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
  21 class Gang(nn.Module):
  22     def __init__(self, models, nb_models_for_generation, mode="groupthink"):
  23         super().__init__()
  24         self.models = nn.ModuleList(models)
  25         self.nb_models_for_generation = nb_models_for_generation
  26         self.mode = mode
  27
  28     def forward(self, bs):
  29         # If first = 0, we are re-starting an auto-regressive process,
  30         # that's the right moment to randomize who gonna do it
  31         if bs.first == 0:
  32             self.models_to_use = [
  33                 self.models[k]
  34                 for k in torch.randperm(len(self.models))[
  35                     : self.nb_models_for_generation
  36                 ]
  37             ]
  38
  39         all_the_logits = torch.cat(
  40             [model(bs).x[None] for model in self.models_to_use], dim=0
  41         )
  42
  43         if self.mode == "groupthink":
  44             y = all_the_logits.mean(dim=0)
  45         elif self.mode == "groupwork":
  46             m = torch.rand(all_the_logits.size(), device=all_the_logits.device)
  47             m = (m.sort(dim=0).indices == 0).long()
  48             y = (y * m).sum(dim=0)
  49         else:
  50             raise ValueError(f"Invalid mode {self.mode}")
  51
  52         return BracketedSequence(y, bs.first, bs.nb)
  53
  54
  55 ######################################################################
  56
  57 # ar_mask is a tensor with 0s and 1s, of same shape as input, with
  58 # 1s where tokens should be generated. The others are kept
  59 # unchanged.
  60
  61
  62 def one_batch_masked_inplace_autoregression(
  63     model,
  64     input,
  65     ar_mask,
  66     seq_logproba,
  67     temperature=1.0,
  68     deterministic_synthesis=False,
  69     forbidden_tokens=None,
  70     forced_biases=None,
  71 ):
  72     to_generate = (ar_mask.sum(0) > 0).nonzero()
  73
  74     if to_generate.min() > 0:
  75         model(
  76             BracketedSequence(input, 0, to_generate.min())
  77         )  # Needed to initialize the model's cache
  78     for s in range(to_generate.min(), to_generate.max() + 1):
  79         output = model(BracketedSequence(input, s, 1)).x
  80
  81         logits = output[:, s]
  82
  83         logits = (logits / temperature).log_softmax(dim=-1)
  84
  85         if forbidden_tokens is not None:
  86             logits = logits.masked_fill(forbidden_tokens, float("-inf"))
  87
  88         if forced_biases is not None:
  89             logits = logits + forced_biases[None, :]
  90
  91         if deterministic_synthesis:
  92             t_next = logits.argmax(-1)
  93         else:
  94             dist = torch.distributions.categorical.Categorical(logits=logits)
  95             t_next = dist.sample()
  96
  97         all_n = torch.arange(t_next.size(0))
  98         seq_logproba += logits[all_n, t_next].sum(dim=-1)
  99
 100         input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 101
 102
 103 def masked_inplace_autoregression(
 104     model,
 105     batch_size,
 106     input,
 107     ar_mask,
 108     seq_logproba,
 109     temperature,
 110     deterministic_synthesis,
 111     forbidden_tokens=None,
 112     logit_biases=None,
 113     progress_bar_desc=None,
 114     device=torch.device("cpu"),
 115 ):
 116     assert input.size() == ar_mask.size()
 117
 118     batches = zip(
 119         input.split(batch_size),
 120         ar_mask.split(batch_size),
 121         seq_logproba.split(batch_size),
 122     )
 123
 124     if progress_bar_desc is not None:
 125         batches = tqdm.tqdm(
 126             batches,
 127             dynamic_ncols=True,
 128             desc=progress_bar_desc,
 129             total=(input.size(0) + batch_size - 1) // batch_size,
 130         )
 131
 132     with torch.autograd.no_grad():
 133         t = model.training
 134         model.eval()
 135
 136         for input, ar_mask, seq_logproba in batches:
 137             one_batch_masked_inplace_autoregression(
 138                 model=model,
 139                 input=input,
 140                 ar_mask=ar_mask,
