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 from mygpt import BracketedSequence
  16
  17 ######################################################################
  18
  19
  20 def masked_inplace_autoregression(
  21     model,
  22     batch_size,
  23     input,
  24     ar_mask,
  25     seq_logproba,
  26     temperature,
  27     deterministic_synthesis,
  28     forbidden_tokens=None,
  29     logit_biases=None,
  30     progress_bar_desc="autoregression",
  31     device=torch.device("cpu"),
  32 ):
  33     assert input.size() == ar_mask.size()
  34
  35     batches = zip(
  36         input.split(batch_size),
  37         ar_mask.split(batch_size),
  38         seq_logproba.split(batch_size),
  39     )
  40
  41     if progress_bar_desc is not None:
  42         batches = tqdm.tqdm(
  43             batches,
  44             dynamic_ncols=True,
  45             desc=progress_bar_desc,
  46             total=(input.size(0) + batch_size - 1) // batch_size,
  47         )
  48
  49     with torch.autograd.no_grad():
  50         t = model.training
  51         model.eval()
  52
  53         for input, ar_mask, seq_logproba in batches:
  54             model.masked_inplace_autoregression(
  55                 input=input,
  56                 ar_mask=ar_mask,
  57                 seq_logproba=seq_logproba,
  58                 temperature=temperature,
  59                 deterministic_synthesis=deterministic_synthesis,
  60                 forbidden_tokens=forbidden_tokens,
  61                 forced_biases=logit_biases,
  62             )
  63
  64         model.train(t)
  65
  66
  67 ######################################################################
  68
  69
  70 class QuizzMachine:
  71     def make_ar_mask(self, input):
  72         b = torch.arange(input.size(1), device=input.device) > input.size(1) // 2
  73         return b.long()[None, :].expand_as(input)
  74
  75     def __init__(
  76         self,
  77         problem,
  78         nb_train_samples,
  79         nb_test_samples,
  80         batch_size,
  81         result_dir=None,
  82         logger=None,
  83         device=torch.device("cpu"),
  84     ):
  85         super().__init__()
  86
  87         self.problem = problem
  88         self.batch_size = batch_size
  89         self.device = device
  90
  91         self.train_w_quizzes = self.problem.generate_seq(nb_train_samples).to(device)
  92         self.test_w_quizzes = self.problem.generate_seq(nb_test_samples).to(device)
  93
  94         self.nb_codes = max(self.train_w_quizzes.max(), self.test_w_quizzes.max()) + 1
  95
  96         self.train_c_quizzes = []
  97         self.test_c_quizzes = []
  98
  99         if result_dir is not None:
 100             self.problem.save_quizzes(
 101                 self.train_w_quizzes[:72], result_dir, f"culture_w_quizzes", logger
 102             )
 103
 104     def batches(self, split="train", desc=None):
 105         assert split in {"train", "test"}
 106         if split == "train":
 107             w_quizzes = self.train_w_quizzes
 108             c_quizzes = self.train_c_quizzes
 109         else:
 110             w_quizzes = self.test_w_quizzes
 111             c_quizzes = self.test_c_quizzes
 112
 113         if len(c_quizzes) > 0:
 114             c_quizzes = torch.cat(c_quizzes, dim=0)
 115             if c_quizzes.size(0) > w_quizzes.size(0) // 2:
 116                 i = torch.randperm(w_quizzes.size(0))[: w_quizzes.size(0) // 2]
 117                 c_quizzes = c_quizzes[i]
 118
 119             i = torch.randperm(w_quizzes.size(0))[
 120                 : w_quizzes.size(0) - c_quizzes.size(0)
 121             ]
 122             w_quizzes = w_quizzes[i]
 123
 124             self.nb_batch_w_quizzes = w_quizzes.size(0)
 125             self.nb_batch_c_quizzes = c_quizzes.size(0)
 126
 127             input = torch.cat([w_quizzes, c_quizzes], dim=0)
 128         else:
 129             input = w_quizzes
 130             self.nb_batch_w_quizzes = w_quizzes.size(0)
 131             self.nb_batch_c_quizzes = 0
 132
 133         # Shuffle
 134         input = input[torch.randperm(input.size(0))]
 135
 136         if desc is None:
 137             desc = f"epoch-{split}"
 138         for batch in tqdm.tqdm(
 139             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 140         ):
 141             yield batch
 142
 143     def vocabulary_size(self):
 144         return self.nb_codes
 145
 146     def produce_results(
 147         self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
 148     ):
 149         def compute_accuracy(input, logger=None):
 150             input = input[:nmax]
 151             ar_mask = self.make_ar_mask(input)
 152             result = input.clone() * (1 - ar_mask)
 153             seq_logproba = torch.empty(input.size(0), device=self.device)
 154
 155             masked_inplace_autoregression(
 156                 model=model,
 157                 batch_size=self.batch_size,
 158                 input=result,
 159                 ar_mask=ar_mask,
 160                 seq_logproba=seq_logproba,
 161                 temperature=1.0,
 162                 deterministic_synthesis=deterministic_synthesis,
 163                 progress_bar_desc=None,
 164                 device=self.device,
 165             )
 166
 167             nb_total, nb_correct = (
 168                 input.size(0),
 169                 (input == result).long().min(dim=1).values.sum(),
 170             )
 171
 172             return nb_total, nb_correct
 173
 174         train_nb_total, train_nb_correct = compute_accuracy(self.train_w_quizzes)
 175
 176         logger(
 177             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}%"
 178         )
 179
 180         test_nb_total, test_nb_correct = compute_accuracy(self.test_w_quizzes, logger)
 181
 182         logger(
 183             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}%"
 184         )
 185
 186         main_test_accuracy = test_nb_correct / test_nb_total
