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