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, lang, 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("--problem", type=str, default="sky")
  83
  84 parser.add_argument("--nb_gpts", type=int, default=5)
  85
  86 parser.add_argument("--min_to_validate", type=int, default=None)
  87
  88 parser.add_argument("--max_to_validate", type=int, default=None)
  89
  90 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
  91
  92 parser.add_argument("--generation_temperature", type=float, default=2.0)
  93
  94 parser.add_argument("--deterministic_validation", action="store_true", default=False)
  95
  96 parser.add_argument("--bidirectional_validation", action="store_true", default=False)
  97
  98 parser.add_argument("--dirty_debug", 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.min_to_validate is None:
 117     args.min_to_validate = args.nb_gpts - 1
 118
 119 if args.max_to_validate is None:
 120     args.max_to_validate = args.nb_gpts - 1
 121
 122 if args.result_dir is None:
 123     args.result_dir = f"results_culture"
 124
 125 ######################################################################
 126
 127 if args.dirty_debug:
 128     args.accuracy_to_make_c_quizzes = 0.0
 129     nb_new_c_quizzes_for_train = 100
 130     nb_new_c_quizzes_for_test = 10
 131
 132 ######################################################################
 133
 134 default_args = {
 135     "model": "37M",
 136     "batch_size": 100,
 137     "nb_train_samples": 100000,
 138     "nb_test_samples": 10000,
 139 }
 140
 141 for k, v in default_args.items():
 142     if getattr(args, k) is None:
 143         setattr(args, k, v)
 144
 145 ######################################################################
 146
 147 default_model_args = {
 148     "17K": {
 149         "dim_model": 32,
 150         "dim_keys": 32,
 151         "dim_hidden": 32,
 152         "nb_heads": 2,
 153         "nb_blocks": 2,
 154     },
 155     "4M": {
 156         "dim_model": 256,
 157         "dim_keys": 32,
 158         "dim_hidden": 1024,
 159         "nb_heads": 4,
 160         "nb_blocks": 6,
 161     },
 162     "37M": {
 163         "dim_model": 512,
 164         "dim_keys": 64,
 165         "dim_hidden": 2048,
 166         "nb_heads": 8,
 167         "nb_blocks": 12,
 168     },
 169     "122M": {
 170         "dim_model": 768,
 171         "dim_keys": 64,
 172         "dim_hidden": 2048,
 173         "nb_heads": 8,
 174         "nb_blocks": 24,
 175     },
 176     "352M": {
 177         "dim_model": 1024,
 178         "dim_keys": 64,
 179         "dim_hidden": 2048,
 180         "nb_heads": 8,
 181         "nb_blocks": 48,
 182     },
 183 }
 184
 185 if args.model in default_model_args:
 186     for k, v in default_model_args[args.model].items():
 187         if getattr(args, k) is None:
 188             setattr(args, k, v)
 189 else:
 190     raise ValueError(f"Unknown model {args.model}")
 191
 192 ######################################################################
 193
 194 try:
 195     os.mkdir(args.result_dir)
 196 except FileExistsError:
 197     print(f"result directory {args.result_dir} already exists")
 198     exit(1)
 199
 200 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 201
 202 if args.seed >= 0:
 203     # torch.backends.cudnn.deterministic = True
 204     # torch.backends.cudnn.benchmark = False
 205     # torch.use_deterministic_algorithms(True)
 206     torch.manual_seed(args.seed)
 207     if torch.cuda.is_available():
 208         torch.cuda.manual_seed_all(args.seed)
 209
 210 ######################################################################
 211
 212
 213 def log_string(s):
 214     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 215
 216     if log_file is not None:
 217         log_file.write(t + s + "\n")
 218         log_file.flush()
 219
 220     print(t + s)
 221     sys.stdout.flush()
 222
 223
 224 log_string(f"argv {' '.join(sys.argv)}")
 225
 226 for n in vars(args):
 227     log_string(f"args.{n} {getattr(args, n)}")
 228
 229
 230 ######################################################################
 231
 232 if args.dirty_debug:
 233     args.nb_train_samples = 2500
 234     args.nb_test_samples = 100
 235
 236 if args.physical_batch_size is None:
 237     args.physical_batch_size = args.batch_size
 238 else:
 239     assert args.batch_size % args.physical_batch_size == 0
 240
 241 assert args.nb_train_samples % args.batch_size == 0
 242 assert args.nb_test_samples % args.batch_size == 0
 243
 244 if args.problem == "sky":
 245     problem = sky.Sky(
 246         height=args.sky_height,
 247         width=args.sky_width,
 248         nb_birds=args.sky_nb_birds,
 249         nb_iterations=args.sky_nb_iterations,
 250         speed=args.sky_speed,
 251     )
 252 elif args.problem == "lang":
 253     problem = lang.Lang(nb_iterations=2)
 254 else:
 255     raise ValueError
 256
 257 quizz_machine = quizz_machine.QuizzMachine(
 258     problem=problem,
 259     nb_train_samples=args.nb_train_samples,
 260     nb_test_samples=args.nb_test_samples,