 141                 seq_logproba=seq_logproba,
 142                 temperature=temperature,
 143                 deterministic_synthesis=deterministic_synthesis,
 144                 forbidden_tokens=forbidden_tokens,
 145                 forced_biases=logit_biases,
 146             )
 147
 148         model.train(t)
 149
 150
 151 ######################################################################
 152
 153
 154 class QuizzMachine:
 155     def make_ar_mask(self, input):
 156         b = torch.arange(input.size(1), device=input.device) > input.size(1) // 2
 157         return b.long()[None, :].expand_as(input)
 158
 159     def __init__(
 160         self,
 161         problem,
 162         nb_train_samples,
 163         nb_test_samples,
 164         batch_size,
 165         result_dir,
 166         logger,
 167         device=torch.device("cpu"),
 168     ):
 169         super().__init__()
 170
 171         self.problem = problem
 172         self.batch_size = batch_size
 173         self.device = device
 174         self.logger = logger
 175
 176         self.train_w_quizzes = self.problem.generate_token_sequences(
 177             nb_train_samples
 178         ).to(device)
 179         self.test_w_quizzes = self.problem.generate_token_sequences(nb_test_samples).to(
 180             device
 181         )
 182
 183         self.nb_codes = max(self.train_w_quizzes.max(), self.test_w_quizzes.max()) + 1
 184
 185         self.train_c_quizzes = []
 186         self.test_c_quizzes = []
 187
 188         if result_dir is not None:
 189             self.problem.save_quizzes(
 190                 self.train_w_quizzes[:72], result_dir, "culture_w_quizzes"
 191             )
 192
 193     def batches(self, split="train", desc=None):
 194         assert split in {"train", "test"}
 195         if split == "train":
 196             w_quizzes = self.train_w_quizzes
 197             c_quizzes = self.train_c_quizzes
 198         else:
 199             w_quizzes = self.test_w_quizzes
 200             c_quizzes = self.test_c_quizzes
 201
 202         if len(c_quizzes) > 0:
 203             c_quizzes = torch.cat(c_quizzes, dim=0)
 204             if c_quizzes.size(0) > w_quizzes.size(0) // 2:
 205                 i = torch.randperm(c_quizzes.size(0))[: w_quizzes.size(0) // 2]
 206                 c_quizzes = c_quizzes[i]
 207
 208             i = torch.randperm(w_quizzes.size(0))[
 209                 : w_quizzes.size(0) - c_quizzes.size(0)
 210             ]
 211             w_quizzes = w_quizzes[i]
 212
 213             self.nb_batch_w_quizzes = w_quizzes.size(0)
 214             self.nb_batch_c_quizzes = c_quizzes.size(0)
 215
 216             input = torch.cat([w_quizzes, c_quizzes], dim=0)
 217         else:
 218             input = w_quizzes
 219             self.nb_batch_w_quizzes = w_quizzes.size(0)
 220             self.nb_batch_c_quizzes = 0
 221
 222         # Shuffle
 223         input = input[torch.randperm(input.size(0))]
 224
 225         if desc is None:
 226             desc = f"epoch-{split}"
 227         for batch in tqdm.tqdm(
 228             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 229         ):
 230             yield batch
 231
 232     def vocabulary_size(self):
 233         return self.nb_codes
 234
 235     def produce_results(
 236         self, n_epoch, model, result_dir, deterministic_synthesis, nmax=1000
 237     ):
 238         def compute_accuracy(input):
 239             input = input[:nmax]
 240             ar_mask = self.make_ar_mask(input)
 241             result = input.clone() * (1 - ar_mask)
 242             seq_logproba = torch.empty(input.size(0), device=self.device)
 243
 244             masked_inplace_autoregression(
 245                 model=model,
 246                 batch_size=self.batch_size,
 247                 input=result,
 248                 ar_mask=ar_mask,
 249                 seq_logproba=seq_logproba,
 250                 temperature=1.0,
 251                 deterministic_synthesis=deterministic_synthesis,
 252                 progress_bar_desc=None,
 253                 device=self.device,
 254             )
 255
 256             nb_total, nb_correct = (
 257                 input.size(0),
