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     forbidden_tokens=None,
  33     forced_biases=None,
  34 ):
  35     to_generate = (ar_mask.sum(0) > 0).nonzero()
  36
  37     if to_generate.min() > 0:
  38         model(
  39             BracketedSequence(input, 0, to_generate.min())
  40         )  # Needed to initialize the model's cache
  41     for s in range(to_generate.min(), to_generate.max() + 1):
  42         output = model(BracketedSequence(input, s, 1)).x
  43
  44         logits = output[:, s]
  45
  46         logits = (logits / temperature).log_softmax(dim=-1)
  47
  48         if forbidden_tokens is not None:
  49             logits = logits.masked_fill(forbidden_tokens, float("-inf"))
  50
  51         if forced_biases is not None:
  52             logits = logits + forced_biases[None, :]
  53
  54         if deterministic_synthesis:
  55             t_next = logits.argmax(-1)
  56         else:
  57             dist = torch.distributions.categorical.Categorical(logits=logits)
  58             t_next = dist.sample()
  59
  60         all_n = torch.arange(t_next.size(0))
  61         seq_logproba += logits[all_n, t_next].sum(dim=-1)
  62
  63         input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
  64
  65
  66 def masked_inplace_autoregression(
  67     model,
  68     batch_size,
  69     input,
  70     ar_mask,
  71     seq_logproba,
  72     temperature,
  73     deterministic_synthesis,
  74     forbidden_tokens=None,
  75     logit_biases=None,
  76     progress_bar_desc=None,
  77     device=torch.device("cpu"),
  78 ):
  79     assert input.size() == ar_mask.size()
  80
  81     batches = zip(
  82         input.split(batch_size),
  83         ar_mask.split(batch_size),
  84         seq_logproba.split(batch_size),
  85     )
  86
  87     if progress_bar_desc is not None:
  88         batches = tqdm.tqdm(
  89             batches,
  90             dynamic_ncols=True,
  91             desc=progress_bar_desc,
  92             total=(input.size(0) + batch_size - 1) // batch_size,
  93         )
  94
  95     with torch.autograd.no_grad():
  96         t = model.training
  97         model.eval()
  98
  99         for input, ar_mask, seq_logproba in batches:
 100             one_batch_masked_inplace_autoregression(
 101                 model=model,
 102                 input=input,
 103                 ar_mask=ar_mask,
 104                 seq_logproba=seq_logproba,
 105                 temperature=temperature,
 106                 deterministic_synthesis=deterministic_synthesis,
 107                 forbidden_tokens=forbidden_tokens,
 108                 forced_biases=logit_biases,
 109             )
 110
 111         model.train(t)
 112
 113
 114 ######################################################################
 115
 116
 117 class QuizzMachine:
 118     def make_ar_mask(self, input):
 119         b = torch.arange(input.size(1), device=input.device) > input.size(1) // 2
 120         return b.long()[None, :].expand_as(input)
 121
 122     def generate_token_sequences(self, nb):
 123         prompts, answers = self.problem.generate_prompts_and_answers(nb)
 124         result = []
 125
 126         for prompt, answer in zip(prompts, answers):
 127             if torch.rand(1) < 0.5:
 128                 a = [torch.tensor([self.token_forward]), prompt, answer]
 129             else:
 130                 a = [torch.tensor([self.token_backward]), answer, prompt]
 131
 132             result.append(torch.cat(a, dim=0)[None, :])
 133
 134         return torch.cat(result, dim=0)
 135
 136     def __init__(
 137         self,
 138         problem,
 139         nb_train_samples,
 140         nb_test_samples,
 141         batch_size,
 142         result_dir,
 143         logger,
 144         device=torch.device("cpu"),
 145     ):
 146         super().__init__()
 147
 148         v = problem.nb_token_values()
 149         self.token_forward = v
 150         self.token_backward = v + 1
 151         self.nb_token_values = v + 2
 152
 153         self.problem = problem
 154         self.batch_size = batch_size
 155         self.device = device
 156         self.logger = logger
 157
 158         self.train_w_quizzes = self.generate_token_sequences(nb_train_samples).to(
 159             device
 160         )
 161
 162         self.test_w_quizzes = self.generate_token_sequences(nb_test_samples).to(device)
 163
 164         self.train_c_quizzes = []
 165         self.test_c_quizzes = []
 166
 167         if result_dir is not None:
 168             self.save_quizzes(
 169                 result_dir, "culture_w_quizzes", self.train_w_quizzes[:72]
 170             )
 171
 172     def save_quizzes(self, result_dir, filename_prefix, quizzes, prediction=False):
 173         print(f"DEBUG {quizzes.size()=}")
 174         l = (quizzes.size(1) - 1) // 2
 175         forward = (quizzes[:, 0] == self.token_forward).long()
 176         backward = (quizzes[:, 0] == self.token_backward).long()
 177         assert forward.equal(1 - backward)
 178         first = quizzes[:, 1 : 1 + l]
 179         second = quizzes[:, 1 + l : 1 + 2 * l]
