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 = sky.seq2img(input.to("cpu"), self.height, self.width)
  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.batch_size = batch_size
  98         self.device = device
  99         self.height = 6
 100         self.width = 8
 101
 102         self.train_w_quizzes = sky.generate_seq(
 103             nb_train_samples, height=self.height, width=self.width
 104         ).to(device)
 105
 106         self.test_w_quizzes = sky.generate_seq(
 107             nb_test_samples, height=self.height, width=self.width
 108         ).to(device)
 109
 110         self.nb_codes = max(self.train_w_quizzes.max(), self.test_w_quizzes.max()) + 1
 111
 112         self.train_c_quizzes = []
 113         self.test_c_quizzes = []
 114
 115         if result_dir is not None:
 116             self.save_quizzes(
 117                 self.train_w_quizzes[:72], result_dir, f"culture_w_quizzes", logger
 118             )
 119
 120     def batches(self, split="train", desc=None):
 121         assert split in {"train", "test"}
 122         if split == "train":
 123             w_quizzes = self.train_w_quizzes
 124             c_quizzes = self.train_c_quizzes
 125         else:
 126             w_quizzes = self.test_w_quizzes
 127             c_quizzes = self.test_c_quizzes
 128
 129         if len(c_quizzes) > 0:
 130             c_quizzes = torch.cat(c_quizzes, dim=0)
 131             if c_quizzes.size(0) > w_quizzes.size(0) // 2:
 132                 i = torch.randperm(w_quizzes.size(0))[: w_quizzes.size(0) // 2]
 133                 c_quizzes = c_quizzes[i]
 134
 135             i = torch.randperm(w_quizzes.size(0))[
 136                 : w_quizzes.size(0) - c_quizzes.size(0)
 137             ]
 138             w_quizzes = w_quizzes[i]
 139
 140             self.nb_batch_w_quizzes = w_quizzes.size(0)
 141             self.nb_batch_c_quizzes = c_quizzes.size(0)
 142
 143             input = torch.cat([w_quizzes, c_quizzes], dim=0)
 144         else:
 145             input = w_quizzes
 146             self.nb_batch_w_quizzes = w_quizzes.size(0)
 147             self.nb_batch_c_quizzes = 0
 148
 149         # Shuffle
 150         input = input[torch.randperm(input.size(0))]
 151
 152         if desc is None:
 153             desc = f"epoch-{split}"
 154         for batch in tqdm.tqdm(
 155             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 156         ):
 157             yield batch
 158
 159     def vocabulary_size(self):
 160         return self.nb_codes
 161
 162     def produce_results(
 163         self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
 164     ):
 165         def compute_accuracy(input, logger=None):
 166             input = input[:nmax]
 167             ar_mask = self.make_ar_mask(input)
 168             result = input.clone() * (1 - ar_mask)
 169             seq_logproba = torch.empty(input.size(0), device=self.device)
 170
 171             masked_inplace_autoregression(
 172                 model=model,
 173                 batch_size=self.batch_size,
 174                 input=result,
 175                 ar_mask=ar_mask,
 176                 seq_logproba=seq_logproba,
 177                 temperature=1.0,
 178                 deterministic_synthesis=deterministic_synthesis,
 179                 progress_bar_desc=None,
 180                 device=self.device,
 181             )
 182
 183             nb_total, nb_correct = (
 184                 input.size(0),
 185                 (input == result).long().min(dim=1).values.sum(),
 186             )
 187
 188             return nb_total, nb_correct
 189
 190         train_nb_total, train_nb_correct = compute_accuracy(self.train_w_quizzes)
 191
 192         logger(
 193             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}%"
 194         )
 195
 196         test_nb_total, test_nb_correct = compute_accuracy(self.test_w_quizzes, logger)
 197
 198         logger(
 199             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}%"
 200         )
 201
 202         main_test_accuracy = test_nb_correct / test_nb_total
 203         logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
 204
 205         ##############################
 206
 207         input = self.test_w_quizzes[:96]
 208         ar_mask = self.make_ar_mask(input)
 209         result = input.clone() * (1 - ar_mask)
 210         seq_logproba = torch.empty(input.size(0), device=self.device)
 211
 212         masked_inplace_autoregression(
 213             model=model,
 214             batch_size=self.batch_size,
 215             input=result,
 216             ar_mask=ar_mask,
 217             seq_logproba=seq_logproba,
 218             temperature=1.0,
 219             deterministic_synthesis=deterministic_synthesis,
 220             progress_bar_desc=None,
 221             device=self.device,
 222         )
 223
 224         self.save_quizzes(
 225             result[:72],
 226             result_dir,
 227             f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
 228             logger,
 229         )
 230
