main.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, sys, argparse, time, tqdm, os, datetime, warnings
   9
  10 import torch, torchvision
  11 from torch import nn
  12 from torch.nn import functional as F
  13
  14 import ffutils
  15 import mygpt, quizz_machine
  16
  17 # world quizzes vs. culture quizzes
  18
  19 ######################################################################
  20
  21 accuracy_to_make_c_quizzes = 0.975
  22 nb_new_c_quizzes_for_train = 1000
  23 nb_new_c_quizzes_for_test = 100
  24
  25 ######################################################################
  26
  27 if torch.cuda.is_available():
  28     device = torch.device("cuda")
  29     torch.backends.cuda.matmul.allow_tf32 = True
  30 else:
  31     device = torch.device("cpu")
  32
  33 ######################################################################
  34
  35 parser = argparse.ArgumentParser(
  36     description="An implementation of GPT with cache.",
  37     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
  38 )
  39
  40 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
  41
  42 parser.add_argument("--result_dir", type=str, default=None)
  43
  44 parser.add_argument("--seed", type=int, default=0)
  45
  46 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
  47
  48 ########################################
  49
  50 parser.add_argument("--nb_epochs", type=int, default=10000)
  51
  52 parser.add_argument("--batch_size", type=int, default=None)
  53
  54 parser.add_argument("--physical_batch_size", type=int, default=None)
  55
  56 parser.add_argument("--nb_train_samples", type=int, default=None)
  57
  58 parser.add_argument("--nb_test_samples", type=int, default=None)
  59
  60 parser.add_argument("--learning_rate", type=float, default=1e-4)
  61
  62 ########################################
  63
  64 parser.add_argument("--model", type=str, default=None)
  65
  66 parser.add_argument("--dim_model", type=int, default=None)
  67
  68 parser.add_argument("--dim_keys", type=int, default=None)
  69
  70 parser.add_argument("--dim_hidden", type=int, default=None)
  71
  72 parser.add_argument("--nb_heads", type=int, default=None)
  73
  74 parser.add_argument("--nb_blocks", type=int, default=None)
  75
  76 parser.add_argument("--dropout", type=float, default=0.1)
  77
  78 ########################################
  79
  80 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  81
  82 parser.add_argument("--nb_gpts", type=int, default=5)
  83
  84 parser.add_argument("--dirty_debug", action="store_true", default=False)
  85
  86 ######################################################################
  87
  88 args = parser.parse_args()
  89
  90 if args.result_dir is None:
  91     args.result_dir = f"results_culture"
  92
  93 ######################################################################
  94
  95 if args.dirty_debug:
  96     accuracy_to_make_c_quizzes = 0.0
  97     nb_new_c_quizzes_for_train = 100
  98     nb_new_c_quizzes_for_test = 10
  99
 100 ######################################################################
 101
 102 default_args = {
 103     "model": "37M",
 104     "batch_size": 100,
 105     "nb_train_samples": 250000,
 106     "nb_test_samples": 10000,
 107 }
 108
 109 for k, v in default_args.items():
 110     if getattr(args, k) is None:
 111         setattr(args, k, v)
 112
 113 ######################################################################
 114
 115 default_model_args = {
 116     "17K": {
 117         "dim_model": 32,
 118         "dim_keys": 32,
 119         "dim_hidden": 32,
 120         "nb_heads": 2,
 121         "nb_blocks": 2,
 122     },
 123     "4M": {
 124         "dim_model": 256,
 125         "dim_keys": 32,
 126         "dim_hidden": 1024,
 127         "nb_heads": 4,
 128         "nb_blocks": 6,
 129     },
 130     "37M": {
 131         "dim_model": 512,
 132         "dim_keys": 64,
 133         "dim_hidden": 2048,
 134         "nb_heads": 8,
 135         "nb_blocks": 12,
 136     },
 137     "122M": {
 138         "dim_model": 768,
 139         "dim_keys": 64,
 140         "dim_hidden": 2048,
 141         "nb_heads": 8,
 142         "nb_blocks": 24,
 143     },
 144     "352M": {
 145         "dim_model": 1024,
 146         "dim_keys": 64,
 147         "dim_hidden": 2048,
 148         "nb_heads": 8,
 149         "nb_blocks": 48,
 150     },
 151 }
 152
 153 if args.model in default_model_args:
 154     for k, v in default_model_args[args.model].items():
 155         if getattr(args, k) is None:
 156             setattr(args, k, v)
 157 else:
 158     raise ValueError(f"Unknown model {args.model}")
 159
 160 ######################################################################
 161
 162 try:
 163     os.mkdir(args.result_dir)
 164 except FileExistsError:
 165     print(f"result directory {args.result_dir} already exists")
 166     exit(1)
 167
 168 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 169
 170 if args.seed >= 0:
 171     # torch.backends.cudnn.deterministic = True
 172     # torch.backends.cudnn.benchmark = False
 173     # torch.use_deterministic_algorithms(True)
 174     torch.manual_seed(args.seed)