 180         prompts = forward[:, None] * first + backward[:, None] * second
 181         answers = forward[:, None] * second + backward[:, None] * first
 182
 183         if prediction:
 184             predicted_prompts = backward
 185             predicted_answers = forward
 186         else:
 187             predicted_prompts = None
 188             predicted_answers = None
 189
 190         self.problem.save_quizzes(
 191             result_dir,
 192             filename_prefix,
 193             prompts,
 194             answers,
 195             predicted_prompts,
 196             predicted_answers,
 197         )
 198
 199     def batches(self, split="train", desc=None):
 200         assert split in {"train", "test"}
 201         if split == "train":
 202             w_quizzes = self.train_w_quizzes
 203             c_quizzes = self.train_c_quizzes
 204         else:
 205             w_quizzes = self.test_w_quizzes
 206             c_quizzes = self.test_c_quizzes
 207
 208         if len(c_quizzes) > 0:
 209             c_quizzes = torch.cat(c_quizzes, dim=0)
 210             if c_quizzes.size(0) > w_quizzes.size(0) // 2:
 211                 i = torch.randperm(c_quizzes.size(0))[: w_quizzes.size(0) // 2]
 212                 c_quizzes = c_quizzes[i]
 213
 214             i = torch.randperm(w_quizzes.size(0))[
 215                 : w_quizzes.size(0) - c_quizzes.size(0)
 216             ]
 217             w_quizzes = w_quizzes[i]
 218
 219             self.nb_batch_w_quizzes = w_quizzes.size(0)
 220             self.nb_batch_c_quizzes = c_quizzes.size(0)
 221
 222             input = torch.cat([w_quizzes, c_quizzes], dim=0)
 223         else:
 224             input = w_quizzes
 225             self.nb_batch_w_quizzes = w_quizzes.size(0)
 226             self.nb_batch_c_quizzes = 0
 227
 228         # Shuffle
 229         input = input[torch.randperm(input.size(0))]
 230
 231         if desc is None:
 232             desc = f"epoch-{split}"
 233         for batch in tqdm.tqdm(
 234             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 235         ):
 236             yield batch
 237
 238     def vocabulary_size(self):
 239         return self.nb_token_values
 240
 241     def produce_results(
 242         self, n_epoch, model, result_dir, deterministic_synthesis, nmax=1000
 243     ):
 244         def compute_accuracy(input):
 245             input = input[:nmax]
 246             ar_mask = self.make_ar_mask(input)
 247             result = input.clone() * (1 - ar_mask)
 248             seq_logproba = torch.empty(input.size(0), device=self.device)
 249
 250             masked_inplace_autoregression(
 251                 model=model,
 252                 batch_size=self.batch_size,
 253                 input=result,
 254                 ar_mask=ar_mask,
 255                 seq_logproba=seq_logproba,
 256                 temperature=1.0,
 257                 deterministic_synthesis=deterministic_synthesis,
 258                 progress_bar_desc=None,
 259                 device=self.device,
 260             )
 261
 262             nb_total, nb_correct = (
 263                 input.size(0),
 264                 (input == result).long().min(dim=1).values.sum(),
 265             )
 266
 267             return nb_total, nb_correct
 268
 269         train_nb_total, train_nb_correct = compute_accuracy(self.train_w_quizzes)
 270
 271         self.logger(
 272             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}%"
 273         )
 274
 275         test_nb_total, test_nb_correct = compute_accuracy(self.test_w_quizzes)
 276
 277         self.logger(
 278             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}%"
 279         )
 280
 281         main_test_accuracy = test_nb_correct / test_nb_total
 282         self.logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
 283
 284         ##############################
 285
 286         input = self.test_w_quizzes[:96]
 287         ar_mask = self.make_ar_mask(input)
 288         result = input.clone() * (1 - ar_mask)
 289         seq_logproba = torch.empty(input.size(0), device=self.device)
 290
 291         masked_inplace_autoregression(
 292             model=model,
 293             batch_size=self.batch_size,
 294             input=result,
 295             ar_mask=ar_mask,
 296             seq_logproba=seq_logproba,
 297             temperature=1.0,
 298             deterministic_synthesis=deterministic_synthesis,
 299             progress_bar_desc=None,
 300             device=self.device,
 301         )
 302
 303         self.save_quizzes(
 304             result_dir,
 305             f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
 306             quizzes=result[:72],
 307             prediction=True,
 308         )
 309
 310         return main_test_accuracy
 311
 312     def renew_w_quizzes(self, nb, for_train=True):
 313         input = self.train_w_quizzes if for_train else self.test_w_quizzes
 314         nb = min(nb, input.size(0))
 315         input[:-nb] = input[nb:].clone()
 316         input[-nb:] = self.generate_token_sequences(nb).to(self.device)
 317