 261     batch_size=args.physical_batch_size,
 262     result_dir=args.result_dir,
 263     logger=log_string,
 264     device=device,
 265 )
 266
 267 ######################################################################
 268
 269 log_string(f"device {device}")
 270
 271 vocabulary_size = quizz_machine.vocabulary_size()
 272
 273 log_string(f"vocabulary_size {vocabulary_size}")
 274
 275 ######################################################################
 276
 277 # Compute the entropy of the training tokens
 278
 279 token_count = 0
 280 for input in quizz_machine.batches(split="train", desc="train-entropy"):
 281     token_count += F.one_hot(input, num_classes=quizz_machine.vocabulary_size()).sum(
 282         (0, 1)
 283     )
 284 token_probas = token_count / token_count.sum()
 285 entropy = -torch.xlogy(token_probas, token_probas).sum()
 286 train_set_perplexity = math.exp(entropy)
 287
 288 ######################################################################
 289 # A bit of paranoia never hurts
 290
 291 if args.max_percents_of_test_in_train >= 0:
 292
 293     def subsets_as_tuples(batches, cs):
 294         s = set()
 295         for batch in batches:
 296             for x in batch:
 297                 s.add(tuple([v.item() for v in x]))
 298                 if len(s) == cs:
 299                     yield s
 300                     s = set()
 301         yield s
 302
 303     nb_test, nb_in_train = 0, 0
 304     for test_subset in subsets_as_tuples(
 305         quizz_machine.batches(split="test", desc="test-check"), 25000
 306     ):
 307         in_train = set()
 308         for train_subset in subsets_as_tuples(
 309             quizz_machine.batches(split="train", desc="train-check"), 25000
 310         ):
 311             in_train.update(test_subset.intersection(train_subset))
 312         nb_in_train += len(in_train)
 313         nb_test += len(test_subset)
 314
 315     log_string(
 316         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 317     )
 318
 319     assert (
 320         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 321     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 322
 323 ##############################
 324
 325
 326 def one_epoch(model, quizz_machine):
 327     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 328
 329     model.train()
 330
 331     nb_train_samples, acc_train_loss = 0, 0.0
 332
 333     for input in quizz_machine.batches(split="train"):
 334         input = input.to(device)
 335
 336         if nb_train_samples % args.batch_size == 0:
 337             optimizer.zero_grad()
 338
 339         output = model(mygpt.BracketedSequence(input)).x
 340         loss = F.cross_entropy(output.transpose(1, 2), input)
 341         acc_train_loss += loss.item() * input.size(0)
 342
 343         nb_train_samples += input.size(0)
 344
 345         loss.backward()
 346
 347         if nb_train_samples % args.batch_size == 0:
 348             optimizer.step()
 349
 350     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 351
 352     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 353
 354
 355 ######################################################################
 356
 357
 358 def run_tests(model, quizz_machine, deterministic_synthesis):
 359     with torch.autograd.no_grad():
 360         model.eval()
 361
 362         nb_test_samples, acc_test_loss = 0, 0.0
 363         nb_samples_accumulated = 0
 364
 365         for input in quizz_machine.batches(split="test"):
 366             input = input.to(device)
 367
 368             bs = model(mygpt.BracketedSequence(input))
 369             output = bs.x
 370
 371             loss = F.cross_entropy(output.transpose(1, 2), input)
 372
 373             acc_test_loss += loss.item() * input.size(0)
 374
 375             nb_test_samples += input.size(0)
 376
 377         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 378
 379         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 380
 381         model.main_test_accuracy = quizz_machine.produce_results(
 382             n_epoch=n_epoch,
 383             model=model,
 384             result_dir=args.result_dir,
 385             deterministic_synthesis=deterministic_synthesis,
 386         )
 387
 388
 389 ######################################################################
 390
 391
 392 def valid_c_quizzes(recorded, criteria):
 393     result = [q[criteria(c)] for q, c in recorded]
 394     return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
 395
 396
 397 ######################################################################
 398
 399
 400 def create_c_quizzes(
 401     models,
 402     quizz_machine,
 403     nb_for_train=1000,
 404     nb_for_test=100,
 405 ):
 406     recorded = []
 407
 408     nb_to_create = nb_for_train + nb_for_test
 409
 410     # ------------------------------------------------------------
 411
 412     standard_validity = lambda nb_correct: torch.logical_and(
 413         nb_correct >= args.min_to_validate, nb_correct <= args.max_to_validate
