
 194             self.answer_len = answers.size(1)
 195
 196         assert prompts.size(1) == self.prompt_len and answers.size(1) == self.answer_len
 197
 198         result = []
 199
 200         for prompt, answer in zip(prompts, answers):
 201             a = [
 202                 torch.tensor([self.token_forward]),
 203                 prompt,
 204                 torch.tensor([self.token_forward]),
 205                 answer,
 206             ]
 207
 208             result.append(torch.cat(a, dim=0)[None, :])
 209
 210         return torch.cat(result, dim=0)
 211
 212     def __init__(
 213         self,
 214         problem,
 215         nb_train_samples,
 216         nb_test_samples,
 217         back_accuracy,
 218         batch_size,
 219         result_dir,
 220         logger,
 221         device=torch.device("cpu"),
 222     ):
 223         super().__init__()
 224
 225         v = problem.nb_token_values()
 226         self.token_forward = v
 227         self.token_backward = v + 1
 228         self.nb_token_values = v + 2
 229
 230         self.problem = problem
 231         self.back_accuracy = back_accuracy
 232         self.batch_size = batch_size
 233         self.device = device
 234         self.logger = logger
 235         self.prompt_len = None
 236         self.answer_len = None
 237
 238         self.train_w_quizzes = self.generate_token_sequences(nb_train_samples)
 239         self.reverse_random_half_in_place(self.train_w_quizzes)
 240         self.train_w_quizzes = self.train_w_quizzes.to(device)
 241
 242         self.test_w_quizzes = self.generate_token_sequences(nb_test_samples).to(device)
 243         self.reverse_random_half_in_place(self.test_w_quizzes)
 244         self.test_w_quizzes = self.test_w_quizzes.to(device)
 245
 246         self.train_c_quizzes = []
 247         self.test_c_quizzes = []
 248
 249         if result_dir is not None:
 250             self.save_quizzes(
 251                 result_dir,
 252                 "culture_w_quizzes",
 253                 self.train_w_quizzes[:72],
 254             )
 255
 256     def save_quizzes(
 257         self,
 258         result_dir,
 259         filename_prefix,
 260         quizzes,
 261         mistakes=None,
 262     ):
 263         quizzes = quizzes.clone().to("cpu")
 264         n_forward = quizzes[quizzes[:, 0] == self.token_forward]
 265         n_backward = quizzes[:, 0] == self.token_backward
 266         backward = quizzes[n_backward]
 267         assert n_forward.size(0) + backward.size(0) == quizzes.size(0)
 268         quizzes[n_backward] = self.reverse_time(quizzes[n_backward])
 269
 270         predicted_prompts = n_backward.long()
 271         predicted_answers = 1 - predicted_prompts
 272         if mistakes is not None:
 273             # 0/-1/+1 ~ not-to-predict / predicted wrong / predicted correct
 274             predicted_prompts *= mistakes.to("cpu")
 275             predicted_answers *= mistakes.to("cpu")
 276         else:
 277             # 0/2 ~ not-to-predict / to predict
 278             predicted_prompts *= 2
 279             predicted_answers *= 2
 280
 281         self.problem.save_quizzes(
 282             result_dir,
 283             filename_prefix,
 284             quizzes[:, 1 : 1 + self.prompt_len],
 285             quizzes[:, 2 + self.prompt_len :],
 286             predicted_prompts,
 287             predicted_answers,
 288         )
 289
 290     def batches(self, split="train", desc=None):
 291         assert split in {"train", "test"}
 292         if split == "train":
 293             w_quizzes = self.train_w_quizzes
 294             c_quizzes = self.train_c_quizzes
 295         else:
 296             w_quizzes = self.test_w_quizzes
 297             c_quizzes = self.test_c_quizzes
 298
 299         if len(c_quizzes) > 0:
 300             c_quizzes = torch.cat(c_quizzes, dim=0)
 301             if c_quizzes.size(0) > w_quizzes.size(0) // 2:
 302                 i = torch.randperm(c_quizzes.size(0))[: w_quizzes.size(0) // 2]
 303                 c_quizzes = c_quizzes[i]
 304
 305             i = torch.randperm(w_quizzes.size(0))[
 306                 : w_quizzes.size(0) - c_quizzes.size(0)
 307             ]
 308             w_quizzes = w_quizzes[i]
 309
 310             self.nb_batch_w_quizzes = w_quizzes.size(0)
 311             self.nb_batch_c_quizzes = c_quizzes.size(0)
 312
 313             input = torch.cat([w_quizzes, c_quizzes], dim=0)
 314         else:
