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 import sky
  70
  71
  72 class QuizzMachine:
  73     def save_image(self, input, result_dir, filename, logger):
  74         img = self.sky.seq2img(input.to("cpu"))
  75         image_name = os.path.join(result_dir, filename)
  76         torchvision.utils.save_image(img.float() / 255.0, image_name, nrow=6, padding=4)
  77         logger(f"wrote {image_name}")
  78
  79     def save_quizzes(self, input, result_dir, filename_prefix, logger):
  80         self.save_image(input, result_dir, filename_prefix + ".png", logger)
  81
  82     def make_ar_mask(self, input):
  83         b = torch.arange(input.size(1), device=input.device) > input.size(1) // 2
  84         return b.long()[None, :].expand_as(input)
  85
  86     def __init__(
  87         self,
  88         nb_train_samples,
  89         nb_test_samples,
  90         batch_size,
  91         result_dir=None,
  92         logger=None,
  93         device=torch.device("cpu"),
  94     ):
  95         super().__init__()
  96
  97         self.sky = sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2)
  98         self.batch_size = batch_size
  99         self.device = device
 100
 101         self.train_w_quizzes = self.sky.generate_seq(nb_train_samples).to(device)
 102         self.test_w_quizzes = self.sky.generate_seq(nb_test_samples).to(device)
 103
 104         self.nb_codes = max(self.train_w_quizzes.max(), self.test_w_quizzes.max()) + 1
 105
 106         self.train_c_quizzes = []
 107         self.test_c_quizzes = []
 108
 109         if result_dir is not None:
 110             self.save_quizzes(
 111                 self.train_w_quizzes[:72], result_dir, f"culture_w_quizzes", logger
 112             )
 113
 114     def batches(self, split="train", desc=None):
 115         assert split in {"train", "test"}
 116         if split == "train":
 117             w_quizzes = self.train_w_quizzes
 118             c_quizzes = self.train_c_quizzes
 119         else:
 120             w_quizzes = self.test_w_quizzes
 121             c_quizzes = self.test_c_quizzes
 122
 123         if len(c_quizzes) > 0:
 124             c_quizzes = torch.cat(c_quizzes, dim=0)
 125             if c_quizzes.size(0) > w_quizzes.size(0) // 2:
 126                 i = torch.randperm(w_quizzes.size(0))[: w_quizzes.size(0) // 2]
 127                 c_quizzes = c_quizzes[i]
 128
 129             i = torch.randperm(w_quizzes.size(0))[
 130                 : w_quizzes.size(0) - c_quizzes.size(0)
 131             ]
 132             w_quizzes = w_quizzes[i]
 133
 134             self.nb_batch_w_quizzes = w_quizzes.size(0)
 135             self.nb_batch_c_quizzes = c_quizzes.size(0)
 136
 137             input = torch.cat([w_quizzes, c_quizzes], dim=0)
 138         else:
 139             input = w_quizzes
 140             self.nb_batch_w_quizzes = w_quizzes.size(0)
 141             self.nb_batch_c_quizzes = 0
 142
 143         # Shuffle
 144         input = input[torch.randperm(input.size(0))]
 145
 146         if desc is None:
 147             desc = f"epoch-{split}"
 148         for batch in tqdm.tqdm(
 149             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 150         ):
 151             yield batch
 152
 153     def vocabulary_size(self):
 154         return self.nb_codes
 155
 156     def produce_results(
 157         self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
 158     ):
 159         def compute_accuracy(input, logger=None):
 160             input = input[:nmax]
 161             ar_mask = self.make_ar_mask(input)
 162             result = input.clone() * (1 - ar_mask)
 163             seq_logproba = torch.empty(input.size(0), device=self.device)
 164
 165             masked_inplace_autoregression(
 166                 model=model,
 167                 batch_size=self.batch_size,
 168                 input=result,
 169                 ar_mask=ar_mask,
 170                 seq_logproba=seq_logproba,
 171                 temperature=1.0,
 172                 deterministic_synthesis=deterministic_synthesis,
 173                 progress_bar_desc=None,
 174                 device=self.device,
 175             )
 176
 177             nb_total, nb_correct = (
 178                 input.size(0),
 179                 (input == result).long().min(dim=1).values.sum(),
 180             )
 181
 182             return nb_total, nb_correct
 183
 184         train_nb_total, train_nb_correct = compute_accuracy(self.train_w_quizzes)
 185
 186         logger(
 187             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}%"
 188         )
 189
 190         test_nb_total, test_nb_correct = compute_accuracy(self.test_w_quizzes, logger)
 191
 192         logger(
 193             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}%"
 194         )
 195
 196         main_test_accuracy = test_nb_correct / test_nb_total
 197         logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
 198
 199         ##############################
 200
 201         input = self.test_w_quizzes[:96]
 202         ar_mask = self.make_ar_mask(input)
 203         result = input.clone() * (1 - ar_mask)
 204         seq_logproba = torch.empty(input.size(0), device=self.device)
 205
 206         masked_inplace_autoregression(
 207             model=model,
 208             batch_size=self.batch_size,
 209             input=result,
 210             ar_mask=ar_mask,
 211             seq_logproba=seq_logproba,