 231         return main_test_accuracy
 232
 233     def renew_w_quizzes(self, nb, for_train=True):
 234         input = self.train_w_quizzes if for_train else self.test_w_quizzes
 235         nb = min(nb, input.size(0))
 236         input[:-nb] = input[nb:].clone()
 237         input[-nb:] = sky.generate_seq(nb, height=self.height, width=self.width).to(
 238             self.device
 239         )
 240
 241     def store_c_quizzes(self, new_c_quizzes, for_train=True):
 242         if for_train:
 243             self.train_c_quizzes.append(new_c_quizzes)
 244         else:
 245             self.test_c_quizzes.append(new_c_quizzes)
 246
 247     def create_c_quizzes(
 248         self,
 249         n_epoch,
 250         result_dir,
 251         logger,
 252         nb,
 253         model,
 254         other_models,
 255         min_ave_seq_logproba,
 256     ):
 257         ###############################################################
 258         # Generate quizzes with model
 259
 260         c_quizzes = torch.empty(
 261             nb, self.height * self.width * 2 + 1, device=self.device, dtype=torch.int64
 262         )
 263
 264         ar_mask = torch.full(c_quizzes.size(), 1, device=self.device)
 265         seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
 266
 267         temperature = 1
 268         d_temperature = 1 / 3
 269
 270         while True:
 271             seq_logproba[...] = 0
 272
 273             masked_inplace_autoregression(
 274                 model=model,
 275                 batch_size=self.batch_size,
 276                 input=c_quizzes,
 277                 ar_mask=ar_mask,
 278                 seq_logproba=seq_logproba,
 279                 temperature=temperature,
 280                 deterministic_synthesis=False,
 281                 progress_bar_desc="sampling c_quizzes",
 282                 device=self.device,
 283             )
 284
 285             ave_seq_logproba = seq_logproba.mean()
 286
 287             logger(f"{ave_seq_logproba=} {min_ave_seq_logproba=}")
 288
 289             if min_ave_seq_logproba is None:
 290                 break
 291
 292             # Oh man that's ugly
 293             if ave_seq_logproba < min_ave_seq_logproba * 1.1:
 294                 if d_temperature > 0:
 295                     d_temperature *= -1 / 3
 296                 temperature += d_temperature
 297             elif ave_seq_logproba > min_ave_seq_logproba:
 298                 if d_temperature < 0:
 299                     d_temperature *= -1 / 3
 300                 temperature += d_temperature
 301             else:
 302                 break
 303
 304             logger(f"chaging temperature to {temperature}")
 305
 306         ###############################################################
 307         # Create the reverse quizzes
 308
 309         l = self.height * self.width
 310         direction = c_quizzes[:, l : l + 1]
 311         direction = sky.token_forward * (
 312             direction == sky.token_backward
 313         ) + sky.token_backward * (direction == sky.token_forward)
 314         reverse_c_quizzes = torch.cat(
 315             [c_quizzes[:, l + 1 :], direction, c_quizzes[:, :l]], dim=1
 316         )
 317
 318         ar_mask = self.make_ar_mask(c_quizzes)
 319         seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
 320
 321         ###############################################################
 322         # Check how many of the other models can solve them in both
 323         # directions
 324
 325         nb_correct = []
 326
 327         for m in other_models:
 328             result = c_quizzes.clone()
 329
 330             masked_inplace_autoregression(
 331                 model=m,
 332                 batch_size=self.batch_size,
 333                 input=result,
 334                 ar_mask=ar_mask,
 335                 seq_logproba=seq_logproba,
 336                 temperature=1.0,
 337                 deterministic_synthesis=True,
 338                 progress_bar_desc="solving c_quizzes",
 339                 device=self.device,
 340             )
 341
 342             correct = (c_quizzes == result).long().min(dim=-1).values
 343
 344             reverse_result = reverse_c_quizzes.clone()
 345
 346             masked_inplace_autoregression(
 347                 model=m,
 348                 batch_size=self.batch_size,
 349                 input=reverse_result,
 350                 ar_mask=ar_mask,
 351                 seq_logproba=seq_logproba,
 352                 temperature=1.0,
 353                 deterministic_synthesis=True,
 354                 progress_bar_desc="solving reversed c_quizzes",
 355                 device=self.device,
 356             )
 357
 358             reverse_correct = (
 359                 (reverse_c_quizzes == reverse_result).long().min(dim=-1).values
 360             )
 361
 362             nb_correct.append((correct * reverse_correct)[None, :])
 363
 364         nb_correct = torch.cat(nb_correct, dim=0).sum(dim=0)
 365
 366         return c_quizzes, nb_correct, seq_logproba.mean()