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
  16 import sky, wireworld, quizz_machine
  17
  18 # world quizzes vs. culture quizzes
  19
  20 ######################################################################
  21
  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")
  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-3)
  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("--both_directions", action="store_true", default=False)
  83
  84 parser.add_argument("--problem", type=str, default="sky")
  85
  86 parser.add_argument("--nb_gpts", type=int, default=5)
  87
  88 parser.add_argument("--min_to_validate", type=int, default=4)
  89
  90 parser.add_argument("--max_to_validate", type=int, default=4)
  91
  92 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
  93
  94 parser.add_argument("--dirty_debug", action="store_true", default=False)
  95
  96 parser.add_argument("--generation_temperature", type=float, default=1.0)
  97
  98 parser.add_argument("--stochastic_validation", action="store_true", default=False)
  99
 100 ######################################################################
 101
 102 parser.add_argument("--sky_height", type=int, default=6)
 103
 104 parser.add_argument("--sky_width", type=int, default=8)
 105
 106 parser.add_argument("--sky_nb_birds", type=int, default=3)
 107
 108 parser.add_argument("--sky_nb_iterations", type=int, default=2)
 109
 110 parser.add_argument("--sky_speed", type=int, default=3)
 111
 112 ######################################################################
 113
 114 args = parser.parse_args()
 115
 116 if args.result_dir is None:
 117     args.result_dir = f"results_culture"
 118
 119 ######################################################################
 120
 121 if args.dirty_debug:
 122     args.accuracy_to_make_c_quizzes = 0.0
 123     nb_new_c_quizzes_for_train = 100
 124     nb_new_c_quizzes_for_test = 10
 125
 126 ######################################################################
 127
 128 default_args = {
 129     "model": "37M",
 130     "batch_size": 100,
 131     "nb_train_samples": 100000,
 132     "nb_test_samples": 10000,
 133 }
 134
 135 for k, v in default_args.items():
 136     if getattr(args, k) is None:
 137         setattr(args, k, v)
 138
 139 ######################################################################
 140
 141 default_model_args = {
 142     "17K": {
 143         "dim_model": 32,
 144         "dim_keys": 32,
 145         "dim_hidden": 32,
 146         "nb_heads": 2,
 147         "nb_blocks": 2,
 148     },
 149     "4M": {
 150         "dim_model": 256,
 151         "dim_keys": 32,
 152         "dim_hidden": 1024,
 153         "nb_heads": 4,
 154         "nb_blocks": 6,
 155     },
 156     "37M": {
 157         "dim_model": 512,
 158         "dim_keys": 64,
 159         "dim_hidden": 2048,
 160         "nb_heads": 8,
 161         "nb_blocks": 12,
 162     },
 163     "122M": {
 164         "dim_model": 768,
 165         "dim_keys": 64,
 166         "dim_hidden": 2048,
 167         "nb_heads": 8,
 168         "nb_blocks": 24,
 169     },
 170     "352M": {
 171         "dim_model": 1024,
 172         "dim_keys": 64,
 173         "dim_hidden": 2048,
 174         "nb_heads": 8,
 175         "nb_blocks": 48,
 176     },
 177 }
 178
 179 if args.model in default_model_args:
 180     for k, v in default_model_args[args.model].items():
 181         if getattr(args, k) is None:
 182             setattr(args, k, v)
 183 else:
 184     raise ValueError(f"Unknown model {args.model}")
 185
 186 ######################################################################
 187
 188 try:
 189     os.mkdir(args.result_dir)
 190 except FileExistsError:
 191     print(f"result directory {args.result_dir} already exists")
 192     exit(1)
 193
 194 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 195
 196 if args.seed >= 0:
 197     # torch.backends.cudnn.deterministic = True
 198     # torch.backends.cudnn.benchmark = False
 199     # torch.use_deterministic_algorithms(True)
 200     torch.manual_seed(args.seed)
 201     if torch.cuda.is_available():
 202         torch.cuda.manual_seed_all(args.seed)
 203
 204 ######################################################################
 205
 206
 207 def log_string(s):
 208     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 209
 210     if log_file is not None:
 211         log_file.write(t + s + "\n")
 212         log_file.flush()
 213
 214     print(t + s)
 215     sys.stdout.flush()
 216
 217
 218 log_string(f"argv {' '.join(sys.argv)}")
 219
 220 for n in vars(args):
 221     log_string(f"args.{n} {getattr(args, n)}")
 222
 223
 224 ######################################################################
 225
 226 if args.dirty_debug:
 227     args.nb_train_samples = 2500
 228     args.nb_test_samples = 100
 229
 230 if args.physical_batch_size is None:
 231     args.physical_batch_size = args.batch_size
 232 else:
 233     assert args.batch_size % args.physical_batch_size == 0
 234
 235 assert args.nb_train_samples % args.batch_size == 0
 236 assert args.nb_test_samples % args.batch_size == 0
 237
 238 if args.problem == "sky":
 239     problem = sky.Sky(
 240         height=args.sky_height,
 241         width=args.sky_width,