 212             temperature=1.0,
 213             deterministic_synthesis=deterministic_synthesis,
 214             progress_bar_desc=None,
 215             device=self.device,
 216         )
 217
 218         self.save_quizzes(
 219             result[:72],
 220             result_dir,
 221             f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
 222             logger,
 223         )
 224
 225         return main_test_accuracy
 226
 227     def renew_w_quizzes(self, nb, for_train=True):
 228         input = self.train_w_quizzes if for_train else self.test_w_quizzes
 229         nb = min(nb, input.size(0))
 230         input[:-nb] = input[nb:].clone()
 231         input[-nb:] = self.sky.generate_seq(nb).to(self.device)
 232
 233     def store_c_quizzes(self, new_c_quizzes, for_train=True):
 234         if for_train:
 235             self.train_c_quizzes.append(new_c_quizzes)
 236         else:
 237             self.test_c_quizzes.append(new_c_quizzes)
 238
 239     def create_c_quizzes(
 240         self,
 241         n_epoch,
 242         result_dir,
 243         logger,
 244         nb,
 245         model,
 246         other_models,
 247         min_ave_seq_logproba,
 248     ):
 249         ###############################################################
 250         # Generate quizzes with model
 251
 252         c_quizzes = torch.empty(
 253             nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
 254         )
 255
 256         ar_mask = torch.full(c_quizzes.size(), 1, device=self.device)
 257         seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
 258
 259         temperature = 1
 260         d_temperature = 1 / 3
 261
 262         while True:
 263             seq_logproba[...] = 0
 264
 265             masked_inplace_autoregression(
 266                 model=model,
 267                 batch_size=self.batch_size,
 268                 input=c_quizzes,
 269                 ar_mask=ar_mask,
 270                 seq_logproba=seq_logproba,
 271                 temperature=temperature,
 272                 deterministic_synthesis=False,
 273                 progress_bar_desc="sampling c_quizzes",
 274                 device=self.device,
 275             )
 276
 277             ave_seq_logproba = seq_logproba.mean()
 278
 279             logger(f"{ave_seq_logproba=} {min_ave_seq_logproba=}")
 280
 281             if min_ave_seq_logproba is None:
 282                 break
 283
 284             # Oh man that's ugly
 285             if ave_seq_logproba < min_ave_seq_logproba * 1.1:
 286                 if d_temperature > 0:
 287                     d_temperature *= -1 / 3
 288                 temperature += d_temperature
 289             elif ave_seq_logproba > min_ave_seq_logproba:
 290                 if d_temperature < 0:
 291                     d_temperature *= -1 / 3
 292                 temperature += d_temperature
 293             else:
 294                 break
 295
 296             logger(f"chaging temperature to {temperature}")
 297
 298         ###############################################################
 299         # Create the reverse quizzes
 300
 301         l = (c_quizzes.size(1) - 1) // 2
 302         direction = c_quizzes[:, l : l + 1]
 303         direction = self.sky.token_forward * (
 304             direction == self.sky.token_backward
 305         ) + self.sky.token_backward * (direction == self.sky.token_forward)
 306         reverse_c_quizzes = torch.cat(
 307             [c_quizzes[:, l + 1 :], direction, c_quizzes[:, :l]], dim=1
 308         )
 309
 310         ar_mask = self.make_ar_mask(c_quizzes)
 311         seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
 312
 313         ###############################################################
 314         # Check how many of the other models can solve them in both
 315         # directions
 316
 317         nb_correct = []
 318
 319         for m in other_models:
 320             result = c_quizzes.clone()
 321
 322             masked_inplace_autoregression(
 323                 model=m,
 324                 batch_size=self.batch_size,
 325                 input=result,
 326                 ar_mask=ar_mask,
 327                 seq_logproba=seq_logproba,
 328                 temperature=1.0,
 329                 deterministic_synthesis=True,
 330                 progress_bar_desc="solving c_quizzes",
 331                 device=self.device,
 332             )
 333
 334             correct = (c_quizzes == result).long().min(dim=-1).values
 335
 336             reverse_result = reverse_c_quizzes.clone()
 337
 338             masked_inplace_autoregression(
 339                 model=m,
 340                 batch_size=self.batch_size,
 341                 input=reverse_result,
 342                 ar_mask=ar_mask,
 343                 seq_logproba=seq_logproba,
 344                 temperature=1.0,
 345                 deterministic_synthesis=True,
 346                 progress_bar_desc="solving reversed c_quizzes",
 347                 device=self.device,
 348             )
 349
 350             reverse_correct = (
 351                 (reverse_c_quizzes == reverse_result).long().min(dim=-1).values
 352             )
 353
 354             nb_correct.append((correct * reverse_correct)[None, :])
 355
 356         nb_correct = torch.cat(nb_correct, dim=0).sum(dim=0)
 357
 358         return c_quizzes, nb_correct, seq_logproba.mean()