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