 242         nb_birds=args.sky_nb_birds,
 243         nb_iterations=args.sky_nb_iterations,
 244         speed=args.sky_speed,
 245     )
 246 elif args.problem == "wireworld":
 247     problem = wireworld.Wireworld(height=8, width=10, nb_iterations=2, speed=5)
 248 else:
 249     raise ValueError
 250
 251 quizz_machine = quizz_machine.QuizzMachine(
 252     problem=problem,
 253     nb_train_samples=args.nb_train_samples,
 254     nb_test_samples=args.nb_test_samples,
 255     batch_size=args.physical_batch_size,
 256     result_dir=args.result_dir,
 257     logger=log_string,
 258     device=device,
 259 )
 260
 261 ######################################################################
 262
 263 log_string(f"device {device}")
 264
 265 vocabulary_size = quizz_machine.vocabulary_size()
 266
 267 log_string(f"vocabulary_size {vocabulary_size}")
 268
 269 ######################################################################
 270
 271 # Compute the entropy of the training tokens
 272
 273 token_count = 0
 274 for input in quizz_machine.batches(split="train", desc="train-entropy"):
 275     token_count += F.one_hot(input, num_classes=quizz_machine.vocabulary_size()).sum(
 276         (0, 1)
 277     )
 278 token_probas = token_count / token_count.sum()
 279 entropy = -torch.xlogy(token_probas, token_probas).sum()
 280 train_set_perplexity = math.exp(entropy)
 281
 282 ######################################################################
 283 # A bit of paranoia never hurts
 284
 285 if args.max_percents_of_test_in_train >= 0:
 286
 287     def subsets_as_tuples(batches, cs):
 288         s = set()
 289         for batch in batches:
 290             for x in batch:
 291                 s.add(tuple([v.item() for v in x]))
 292                 if len(s) == cs:
 293                     yield s
 294                     s = set()
 295         yield s
 296
 297     nb_test, nb_in_train = 0, 0
 298     for test_subset in subsets_as_tuples(
 299         quizz_machine.batches(split="test", desc="test-check"), 25000
 300     ):
 301         in_train = set()
 302         for train_subset in subsets_as_tuples(
 303             quizz_machine.batches(split="train", desc="train-check"), 25000
 304         ):
 305             in_train.update(test_subset.intersection(train_subset))
 306         nb_in_train += len(in_train)
 307         nb_test += len(test_subset)
 308
 309     log_string(
 310         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 311     )
 312
 313     assert (
 314         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 315     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 316
 317 ##############################
 318
 319
 320 def one_epoch(model, quizz_machine):
 321     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 322
 323     model.train()
 324
 325     nb_train_samples, acc_train_loss = 0, 0.0
 326
 327     for input in quizz_machine.batches(split="train"):
 328         input = input.to(device)
 329
 330         if nb_train_samples % args.batch_size == 0:
 331             optimizer.zero_grad()
 332
 333         output = model(mygpt.BracketedSequence(input)).x
 334         loss = F.cross_entropy(output.transpose(1, 2), input)
 335         acc_train_loss += loss.item() * input.size(0)
 336
 337         nb_train_samples += input.size(0)
 338
 339         loss.backward()
 340
 341         if nb_train_samples % args.batch_size == 0:
 342             optimizer.step()
 343
 344     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 345
 346     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 347
 348
 349 ######################################################################
 350
 351
 352 def run_tests(model, quizz_machine, deterministic_synthesis):
 353     with torch.autograd.no_grad():
 354         model.eval()
 355
 356         nb_test_samples, acc_test_loss = 0, 0.0
 357         nb_samples_accumulated = 0
 358
 359         for input in quizz_machine.batches(split="test"):
 360             input = input.to(device)
 361
 362             bs = model(mygpt.BracketedSequence(input))
 363             output = bs.x
 364
 365             loss = F.cross_entropy(output.transpose(1, 2), input)
 366
 367             acc_test_loss += loss.item() * input.size(0)
 368
 369             nb_test_samples += input.size(0)
 370
 371         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 372
 373         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 374
 375         model.main_test_accuracy = quizz_machine.produce_results(
 376             n_epoch=n_epoch,
 377             model=model,
 378             result_dir=args.result_dir,
 379             deterministic_synthesis=deterministic_synthesis,
 380         )
 381
 382
 383 ######################################################################
 384
 385
 386 def valid_c_quizzes(recorded, criteria):
 387     result = [q[criteria(c)] for q, c in recorded]
 388     return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
 389
 390
 391 ######################################################################
 392
 393
 394 def create_c_quizzes(
 395     models,
 396     quizz_machine,
 397     nb_for_train=1000,
 398     nb_for_test=100,
 399 ):
 400     recorded = []
 401
 402     nb_to_create = nb_for_train + nb_for_test
