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 make_ar_mask(self, input, first, nb):
 109         i = torch.arange(input.size(1), device=input.device)
 110         b = torch.logical_and(i >= first, i < first + nb)
 111         return b.long()[None, :].expand_as(input)
 112
 113     def generate_token_sequences(self, nb):
 114         prompts, answers = self.problem.generate_prompts_and_answers(nb)
 115
 116         if self.prompt_len is None:
 117             self.prompt_len = prompts.size(1)
 118
 119         if self.answer_len is None:
 120             self.answer_len = answers.size(1)
 121
 122         assert prompts.size(1) == self.prompt_len and answers.size(1) == self.answer_len
 123
 124         result = []
 125
 126         for prompt, answer in zip(prompts, answers):
 127             if torch.rand(1) < 0.5:
 128                 a = [
 129                     torch.tensor([self.token_forward]),
 130                     prompt,
 131                     torch.tensor([self.token_forward]),
 132                     answer,
 133                 ]
 134             else:
 135                 a = [
 136                     torch.tensor([self.token_backward]),
 137                     answer,
 138                     torch.tensor([self.token_backward]),
 139                     prompt,
 140                 ]
 141
 142             result.append(torch.cat(a, dim=0)[None, :])
 143
 144         return torch.cat(result, dim=0)
 145
 146     def __init__(
 147         self,
 148         problem,
 149         nb_train_samples,
 150         nb_test_samples,
 151         batch_size,
 152         result_dir,
 153         logger,
 154         device=torch.device("cpu"),
 155     ):
 156         super().__init__()
 157
 158         v = problem.nb_token_values()
 159         self.token_forward = v
 160         self.token_backward = v + 1
 161         self.nb_token_values = v + 2
 162
 163         self.problem = problem
 164         self.batch_size = batch_size
 165         self.device = device
 166         self.logger = logger
 167         self.prompt_len = None
 168         self.answer_len = None
 169
 170         self.train_w_quizzes = self.generate_token_sequences(nb_train_samples).to(
 171             device
 172         )
 173
 174         self.test_w_quizzes = self.generate_token_sequences(nb_test_samples).to(device)
 175
 176         self.train_c_quizzes = []
 177         self.test_c_quizzes = []
 178
 179         if result_dir is not None:
 180             self.save_quizzes(
 181                 result_dir, "culture_w_quizzes", self.train_w_quizzes[:72]
 182             )
 183
 184     def save_quizzes(self, result_dir, filename_prefix, quizzes, prediction=False):
 185         l = (quizzes.size(1) - 1) // 2
 186         forward = (quizzes[:, 0] == self.token_forward).long()
 187         backward = (quizzes[:, 0] == self.token_backward).long()
 188         assert forward.equal(1 - backward)
 189         first = quizzes[:, 1 : 1 + l]
 190         second = quizzes[:, 1 + l : 1 + 2 * l]
 191         prompts = forward[:, None] * first + backward[:, None] * second
 192         answers = forward[:, None] * second + backward[:, None] * first
 193
 194         if prediction:
 195             predicted_prompts = backward
 196             predicted_answers = forward
 197         else:
 198             predicted_prompts = None
 199             predicted_answers = None
 200
 201         self.problem.save_quizzes(
 202             result_dir,
 203             filename_prefix,
 204             prompts,
 205             answers,
 206             predicted_prompts,
 207             predicted_answers,
 208         )
 209
 210     def batches(self, split="train", desc=None):
 211         assert split in {"train", "test"}
 212         if split == "train":
 213             w_quizzes = self.train_w_quizzes
 214             c_quizzes = self.train_c_quizzes
 215         else:
 216             w_quizzes = self.test_w_quizzes
 217             c_quizzes = self.test_c_quizzes
 218
 219         if len(c_quizzes) > 0:
 220             c_quizzes = torch.cat(c_quizzes, dim=0)
 221             if c_quizzes.size(0) > w_quizzes.size(0) // 2:
 222                 i = torch.randperm(c_quizzes.size(0))[: w_quizzes.size(0) // 2]
 223                 c_quizzes = c_quizzes[i]
 224
 225             i = torch.randperm(w_quizzes.size(0))[
 226                 : w_quizzes.size(0) - c_quizzes.size(0)
 227             ]
 228             w_quizzes = w_quizzes[i]
 229
 230             self.nb_batch_w_quizzes = w_quizzes.size(0)
 231             self.nb_batch_c_quizzes = c_quizzes.size(0)
 232
 233             input = torch.cat([w_quizzes, c_quizzes], dim=0)
 234         else:
 235             input = w_quizzes
