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, log_prefix=None):
 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             correct = torch.empty(input.size(0), dtype=torch.int64, device=input.device)
 346
 347             n_forward = input[:, 0] == self.token_forward
 348             n_backward = input[:, 0] == self.token_backward
 349
 350             correct[n_forward] = (
 351                 (input[n_forward] == result[n_forward]).long().min(dim=1).values
 352             )
 353
 354             if self.back_accuracy and n_backward.any():
 355                 # accuracy of B->A*->B*=B instead of B->A*=A
 356                 back_input = self.reverse_time(result[n_backward])
 357                 back_input[:, 2 + self.prompt_len :] = input[
 358                     n_backward, 1 : 1 + self.answer_len
 359                 ]
 360                 result[n_backward], correct[n_backward] = compute_accuracy(back_input)
 361
 362             if log_prefix is not None:
 363                 nb_correct = correct[n_forward].sum()
 364                 nb_total = correct[n_forward].size(0)
 365                 back_nb_correct = correct[n_backward].sum()
 366                 back_nb_total = correct[n_backward].size(0)
 367
 368                 self.logger(
 369                     f"accuracy {log_prefix} {n_epoch} {model.id=} {nb_correct} / {nb_total}"
 370                 )
 371
 372                 self.logger(
 373                     f"back_accuracy {log_prefix} {n_epoch} {model.id=} {back_nb_correct} / {back_nb_total}"
 374                 )
 375
 376             return result, correct
 377
 378         compute_accuracy(self.train_w_quizzes[:nmax], log_prefix="train")
 379
 380         result, correct = compute_accuracy(
 381             self.test_w_quizzes[:nmax], log_prefix="test"
 382         )
 383
 384         main_test_accuracy = correct.sum() / correct.size(0)
 385         self.logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
 386
 387         ##############################
 388
 389         self.save_quizzes(
 390             result_dir,
 391             f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
 392             quizzes=result[:72],
 393             show_to_be_predicted=True,
 394             mistakes=correct[:72] * 2 - 1,
 395         )
 396
 397         return main_test_accuracy
 398
 399     def renew_w_quizzes(self, nb, for_train=True):
 400         input = self.train_w_quizzes if for_train else self.test_w_quizzes
 401         nb = min(nb, input.size(0))
 402         input[:-nb] = input[nb:].clone()
 403         input[-nb:] = self.generate_token_sequences(nb).to(self.device)
 404
 405     def store_c_quizzes(self, new_c_quizzes, for_train=True):
 406         if for_train:
 407             self.train_c_quizzes.append(new_c_quizzes)
 408         else:
 409             self.test_c_quizzes.append(new_c_quizzes)
 410
 411     def compute_correctness(
 412         self,
 413         c_quizzes,
 414         models_for_validation,
 415         bidirectional_validation=False,
 416         deterministic_validation=True,
 417     ):
 418         if bidirectional_validation:
 419             backward_c_quizzes = self.forward_to_backward(c_quizzes)
 420
 421         seq_logproba = torch.zeros(
 422             c_quizzes.size(0),
 423             max([m.id for m in models_for_validation]) + 1,
 424             device=self.device,
 425         )
 426
 427         nb_correct = 0
 428
 429         for model in models_for_validation:
 430             result = c_quizzes.clone()
 431
 432             seq_logproba[...] = 0.0
 433
 434             ar_mask = self.make_ar_mask(result)
 435
 436             masked_inplace_autoregression(
 437                 model=model,
 438                 batch_size=self.batch_size,
 439                 input=result,
 440                 ar_mask=ar_mask,
 441                 seq_logproba=seq_logproba[:, model.id],
 442                 temperature=1.0,
 443                 deterministic_synthesis=deterministic_validation,
 444                 # progress_bar_desc="solving c_quizzes",
 445                 device=self.device,
 446             )
 447
 448             correct = (c_quizzes == result).long().min(dim=-1).values
 449
 450             if bidirectional_validation:
 451                 backward_result = backward_c_quizzes.clone()
 452
 453                 ar_mask = self.make_ar_mask(backward_result)
 454
 455                 masked_inplace_autoregression(
 456                     model=model,
 457                     batch_size=self.batch_size,
 458                     input=backward_result,
 459                     ar_mask=ar_mask,
 460                     seq_logproba=seq_logproba[:, model.id],
 461                     temperature=1.0,
 462                     deterministic_synthesis=deterministic_validation,
 463                     # progress_bar_desc="solving backward c_quizzes",
 464                     device=self.device,
 465                 )
 466
 467                 backward_correct = (
 468                     (backward_c_quizzes == backward_result).long().min(dim=-1).values
 469                 )
 470
 471                 correct *= backward_correct
 472
 473             # endif
 474
 475             nb_correct += correct
 476
 477         return nb_correct, seq_logproba
 478
 479     ###############################################################
 480
 481     def generate_quizzes(self, nb, model_for_generation, temperature=1.0):
 482         c_quizzes = torch.empty(
 483             nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
 484         )
 485
 486         seq_logproba = torch.zeros(nb, device=self.device)
 487
 488         # First, we generate the answer at high temperature
 489
 490         c_quizzes[:, 0] = self.token_backward
 491         c_quizzes[:, 1 + self.answer_len] = self.token_backward
 492
 493         masked_inplace_autoregression(
 494             model=model_for_generation,
 495             batch_size=self.batch_size,
 496             input=c_quizzes,
 497             ar_mask=self.make_ar_mask(c_quizzes, first=True),
 498             seq_logproba=seq_logproba,
 499             temperature=temperature,
 500             deterministic_synthesis=False,
 501             device=self.device,
 502         )
 503
 504         # Then, we generate the prompt at low temperature
 505
 506         masked_inplace_autoregression(
 507             model=model_for_generation,
 508             batch_size=self.batch_size,
 509             input=c_quizzes,
 510             ar_mask=self.make_ar_mask(c_quizzes),
 511             seq_logproba=seq_logproba,
 512             temperature=1 / temperature,
 513             deterministic_synthesis=False,
 514             device=self.device,
 515         )
 516
 517         # Then we return the quizz, and re-generate the response, now
 518         # at low temperature
 519
 520         c_quizzes = self.reverse_time(c_quizzes)
 521
 522         masked_inplace_autoregression(
 523             model=model_for_generation,
 524             batch_size=self.batch_size,
 525             input=c_quizzes,
 526             ar_mask=self.make_ar_mask(c_quizzes),
 527             seq_logproba=seq_logproba,
 528             temperature=1 / temperature,
 529             deterministic_synthesis=False,
 530             device=self.device,
 531         )
 532
 533         return c_quizzes