 175     if torch.cuda.is_available():
 176         torch.cuda.manual_seed_all(args.seed)
 177
 178 ######################################################################
 179
 180
 181 def log_string(s):
 182     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 183
 184     if log_file is not None:
 185         log_file.write(t + s + "\n")
 186         log_file.flush()
 187
 188     print(t + s)
 189     sys.stdout.flush()
 190
 191
 192 log_string(f"argv {' '.join(sys.argv)}")
 193
 194 for n in vars(args):
 195     log_string(f"args.{n} {getattr(args, n)}")
 196
 197
 198 ######################################################################
 199
 200 if args.dirty_debug:
 201     args.nb_train_samples = 2500
 202     args.nb_test_samples = 100
 203
 204 if args.physical_batch_size is None:
 205     args.physical_batch_size = args.batch_size
 206 else:
 207     assert args.batch_size % args.physical_batch_size == 0
 208
 209 assert args.nb_train_samples % args.batch_size == 0
 210 assert args.nb_test_samples % args.batch_size == 0
 211
 212 quizz_machine = quizz_machine.QuizzMachine(
 213     nb_train_samples=args.nb_train_samples,
 214     nb_test_samples=args.nb_test_samples,
 215     batch_size=args.physical_batch_size,
 216     result_dir=args.result_dir,
 217     logger=log_string,
 218     device=device,
 219 )
 220
 221 ######################################################################
 222
 223 log_string(f"device {device}")
 224
 225 vocabulary_size = quizz_machine.vocabulary_size()
 226
 227 log_string(f"vocabulary_size {vocabulary_size}")
 228
 229 ######################################################################
 230
 231 # Compute the entropy of the training tokens
 232
 233 token_count = 0
 234 for input in quizz_machine.batches(split="train", desc="train-entropy"):
 235     token_count += F.one_hot(input, num_classes=quizz_machine.vocabulary_size()).sum(
 236         (0, 1)
 237     )
 238 token_probas = token_count / token_count.sum()
 239 entropy = -torch.xlogy(token_probas, token_probas).sum()
 240 train_set_perplexity = math.exp(entropy)
 241
 242 ######################################################################
 243 # A bit of paranoia never hurts
 244
 245 if args.max_percents_of_test_in_train >= 0:
 246
 247     def subsets_as_tuples(batches, cs):
 248         s = set()
 249         for batch in batches:
 250             for x in batch:
 251                 s.add(tuple([v.item() for v in x]))
 252                 if len(s) == cs:
 253                     yield s
 254                     s = set()
 255         yield s
 256
 257     nb_test, nb_in_train = 0, 0
 258     for test_subset in subsets_as_tuples(
 259         quizz_machine.batches(split="test", desc="test-check"), 25000
 260     ):
 261         in_train = set()
 262         for train_subset in subsets_as_tuples(
 263             quizz_machine.batches(split="train", desc="train-check"), 25000
 264         ):
 265             in_train.update(test_subset.intersection(train_subset))
 266         nb_in_train += len(in_train)
 267         nb_test += len(test_subset)
 268
 269     log_string(
 270         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 271     )
 272
 273     assert (
 274         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 275     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 276
 277 ##############################
 278
 279
 280 def one_epoch(model, quizz_machine):
 281     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 282
 283     model.train()
 284
 285     nb_train_samples, acc_train_loss = 0, 0.0
 286
 287     for input in quizz_machine.batches(split="train"):
 288         input = input.to(device)
 289
 290         if nb_train_samples % args.batch_size == 0:
 291             optimizer.zero_grad()
 292
 293         output = model(mygpt.BracketedSequence(input)).x
 294         loss = F.cross_entropy(output.transpose(1, 2), input)
 295         acc_train_loss += loss.item() * input.size(0)
 296
 297         nb_train_samples += input.size(0)
 298
 299         loss.backward()
 300
 301         if nb_train_samples % args.batch_size == 0:
 302             optimizer.step()
 303
 304     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 305
 306     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 307
 308
 309 ######################################################################
 310
 311
 312 def run_tests(model, quizz_machine, deterministic_synthesis):
 313     with torch.autograd.no_grad():
 314         model.eval()
 315
 316         nb_test_samples, acc_test_loss = 0, 0.0
 317         nb_samples_accumulated = 0
 318
 319         for input in quizz_machine.batches(split="test"):
 320             input = input.to(device)
 321
 322             bs = model(mygpt.BracketedSequence(input))
 323             output = bs.x
 324
 325             loss = F.cross_entropy(output.transpose(1, 2), input)
 326
 327             acc_test_loss += loss.item() * input.size(0)
 328
 329             nb_test_samples += input.size(0)
 330
 331         main_test_accuracy = quizz_machine.produce_results(
 332             n_epoch=n_epoch,