 236             self.nb_batch_w_quizzes = w_quizzes.size(0)
 237             self.nb_batch_c_quizzes = 0
 238
 239         # Shuffle
 240         input = input[torch.randperm(input.size(0))]
 241
 242         if desc is None:
 243             desc = f"epoch-{split}"
 244         for batch in tqdm.tqdm(
 245             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 246         ):
 247             yield batch
 248
 249     def vocabulary_size(self):
 250         return self.nb_token_values
 251
 252     def produce_results(
 253         self, n_epoch, model, result_dir, deterministic_synthesis, nmax=1000
 254     ):
 255         def compute_accuracy(input):
 256             input = input[:nmax]
 257             ar_mask = self.make_ar_mask(input, 2 + self.prompt_len, self.answer_len)
 258             result = input.clone() * (1 - ar_mask)
 259             seq_logproba = torch.empty(input.size(0), device=self.device)
 260
 261             masked_inplace_autoregression(
 262                 model=model,
 263                 batch_size=self.batch_size,
 264                 input=result,
 265                 ar_mask=ar_mask,
 266                 seq_logproba=seq_logproba,
 267                 temperature=1.0,
 268                 deterministic_synthesis=deterministic_synthesis,
 269                 progress_bar_desc=None,
 270                 device=self.device,
 271             )
 272
 273             nb_total = input.size(0)
 274             nb_correct = (input == result).long().min(dim=1).values.sum()
 275
 276             return nb_total, nb_correct
 277
 278         train_nb_total, train_nb_correct = compute_accuracy(self.train_w_quizzes)
 279
 280         self.logger(
 281             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}%"
 282         )
 283
 284         test_nb_total, test_nb_correct = compute_accuracy(self.test_w_quizzes)
 285
 286         self.logger(
 287             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}%"
 288         )
 289
 290         main_test_accuracy = test_nb_correct / test_nb_total
 291         self.logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
 292
 293         ##############################
 294
 295         input = self.test_w_quizzes[:96]
 296         ar_mask = self.make_ar_mask(input, 2 + self.prompt_len, self.answer_len)
 297         result = input.clone() * (1 - ar_mask)
 298         seq_logproba = torch.empty(input.size(0), device=self.device)
 299
 300         masked_inplace_autoregression(
 301             model=model,
 302             batch_size=self.batch_size,
 303             input=result,
 304             ar_mask=ar_mask,
 305             seq_logproba=seq_logproba,
 306             temperature=1.0,
 307             deterministic_synthesis=deterministic_synthesis,
 308             progress_bar_desc=None,
 309             device=self.device,
 310         )
 311
 312         self.save_quizzes(
 313             result_dir,
 314             f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
 315             quizzes=result[:72],
 316             prediction=True,
 317         )
 318
 319         return main_test_accuracy
 320
 321     def renew_w_quizzes(self, nb, for_train=True):
 322         input = self.train_w_quizzes if for_train else self.test_w_quizzes
 323         nb = min(nb, input.size(0))
 324         input[:-nb] = input[nb:].clone()
 325         input[-nb:] = self.generate_token_sequences(nb).to(self.device)
 326
 327     def store_c_quizzes(self, new_c_quizzes, for_train=True):
 328         if for_train:
 329             self.train_c_quizzes.append(new_c_quizzes)
 330         else:
 331             self.test_c_quizzes.append(new_c_quizzes)
 332
 333     def forward_to_backward(self, c_quizzes):
 334         prompts = c_quizzes[:, 1 : 1 + self.prompt_len]
 335         answers = c_quizzes[:, 2 + self.prompt_len :]
 336         return torch.cat(
 337             [
 338                 c_quizzes.new_full((c_quizzes, 1), self.token_backward),
 339                 answers,
 340                 c_quizzes.new_full((c_quizzes, 1), self.token_backward),
 341                 prompts,
 342             ],
 343             dim=1,
 344         )
 345
 346     def backward_to_forward(self, c_quizzes):
 347         answers = c_quizzes[:, 1 : 1 + self.answer_len :]
 348         prompts = c_quizzes[:, 2 + self.answer_len :]
 349         return torch.cat(
 350             [
 351                 c_quizzes.new_full((c_quizzes.size(0), 1), self.token_forward),
 352                 prompts,
 353                 c_quizzes.new_full((c_quizzes.size(0), 1), self.token_forward),
 354                 answers,
 355             ],
 356             dim=1,
 357         )
 358
 359     def compute_correctness(
 360         self,
