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