 258                 (input == result).long().min(dim=1).values.sum(),
 259             )
 260
 261             return nb_total, nb_correct
 262
 263         train_nb_total, train_nb_correct = compute_accuracy(self.train_w_quizzes)
 264
 265         self.logger(
 266             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}%"
 267         )
 268
 269         test_nb_total, test_nb_correct = compute_accuracy(self.test_w_quizzes)
 270
 271         self.logger(
 272             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}%"
 273         )
 274
 275         main_test_accuracy = test_nb_correct / test_nb_total
 276         self.logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
 277
 278         ##############################
 279
 280         input = self.test_w_quizzes[:96]
 281         ar_mask = self.make_ar_mask(input)
 282         result = input.clone() * (1 - ar_mask)
 283         seq_logproba = torch.empty(input.size(0), device=self.device)
 284
 285         masked_inplace_autoregression(
 286             model=model,
 287             batch_size=self.batch_size,
 288             input=result,
 289             ar_mask=ar_mask,
 290             seq_logproba=seq_logproba,
 291             temperature=1.0,
 292             deterministic_synthesis=deterministic_synthesis,
 293             progress_bar_desc=None,
 294             device=self.device,
 295         )
 296
 297         self.problem.save_quizzes(
 298             result[:72], result_dir, f"culture_prediction_{n_epoch:04d}_{model.id:02d}"
 299         )
 300
 301         return main_test_accuracy
 302
 303     def renew_w_quizzes(self, nb, for_train=True):
 304         input = self.train_w_quizzes if for_train else self.test_w_quizzes
 305         nb = min(nb, input.size(0))
 306         input[:-nb] = input[nb:].clone()
 307         input[-nb:] = self.problem.generate_token_sequences(nb).to(self.device)
 308
 309     def store_c_quizzes(self, new_c_quizzes, for_train=True):
 310         if for_train:
 311             self.train_c_quizzes.append(new_c_quizzes)
 312         else:
 313             self.test_c_quizzes.append(new_c_quizzes)
 314
 315     def reverse_time(self, c_quizzes):
 316         token_forward, token_backward = self.problem.direction_tokens()
 317
 318         l = (c_quizzes.size(1) - 1) // 2
 319         direction = c_quizzes[:, l : l + 1]
 320         direction = self.problem.token_forward * (
 321             direction == self.problem.token_backward
 322         ) + self.problem.token_backward * (direction == self.problem.token_forward)
 323
 324         return torch.cat([c_quizzes[:, l + 1 :], direction, c_quizzes[:, :l]], dim=1)
 325
 326     def comput_correctness(self, c_quizzes, models_for_validation):
 327         reversed_c_quizzes = self.reverse_time(c_quizzes)
 328
 329         ar_mask = self.make_ar_mask(c_quizzes)
 330         seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
 331
 332         # Check how many of models can solve the quizzes in both directions
 333
 334         nb_correct = 0
 335
 336         for model in models_for_validation:
 337             result = c_quizzes.clone()
 338
 339             masked_inplace_autoregression(
 340                 model=model,
 341                 batch_size=self.batch_size,
 342                 input=result,
 343                 ar_mask=ar_mask,
 344                 seq_logproba=seq_logproba,
 345                 temperature=1.0,
 346                 deterministic_synthesis=True,
 347                 # progress_bar_desc="solving c_quizzes",
 348                 device=self.device,
 349             )
 350
 351             correct = (c_quizzes == result).long().min(dim=-1).values
 352
 353             reversed_result = reversed_c_quizzes.clone()
 354
 355             masked_inplace_autoregression(
 356                 model=model,
 357                 batch_size=self.batch_size,
 358                 input=reversed_result,
 359                 ar_mask=ar_mask,
 360                 seq_logproba=seq_logproba,
 361                 temperature=1.0,
 362                 deterministic_synthesis=True,
 363                 # progress_bar_desc="solving reversed c_quizzes",