 333             model=model,
 334             result_dir=args.result_dir,
 335             logger=log_string,
 336             deterministic_synthesis=deterministic_synthesis,
 337         )
 338
 339         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 340
 341         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 342
 343     model.main_test_accuracy = main_test_accuracy
 344
 345
 346 ######################################################################
 347
 348
 349 def create_c_quizzes(
 350     model,
 351     other_models,
 352     quizz_machine,
 353     nb_for_train=1000,
 354     nb_for_test=100,
 355     min_ave_seq_logproba=None,
 356 ):
 357     kept = []
 358
 359     sum_logits, sum_nb_c_quizzes = 0, 0
 360
 361     while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
 362         nb_to_generate = 4 * (nb_for_train + nb_for_test)
 363
 364         new_c_quizzes, nb_correct, ave_seq_logproba = quizz_machine.create_c_quizzes(
 365             n_epoch=n_epoch,
 366             result_dir=args.result_dir,
 367             logger=log_string,
 368             nb=nb_to_generate,
 369             model=model,
 370             other_models=other_models,
 371             min_ave_seq_logproba=min_ave_seq_logproba,
 372         )
 373
 374         sum_logits += new_c_quizzes.size(0) * ave_seq_logproba
 375         sum_nb_c_quizzes += new_c_quizzes.size(0)
 376
 377         to_keep = new_c_quizzes[nb_correct == len(other_models) - 1]
 378
 379         if args.dirty_debug:
 380             to_keep = new_c_quizzes
 381
 382         log_string(
 383             f"keep {to_keep.size(0)}/{new_c_quizzes.size(0)} c_quizzes ({to_keep.size(0)*100/new_c_quizzes.size(0):.02f}%)"
 384         )
 385
 386         kept.append(to_keep)
 387
 388     new_c_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
 389
 390     quizz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 391     quizz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 392
 393     quizz_machine.save_quizzes(
 394         new_c_quizzes[:72],
 395         args.result_dir,
 396         f"culture_c_quiz_{n_epoch:04d}_{model.id:02d}",
 397         log_string,
 398     )
 399
 400     return sum_logits / sum_nb_c_quizzes
 401
 402
 403 ######################################################################
 404
 405 models = []
 406
 407 for k in range(args.nb_gpts):
 408     model = mygpt.MyGPT(
 409         vocabulary_size=vocabulary_size,
 410         dim_model=args.dim_model,
 411         dim_keys=args.dim_keys,
 412         dim_hidden=args.dim_hidden,
 413         nb_heads=args.nb_heads,
 414         nb_blocks=args.nb_blocks,
 415         causal=True,
 416         dropout=args.dropout,
 417     ).to(device)
 418
 419     model.main_test_accuracy = 0.0
 420     model.id = k
 421
 422     models.append(model)
 423
 424
 425 nb_parameters = sum(p.numel() for p in models[0].parameters())
 426 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 427
 428 ######################################################################
 429
 430 min_ave_seq_logproba = None
 431
 432 for n_epoch in range(args.nb_epochs):
 433     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 434
 435     a = [(model.id, float(model.main_test_accuracy)) for model in models]
 436     a.sort(key=lambda p: p[0])
 437     log_string(f"current accuracies {a}")
 438
 439     # select the model with lowest accuracy
 440     models.sort(key=lambda model: model.main_test_accuracy)
 441     model = models[0]
 442
 443     log_string(
 444         f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
 445     )
 446
 447     # improve it
 448     one_epoch(model, quizz_machine)
 449
 450     quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
 451
 452     log_string(
 453         f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
 454     )
 455
 456     # test it
 457     run_tests(model, quizz_machine, deterministic_synthesis=False)
 458
 459     log_string(
 460         f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
 461     )
 462
 463     if min([m.main_test_accuracy for m in models]) >= accuracy_to_make_c_quizzes:
 464         other_models = models.copy()
 465         other_models.remove(model)
 466
 467         ave_seq_logproba = create_c_quizzes(
 468             model,
 469             other_models,
 470             quizz_machine,
 471             nb_for_train=nb_new_c_quizzes_for_train,
 472             nb_for_test=nb_new_c_quizzes_for_test,
 473             min_ave_seq_logproba=min_ave_seq_logproba,
 474         )
 475
 476         # We keep the first average logits as a reference
 477         if min_ave_seq_logproba is None:
 478             min_ave_seq_logproba = ave_seq_logproba
 479         else:
 480             log_string(
 481                 f"min_ave_seq_logproba {min_ave_seq_logproba} ave_seq_logproba {ave_seq_logproba}"
 482             )
 483
 484         # We update everyone
 485         for model in models:
 486             run_tests(model, quizz_machine, deterministic_synthesis=False)
 487
 488
 489 ######################################################################