 361         c_quizzes,
 362         models_for_validation,
 363         bidirectional_validation=False,
 364         deterministic_validation=True,
 365     ):
 366         if bidirectional_validation:
 367             backward_c_quizzes = self.forward_to_backward(c_quizzes)
 368
 369         seq_logproba = torch.zeros(
 370             c_quizzes.size(0),
 371             max([m.id for m in models_for_validation]) + 1,
 372             device=self.device,
 373         )
 374
 375         nb_correct = 0
 376
 377         for model in models_for_validation:
 378             result = c_quizzes.clone()
 379
 380             seq_logproba[...] = 0.0
 381
 382             ar_mask = self.make_ar_mask(result, 2 + self.prompt_len, self.answer_len)
 383
 384             masked_inplace_autoregression(
 385                 model=model,
 386                 batch_size=self.batch_size,
 387                 input=result,
 388                 ar_mask=ar_mask,
 389                 seq_logproba=seq_logproba[:, model.id],
 390                 temperature=1.0,
 391                 deterministic_synthesis=deterministic_validation,
 392                 # progress_bar_desc="solving c_quizzes",
 393                 device=self.device,
 394             )
 395
 396             correct = (c_quizzes == result).long().min(dim=-1).values
 397
 398             if bidirectional_validation:
 399                 backward_result = backward_c_quizzes.clone()
 400
 401                 ar_mask = self.make_ar_mask(
 402                     backward_result, 2 + self.answer_len, self.prompt_len
 403                 )
 404
 405                 masked_inplace_autoregression(
 406                     model=model,
 407                     batch_size=self.batch_size,
 408                     input=backward_result,
 409                     ar_mask=ar_mask,
 410                     seq_logproba=seq_logproba[:, model.id],
 411                     temperature=1.0,
 412                     deterministic_synthesis=deterministic_validation,
 413                     # progress_bar_desc="solving backward c_quizzes",
 414                     device=self.device,
 415                 )
 416
 417                 backward_correct = (
 418                     (backward_c_quizzes == backward_result).long().min(dim=-1).values
 419                 )
 420
 421                 correct *= backward_correct
 422
 423             # endif
 424
 425             nb_correct += correct
 426
 427         return nb_correct, seq_logproba
 428
 429     ###############################################################
 430
 431     def generate_quizzes(self, nb, model_for_generation, temperature=1.0):
 432         c_quizzes = torch.empty(
 433             nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
 434         )
 435
 436         ar_mask_first = torch.zeros(c_quizzes.size(), device=self.device)
 437         ar_mask_first[:, : ar_mask_first.size(1) // 2 + 1] = 1
 438         ar_mask_second = 1 - ar_mask_first
 439         ar_mask_first[:, 0] = 0
 440         ar_mask_second[:, 0] = 0
 441
 442         seq_logproba = torch.zeros(ar_mask_first.size(0), device=self.device)
 443
 444         # First, we generate the answer at high temperature
 445
 446         c_quizzes[:, 0] = self.token_backward
 447
 448         masked_inplace_autoregression(
 449             model=model_for_generation,
 450             batch_size=self.batch_size,
 451             input=c_quizzes,
 452             ar_mask=ar_mask_first,
 453             seq_logproba=seq_logproba,
 454             temperature=temperature,
 455             deterministic_synthesis=False,
 456             device=self.device,
 457         )
 458
 459         # Then, we generate the prompt at low temperature
 460
 461         masked_inplace_autoregression(
 462             model=model_for_generation,
 463             batch_size=self.batch_size,
 464             input=c_quizzes,
 465             ar_mask=ar_mask_second,
 466             seq_logproba=seq_logproba,
 467             temperature=1 / temperature,
 468             deterministic_synthesis=False,
 469             device=self.device,
 470         )
 471
 472         # Then we return the quizz, and re-generate the response, now
 473         # at low temperature
 474
 475         c_quizzes = self.backward_to_forward(c_quizzes)
 476
 477         masked_inplace_autoregression(
 478             model=model_for_generation,
 479             batch_size=self.batch_size,
 480             input=c_quizzes,
 481             ar_mask=ar_mask_second,
 482             seq_logproba=seq_logproba,
 483             temperature=1 / temperature,
 484             deterministic_synthesis=False,
 485             device=self.device,
 486         )
 487
 488         return c_quizzes