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