 414     )
 415
 416     file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
 417     with open(file_name, "w") as logp_file:
 418         while valid_c_quizzes(recorded, standard_validity).size(0) < nb_to_create:
 419             # Select a model at random to generate the new quizzes
 420
 421             model_for_generation = models[torch.randint(len(models), (1,))]
 422
 423             c_quizzes = quizz_machine.generate_quizzes(
 424                 nb_to_create,
 425                 model_for_generation=model_for_generation,
 426                 temperature=args.generation_temperature,
 427             )
 428
 429             nb_correct, seq_logproba = quizz_machine.compute_correctness(
 430                 c_quizzes,
 431                 models,
 432                 bidirectional_validation=args.bidirectional_validation,
 433                 deterministic_validation=args.deterministic_validation,
 434             )
 435
 436             for n, l in zip(nb_correct, seq_logproba):
 437                 s = " ".join([str(x.item()) for x in l])
 438                 logp_file.write(f"{n} {s}\n")
 439
 440             if args.dirty_debug:
 441                 nb_correct = torch.randint(
 442                     len(models) + 1, nb_correct.size(), device=c_quizzes.device
 443                 )
 444
 445             recorded.append((c_quizzes, nb_correct))
 446
 447             nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0)
 448             nv = " ".join([str(x.item()) for x in nv])
 449
 450             nb_validated = valid_c_quizzes(recorded, standard_validity).size(0)
 451
 452             log_string(
 453                 f"keep c_quizzes model {model_for_generation.id} kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
 454             )
 455
 456     # store the new c_quizzes which have been validated
 457
 458     new_c_quizzes = valid_c_quizzes(recorded, standard_validity)
 459
 460     quizz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 461     quizz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 462
 463     # save a bunch of images to investigate what quizzes with a
 464     # certain nb of correct predictions look like
 465
 466     for n in range(len(models) + 1):
 467         s = (
 468             "_validated"
 469             if n >= args.min_to_validate and n <= args.max_to_validate
 470             else ""
 471         )
 472
 473         q = valid_c_quizzes(recorded, criteria=lambda nb_correct: nb_correct == n)[:72]
 474
 475         if q.size(0) > 0:
 476             quizz_machine.save_quizzes(
 477                 args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
 478             )
 479
 480
 481 ######################################################################
 482
 483 models = []
 484
 485 for k in range(args.nb_gpts):
 486     model = mygpt.MyGPT(
 487         vocabulary_size=vocabulary_size,
 488         dim_model=args.dim_model,
 489         dim_keys=args.dim_keys,
 490         dim_hidden=args.dim_hidden,
 491         nb_heads=args.nb_heads,
 492         nb_blocks=args.nb_blocks,
 493         causal=True,
 494         dropout=args.dropout,
 495     ).to(device)
 496
 497     model.main_test_accuracy = 0.0
 498     model.id = k
 499
 500     models.append(model)
 501
 502
 503 nb_parameters = sum(p.numel() for p in models[0].parameters())
 504 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 505
 506 ######################################################################
 507
 508 for n_epoch in range(args.nb_epochs):
 509     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 510
 511     # Select, improve, and eval the worst model
 512
 513     weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))
 514
 515     log_string(
 516         f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
 517     )
 518
 519     one_epoch(weakest_model, quizz_machine)
 520
 521     log_string(
 522         f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
 523     )
 524
 525     run_tests(weakest_model, quizz_machine, deterministic_synthesis=False)
 526
 527     log_string(
 528         f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
 529     )
 530
 531     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
 532     log_string(f"current_test_accuracies {cta}")
 533
 534     # Replace a fraction of the w_quizzes with fresh ones
 535
 536     quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
 537
 538     # If all the models are good enough, generate new quizzes and
 539     # re-compute the test errors
 540
 541     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
 542         create_c_quizzes(
 543             models,
 544             quizz_machine,
 545             nb_for_train=nb_new_c_quizzes_for_train,
 546             nb_for_test=nb_new_c_quizzes_for_test,
 547         )
 548
 549         for model in models:
 550             run_tests(model, quizz_machine, deterministic_synthesis=False)
 551
 552
 553 ######################################################################