 315             input = w_quizzes
 316             self.nb_batch_w_quizzes = w_quizzes.size(0)
 317             self.nb_batch_c_quizzes = 0
 318
 319         # Shuffle
 320         input = input[torch.randperm(input.size(0))]
 321
 322         if desc is None:
 323             desc = f"epoch-{split}"
 324         for batch in tqdm.tqdm(
 325             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 326         ):
 327             yield batch
 328
 329     def vocabulary_size(self):
 330         return self.nb_token_values
 331
 332     def produce_results(
 333         self, n_epoch, model, result_dir, deterministic_synthesis, nmax=1000
 334     ):
 335         def compute_accuracy(input, log_prefix=None):
 336             ar_mask = self.make_ar_mask(input)
 337             result = input.clone() * (1 - ar_mask)
 338             seq_logproba = torch.empty(input.size(0), device=self.device)
 339
 340             masked_inplace_autoregression(
 341                 model=model,
 342                 batch_size=self.batch_size,
 343                 input=result,
 344                 ar_mask=ar_mask,
 345                 seq_logproba=seq_logproba,
 346                 temperature=1.0,
 347                 deterministic_synthesis=deterministic_synthesis,
 348                 progress_bar_desc=None,
 349                 device=self.device,
 350             )
 351
 352             correct = torch.empty(input.size(0), dtype=torch.int64, device=input.device)
 353
 354             n_forward = input[:, 0] == self.token_forward
 355             n_backward = input[:, 0] == self.token_backward
 356
 357             correct[n_forward] = (
 358                 (input[n_forward] == result[n_forward]).long().min(dim=1).values
 359             )
 360
 361             if self.back_accuracy and n_backward.any():
 362                 # accuracy of B->A*->B*=B instead of B->A*=A
 363                 back_input = self.reverse_time(result[n_backward])
 364                 back_input[:, 2 + self.prompt_len :] = input[
 365                     n_backward, 1 : 1 + self.answer_len
 366                 ]
 367                 _, correct[n_backward] = compute_accuracy(back_input)
 368
 369             if log_prefix is not None:
 370                 forward_nb_correct = correct[n_forward].sum()
 371                 forward_nb_total = correct[n_forward].size(0)
 372                 backward_nb_correct = correct[n_backward].sum()
 373                 backward_nb_total = correct[n_backward].size(0)
 374
 375                 self.logger(
 376                     f"{log_prefix}_forward_accuracy {n_epoch} model {model.id} nb_correct {forward_nb_correct} / {forward_nb_total} ({forward_nb_correct*100/forward_nb_total} %)"
 377                 )
 378
 379                 self.logger(
 380                     f"{log_prefix}_backward_accuracy {n_epoch} model {model.id} nb_correct {backward_nb_correct} / {backward_nb_total} ({backward_nb_correct*100/backward_nb_total} %)"
 381                 )
 382
 383             return result, correct
 384
 385         compute_accuracy(self.train_w_quizzes[:nmax], log_prefix="train")
 386
 387         test_result, test_correct = compute_accuracy(
 388             self.test_w_quizzes[:nmax], log_prefix="test"
 389         )
 390
 391         main_test_accuracy = test_correct.sum() / test_correct.size(0)
 392         self.logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
 393
 394         ##############################
 395
 396         self.save_quizzes(
 397             result_dir,
 398             f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
 399             quizzes=test_result[:72],
 400             mistakes=test_correct[:72] * 2 - 1,
 401         )
 402
 403         return main_test_accuracy
 404
 405     def renew_w_quizzes(self, nb, for_train=True):
 406         input = self.train_w_quizzes if for_train else self.test_w_quizzes
 407         nb = min(nb, input.size(0))
 408         input[:-nb] = input[nb:].clone()
 409         fresh_w_quizzes = self.generate_token_sequences(nb)
 410         self.reverse_random_half_in_place(fresh_w_quizzes)
 411         input[-nb:] = fresh_w_quizzes.to(self.device)
 412
 413     def store_c_quizzes(self, new_c_quizzes, for_train=True):
 414         if for_train:
 415             self.train_c_quizzes.append(new_c_quizzes)
 416         else:
 417             self.test_c_quizzes.append(new_c_quizzes)
 418
 419     def compute_correctness(