 364                 device=self.device,
 365             )
 366
 367             reversed_correct = (
 368                 (reversed_c_quizzes == reversed_result).long().min(dim=-1).values
 369             )
 370
 371             nb_correct += correct * reversed_correct
 372
 373         return nb_correct
 374
 375     ###############################################################
 376
 377     def generate_quizzes(
 378         self, nb, model_for_generation, min_ave_seq_logproba, reverse_cleanup=False
 379     ):
 380         c_quizzes = torch.empty(
 381             nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
 382         )
 383
 384         ar_mask_prompt = torch.zeros(c_quizzes.size(), device=self.device)
 385         ar_mask_prompt[:, : ar_mask_prompt.size(1) // 2 + 1] = 1
 386         ar_mask_solve = 1 - ar_mask_prompt
 387         seq_logproba = torch.empty(ar_mask_prompt.size(0), device=self.device)
 388
 389         # warnings.warn("noise injection", RuntimeWarning)
 390         temperature = 1
 391         # noise_std = torch.rand(1).item()
 392         # self.logger(f"{noise_std=}")
 393
 394         # mygpt.set_noise_injection(model_for_generation, noise_std)
 395
 396         masked_inplace_autoregression(
 397             model=model_for_generation,
 398             batch_size=self.batch_size,
 399             input=c_quizzes,
 400             ar_mask=ar_mask_prompt,
 401             seq_logproba=seq_logproba,
 402             temperature=temperature,
 403             deterministic_synthesis=False,
 404             # progress_bar_desc="sampling c_quizzes",
 405             device=self.device,
 406         )
 407
 408         # mygpt.set_noise_injection(model_for_generation, 0.0)
 409
 410         ave_seq_logproba = seq_logproba.mean()
 411
 412         masked_inplace_autoregression(
 413             model=model_for_generation,
 414             batch_size=self.batch_size,
 415             input=c_quizzes,
 416             ar_mask=ar_mask_solve,
 417             seq_logproba=seq_logproba,
 418             temperature=temperature,
 419             deterministic_synthesis=True,
 420             # progress_bar_desc="sampling c_quizzes",
 421             device=self.device,
 422         )
 423
 424         if reverse_cleanup:
 425             c_quizzes = self.reverse_time(c_quizzes)
 426             masked_inplace_autoregression(
 427                 model=model_for_generation,
 428                 batch_size=self.batch_size,
 429                 input=c_quizzes,
 430                 ar_mask=ar_mask_solve,
 431                 seq_logproba=seq_logproba,
 432                 temperature=temperature,
 433                 deterministic_synthesis=True,
 434                 # progress_bar_desc="sampling c_quizzes",
 435                 device=self.device,
 436             )
 437
 438         return c_quizzes, seq_logproba.mean()
 439
 440     ######################################################################
 441
 442     def create_c_quizzes(
 443         self,
 444         nb,
 445         model_for_generation,
 446         models_for_validation,
 447         min_ave_seq_logproba,
 448         n_epoch,
 449         result_dir,
 450     ):
 451         c_quizzes, ave_seq_logproba = self.generate_quizzes(
 452             nb, model_for_generation, min_ave_seq_logproba
 453         )
 454
 455         nb_correct = self.comput_correctness(c_quizzes, models_for_validation)
 456
 457         return c_quizzes, nb_correct, ave_seq_logproba
 458
 459     ######################################################################
 460
 461     def gang_create_c_quizzes(
 462         self,
 463         nb,
 464         nb_models_for_generation,
 465         models,
 466         mode,
 467         min_ave_seq_logproba,
 468         n_epoch,
 469         result_dir,
 470     ):
 471         model_for_generation = Gang(models, nb_models_for_generation, mode)
 472         models_for_validation = models
 473         return self.create_c_quizzes(
 474             nb,
 475             model_for_generation,
 476             models_for_validation,
 477             min_ave_seq_logproba,
 478             n_epoch,
 479             result_dir,
 480         )