 403
 404     # ------------------------------------------------------------
 405
 406     standard_validity = lambda nb_correct: torch.logical_and(
 407         nb_correct >= args.min_to_validate, nb_correct <= args.max_to_validate
 408     )
 409
 410     file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
 411     with open(file_name, "w") as logp_file:
 412         while valid_c_quizzes(recorded, standard_validity).size(0) < nb_to_create:
 413             # Select a model at random to generate the new quizzes
 414
 415             model_for_generation = models[torch.randint(len(models), (1,))]
 416
 417             c_quizzes = quizz_machine.generate_quizzes(
 418                 nb_to_create,
 419                 model_for_generation=model_for_generation,
 420                 temperature=args.generation_temperature,
 421             )
 422
 423             nb_correct, seq_logproba = quizz_machine.compute_correctness(
 424                 c_quizzes,
 425                 models,
 426                 both_directions=args.both_directions,
 427                 deterministic_validation=not args.stochastic_validation,
 428             )
 429
 430             for n, l in zip(nb_correct, seq_logproba):
 431                 s = " ".join([str(x.item()) for x in l])
 432                 logp_file.write(f"{n} {s}\n")
 433
 434             if args.dirty_debug:
 435                 nb_correct = torch.randint(
 436                     len(models) + 1, nb_correct.size(), device=c_quizzes.device
 437                 )
 438
 439             recorded.append((c_quizzes, nb_correct))
 440
 441             nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0)
 442             nv = " ".join([str(x.item()) for x in nv])
 443
 444             nb_validated = valid_c_quizzes(recorded, standard_validity).size(0)
 445
 446             log_string(
 447                 f"keep c_quizzes model {model_for_generation.id} kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
 448             )
 449
 450     # store the new c_quizzes which have been validated
 451
 452     new_c_quizzes = valid_c_quizzes(recorded, standard_validity)
 453
 454     quizz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 455     quizz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 456
 457     # save a bunch of images to investigate what quizzes with a
 458     # certain nb of correct predictions look like
 459
 460     for n in range(len(models) + 1):
 461         s = (
 462             "_validated"
 463             if n >= args.min_to_validate and n <= args.max_to_validate
 464             else ""
 465         )
 466
 467         q = valid_c_quizzes(recorded, criteria=lambda nb_correct: nb_correct == n)[:72]
 468
 469         if q.size(0) > 0:
 470             quizz_machine.save_quizzes(
 471                 args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
 472             )
 473
 474
 475 ######################################################################
 476
 477 models = []
 478
 479 for k in range(args.nb_gpts):
 480     model = mygpt.MyGPT(
 481         vocabulary_size=vocabulary_size,
 482         dim_model=args.dim_model,
 483         dim_keys=args.dim_keys,
 484         dim_hidden=args.dim_hidden,
 485         nb_heads=args.nb_heads,
 486         nb_blocks=args.nb_blocks,
 487         causal=True,
 488         dropout=args.dropout,
 489     ).to(device)
 490
 491     model.main_test_accuracy = 0.0
 492     model.id = k
 493
 494     models.append(model)
 495
 496
 497 nb_parameters = sum(p.numel() for p in models[0].parameters())
 498 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 499
 500 ######################################################################
 501
 502 for n_epoch in range(args.nb_epochs):
 503     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 504
 505     # Select, improve, and eval the worst model
 506
 507     weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))
 508
 509     log_string(
 510         f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
 511     )
 512
 513     one_epoch(weakest_model, quizz_machine)
 514
 515     log_string(
 516         f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
 517     )
 518
 519     run_tests(weakest_model, quizz_machine, deterministic_synthesis=False)
 520
 521     log_string(
 522         f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
 523     )
 524
 525     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
 526     log_string(f"current_test_accuracies {cta}")
 527
 528     # Replace a fraction of the w_quizzes with fresh ones
 529
 530     quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
 531
 532     # If all the models are good enough, generate new quizzes and
 533     # re-compute the test errors
 534
 535     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
 536         create_c_quizzes(
 537             models,
 538             quizz_machine,
 539             nb_for_train=nb_new_c_quizzes_for_train,
 540             nb_for_test=nb_new_c_quizzes_for_test,
 541         )
 542
 543         for model in models:
 544             run_tests(model, quizz_machine, deterministic_synthesis=False)
 545
 546
 547 ######################################################################