 420         self,
 421         c_quizzes,
 422         models_for_validation,
 423         bidirectional_validation=False,
 424         deterministic_validation=True,
 425     ):
 426         if bidirectional_validation:
 427             backward_c_quizzes = self.forward_to_backward(c_quizzes)
 428
 429         seq_logproba = torch.zeros(
 430             c_quizzes.size(0),
 431             max([m.id for m in models_for_validation]) + 1,
 432             device=self.device,
 433         )
 434
 435         nb_correct = 0
 436
 437         seq_logproba[...] = 0.0
 438
 439         for model in models_for_validation:
 440             result = c_quizzes.clone()
 441
 442             ar_mask = self.make_ar_mask(result)
 443
 444             masked_inplace_autoregression(
 445                 model=model,
 446                 batch_size=self.batch_size,
 447                 input=result,
 448                 ar_mask=ar_mask,
 449                 seq_logproba=seq_logproba[:, model.id],
 450                 temperature=1.0,
 451                 deterministic_synthesis=deterministic_validation,
 452                 # progress_bar_desc="solving c_quizzes",
 453                 device=self.device,
 454             )
 455
 456             correct = (c_quizzes == result).long().min(dim=-1).values
 457
 458             if bidirectional_validation:
 459                 backward_result = backward_c_quizzes.clone()
 460
 461                 ar_mask = self.make_ar_mask(backward_result)
 462
 463                 masked_inplace_autoregression(
 464                     model=model,
 465                     batch_size=self.batch_size,
 466                     input=backward_result,
 467                     ar_mask=ar_mask,
 468                     seq_logproba=seq_logproba[:, model.id],
 469                     temperature=1.0,
 470                     deterministic_synthesis=deterministic_validation,
 471                     # progress_bar_desc="solving backward c_quizzes",
 472                     device=self.device,
 473                 )
 474
 475                 backward_correct = (
 476                     (backward_c_quizzes == backward_result).long().min(dim=-1).values
 477                 )
 478
 479                 correct *= backward_correct
 480
 481             # endif
 482
 483             nb_correct += correct
 484
 485         return nb_correct, seq_logproba
 486
 487     ###############################################################
 488
 489     def generate_quizzes(self, nb, model_for_generation, temperature=1.0):
 490         c_quizzes = torch.empty(
 491             nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
 492         )
 493
 494         seq_logproba = torch.zeros(nb, device=self.device)
 495
 496         # First, we generate the answer at high temperature
 497
 498         c_quizzes[:, 0] = self.token_backward
 499         c_quizzes[:, 1 + self.answer_len] = self.token_backward
 500
 501         masked_inplace_autoregression(
 502             model=model_for_generation,
 503             batch_size=self.batch_size,
 504             input=c_quizzes,
 505             ar_mask=self.make_ar_mask(c_quizzes, first=True),
 506             seq_logproba=seq_logproba,
 507             temperature=temperature,
 508             deterministic_synthesis=False,
 509             device=self.device,
 510         )
 511
 512         # Then, we generate the prompt at low temperature
 513
 514         masked_inplace_autoregression(
 515             model=model_for_generation,
 516             batch_size=self.batch_size,
 517             input=c_quizzes,
 518             ar_mask=self.make_ar_mask(c_quizzes),
 519             seq_logproba=seq_logproba,
 520             temperature=1 / temperature,
 521             deterministic_synthesis=False,
 522             device=self.device,
 523         )
 524
 525         # Then we return the quizz, and re-generate the response, now
 526         # at low temperature
 527
 528         c_quizzes = self.reverse_time(c_quizzes)
 529
 530         masked_inplace_autoregression(
 531             model=model_for_generation,
 532             batch_size=self.batch_size,
 533             input=c_quizzes,
 534             ar_mask=self.make_ar_mask(c_quizzes),
 535             seq_logproba=seq_logproba,
 536             temperature=1 / temperature,
 537             deterministic_synthesis=False,
 538             device=self.device,
 539         )
 540
 541         return c_quizzes