 187         logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
 188
 189         ##############################
 190
 191         input = self.test_w_quizzes[:96]
 192         ar_mask = self.make_ar_mask(input)
 193         result = input.clone() * (1 - ar_mask)
 194         seq_logproba = torch.empty(input.size(0), device=self.device)
 195
 196         masked_inplace_autoregression(
 197             model=model,
 198             batch_size=self.batch_size,
 199             input=result,
 200             ar_mask=ar_mask,
 201             seq_logproba=seq_logproba,
 202             temperature=1.0,
 203             deterministic_synthesis=deterministic_synthesis,
 204             progress_bar_desc=None,
 205             device=self.device,
 206         )
 207
 208         self.problem.save_quizzes(
 209             result[:72],
 210             result_dir,
 211             f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
 212             logger,
 213         )
 214
 215         return main_test_accuracy
 216
 217     def renew_w_quizzes(self, nb, for_train=True):
 218         input = self.train_w_quizzes if for_train else self.test_w_quizzes
 219         nb = min(nb, input.size(0))
 220         input[:-nb] = input[nb:].clone()
 221         input[-nb:] = self.problem.generate_seq(nb).to(self.device)
 222
 223     def store_c_quizzes(self, new_c_quizzes, for_train=True):
 224         if for_train:
 225             self.train_c_quizzes.append(new_c_quizzes)
 226         else:
 227             self.test_c_quizzes.append(new_c_quizzes)
 228
 229     def create_c_quizzes(
 230         self,
 231         n_epoch,
 232         result_dir,
 233         logger,
 234         nb,
 235         model,
 236         other_models,
 237         min_ave_seq_logproba,
 238     ):
 239         ###############################################################
 240         # Generate quizzes with model
 241
 242         c_quizzes = torch.empty(
 243             nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
 244         )
 245
 246         ar_mask = torch.full(c_quizzes.size(), 1, device=self.device)
 247         seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
 248
 249         temperature = 1
 250         d_temperature = 1 / 3
 251
 252         while True:
 253             seq_logproba[...] = 0
 254
 255             masked_inplace_autoregression(
 256                 model=model,
 257                 batch_size=self.batch_size,
 258                 input=c_quizzes,
 259                 ar_mask=ar_mask,
 260                 seq_logproba=seq_logproba,
 261                 temperature=temperature,
 262                 deterministic_synthesis=False,
 263                 progress_bar_desc="sampling c_quizzes",
 264                 device=self.device,
 265             )
 266
 267             ave_seq_logproba = seq_logproba.mean()
 268
 269             if min_ave_seq_logproba is None:
 270                 break
 271
 272             # Oh man that's ugly
 273             if ave_seq_logproba < min_ave_seq_logproba:
 274                 if d_temperature > 0:
 275                     d_temperature *= -1 / 3
 276                 temperature += d_temperature
 277             elif ave_seq_logproba > min_ave_seq_logproba * 0.99:
 278                 if d_temperature < 0:
 279                     d_temperature *= -1 / 3
 280                 temperature += d_temperature
 281             else:
 282                 break
 283
 284             logger(f"chaging temperature to {temperature}")
 285
 286         ###############################################################
 287         # Create the reverse quizzes
 288
 289         token_forward, token_backward = self.problem.direction_tokens()
 290
 291         l = (c_quizzes.size(1) - 1) // 2
 292         direction = c_quizzes[:, l : l + 1]
 293         direction = self.problem.token_forward * (
 294             direction == self.problem.token_backward
 295         ) + self.problem.token_backward * (direction == self.problem.token_forward)
 296         reverse_c_quizzes = torch.cat(
 297             [c_quizzes[:, l + 1 :], direction, c_quizzes[:, :l]], dim=1
 298         )
 299
 300         ar_mask = self.make_ar_mask(c_quizzes)
 301         seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
 302
 303         ###############################################################
 304         # Check how many of the other models can solve them in both
 305         # directions
 306
 307         nb_correct = []
 308
 309         for m in other_models:
 310             result = c_quizzes.clone()
 311
 312             masked_inplace_autoregression(
 313                 model=m,
 314                 batch_size=self.batch_size,
 315                 input=result,
 316                 ar_mask=ar_mask,
 317                 seq_logproba=seq_logproba,
 318                 temperature=1.0,
 319                 deterministic_synthesis=True,
 320                 progress_bar_desc="solving c_quizzes",
 321                 device=self.device,
 322             )
 323
 324             correct = (c_quizzes == result).long().min(dim=-1).values
 325
 326             reverse_result = reverse_c_quizzes.clone()
 327
 328             masked_inplace_autoregression(
 329                 model=m,
 330                 batch_size=self.batch_size,
 331                 input=reverse_result,
 332                 ar_mask=ar_mask,
 333                 seq_logproba=seq_logproba,
 334                 temperature=1.0,
 335                 deterministic_synthesis=True,
 336                 progress_bar_desc="solving reversed c_quizzes",
 337                 device=self.device,
 338             )
 339
 340             reverse_correct = (
 341                 (reverse_c_quizzes == reverse_result).long().min(dim=-1).values
 342             )
 343
 344             nb_correct.append((correct * reverse_correct)[None, :])
 345
 346         nb_correct = torch.cat(nb_correct, dim=0).sum(dim=0)
 347
 348         return c_quizzes, nb_correct, seq_logproba.mean()