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