 318     def store_c_quizzes(self, new_c_quizzes, for_train=True):
 319         if for_train:
 320             self.train_c_quizzes.append(new_c_quizzes)
 321         else:
 322             self.test_c_quizzes.append(new_c_quizzes)
 323
 324     def reverse_time(self, c_quizzes):
 325         l = (c_quizzes.size(1) - 1) // 2
 326         direction = c_quizzes[:, 0:1]
 327         direction = self.token_forward * (
 328             direction == self.token_backward
 329         ) + self.token_backward * (direction == self.token_forward)
 330
 331         return torch.cat(
 332             [direction, c_quizzes[:, l + 1 :], c_quizzes[:, 1 : l + 1]], dim=1
 333         )
 334
 335     def compute_correctness(
 336         self, c_quizzes, models_for_validation, both_directions=True
 337     ):
 338         reversed_c_quizzes = self.reverse_time(c_quizzes)
 339
 340         ar_mask = self.make_ar_mask(c_quizzes)
 341         seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
 342
 343         # Check how many of models can solve the quizzes in both directions
 344
 345         nb_correct = 0
 346
 347         for model in models_for_validation:
 348             result = c_quizzes.clone()
 349
 350             masked_inplace_autoregression(
 351                 model=model,
 352                 batch_size=self.batch_size,
 353                 input=result,
 354                 ar_mask=ar_mask,
 355                 seq_logproba=seq_logproba,
 356                 temperature=1.0,
 357                 deterministic_synthesis=True,
 358                 # progress_bar_desc="solving c_quizzes",
 359                 device=self.device,
 360             )
 361
 362             correct = (c_quizzes == result).long().min(dim=-1).values
 363
 364             if both_directions:
 365                 reversed_result = reversed_c_quizzes.clone()
 366
 367                 masked_inplace_autoregression(
 368                     model=model,
 369                     batch_size=self.batch_size,
 370                     input=reversed_result,
 371                     ar_mask=ar_mask,
 372                     seq_logproba=seq_logproba,
 373                     temperature=1.0,
 374                     deterministic_synthesis=True,
 375                     # progress_bar_desc="solving reversed c_quizzes",
 376                     device=self.device,
 377                 )
 378
 379                 reversed_correct = (
 380                     (reversed_c_quizzes == reversed_result).long().min(dim=-1).values
 381                 )
 382
 383                 correct *= reversed_correct
 384
 385             # endif
 386
 387             nb_correct += correct
 388
 389         return nb_correct
 390
 391     ###############################################################
 392
 393     def generate_quizzes(self, nb, model_for_generation, reverse_cleanup=False):
 394         c_quizzes = torch.empty(
 395             nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
 396         )
 397
 398         ar_mask_first = torch.zeros(c_quizzes.size(), device=self.device)
 399         ar_mask_first[:, : ar_mask_first.size(1) // 2 + 1] = 1
 400         ar_mask_second = 1 - ar_mask_first
 401         ar_mask_first[:, 0] = 0
 402         ar_mask_second[:, 0] = 0
 403
 404         seq_logproba = torch.empty(ar_mask_first.size(0), device=self.device)
 405
 406         if reverse_cleanup:
 407             temperature = 10.0
 408         else:
 409             temperature = 1.0
 410
 411         # First, we generate the answer at high temperature
 412
 413         c_quizzes[:, 0] = self.token_backward
 414
 415         masked_inplace_autoregression(
 416             model=model_for_generation,
 417             batch_size=self.batch_size,
 418             input=c_quizzes,
 419             ar_mask=ar_mask_first,
 420             seq_logproba=seq_logproba,
 421             temperature=temperature,
 422             deterministic_synthesis=False,
 423             device=self.device,
 424         )
 425
 426         ave_seq_logproba = seq_logproba.mean()
 427
 428         # Then, we generate the prompt deterministically
 429
 430         masked_inplace_autoregression(
 431             model=model_for_generation,
 432             batch_size=self.batch_size,
 433             input=c_quizzes,
 434             ar_mask=ar_mask_second,
 435             seq_logproba=seq_logproba,
 436             temperature=temperature,
 437             deterministic_synthesis=True,
 438             device=self.device,
 439         )
 440
 441         # Then we return the quizz, and re-generate the response, now
 442         # deterministically
 443
 444         c_quizzes = self.reverse_time(c_quizzes)
 445
 446         masked_inplace_autoregression(
 447             model=model_for_generation,
 448             batch_size=self.batch_size,
 449             input=c_quizzes,
 450             ar_mask=ar_mask_second,
 451             seq_logproba=seq_logproba,
 452             temperature=temperature,
 453             deterministic_synthesis=True,
 454             device=self.device,
 455         )
 456
 457         return c_quizzes, seq_logproba.mean()