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, reasoning, 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     back_accuracy = False
 253 elif args.problem == "reasoning":
 254     problem = reasoning.Reasoning(device=device)
 255     back_accuracy = True
 256 else:
 257     raise ValueError
 258
 259 quizz_machine = quizz_machine.QuizzMachine(
 260     problem=problem,
 261     nb_train_samples=args.nb_train_samples,
 262     nb_test_samples=args.nb_test_samples,
 263     back_accuracy=back_accuracy,
 264     batch_size=args.physical_batch_size,
 265     result_dir=args.result_dir,
 266     logger=log_string,
 267     device=device,
 268 )
 269
 270 ######################################################################
 271
 272 log_string(f"device {device}")
 273
 274 vocabulary_size = quizz_machine.vocabulary_size()
 275
 276 log_string(f"vocabulary_size {vocabulary_size}")
 277
 278 ######################################################################
 279
 280 # Compute the entropy of the training tokens
 281
 282 token_count = 0
 283 for input in quizz_machine.batches(split="train", desc="train-entropy"):
 284     token_count += F.one_hot(input, num_classes=quizz_machine.vocabulary_size()).sum(
 285         (0, 1)
 286     )
 287 token_probas = token_count / token_count.sum()
 288 entropy = -torch.xlogy(token_probas, token_probas).sum()
 289 train_set_perplexity = math.exp(entropy)
 290
 291 ######################################################################
 292 # A bit of paranoia never hurts
 293
 294 if args.max_percents_of_test_in_train >= 0:
 295
 296     def subsets_as_tuples(batches, cs):
 297         s = set()
 298         for batch in batches:
 299             for x in batch:
 300                 s.add(tuple([v.item() for v in x]))
 301                 if len(s) == cs:
 302                     yield s
 303                     s = set()
 304         yield s
 305
 306     nb_test, nb_in_train = 0, 0
 307     for test_subset in subsets_as_tuples(
 308         quizz_machine.batches(split="test", desc="test-check"), 25000
 309     ):
 310         in_train = set()
 311         for train_subset in subsets_as_tuples(
 312             quizz_machine.batches(split="train", desc="train-check"), 25000
 313         ):
 314             in_train.update(test_subset.intersection(train_subset))
 315         nb_in_train += len(in_train)
 316         nb_test += len(test_subset)
 317
 318     log_string(
 319         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 320     )
 321
 322     assert (
 323         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 324     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 325
 326 ##############################
 327
 328
 329 def one_epoch(model, quizz_machine):
 330     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 331
 332     model.train()
 333
 334     nb_train_samples, acc_train_loss = 0, 0.0
 335
 336     for input in quizz_machine.batches(split="train"):
 337         input = input.to(device)
 338
 339         if nb_train_samples % args.batch_size == 0:
 340             optimizer.zero_grad()
 341
 342         output = model(mygpt.BracketedSequence(input)).x
 343         loss = F.cross_entropy(output.transpose(1, 2), input)
 344         acc_train_loss += loss.item() * input.size(0)
 345
 346         nb_train_samples += input.size(0)
 347
 348         loss.backward()
 349
 350         if nb_train_samples % args.batch_size == 0:
 351             optimizer.step()
 352
 353     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 354
 355     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 356
 357
 358 ######################################################################
 359
 360
 361 def run_tests(model, quizz_machine, deterministic_synthesis):
 362     with torch.autograd.no_grad():
 363         model.eval()
 364
 365         nb_test_samples, acc_test_loss = 0, 0.0
 366         nb_samples_accumulated = 0
 367
 368         for input in quizz_machine.batches(split="test"):
 369             input = input.to(device)
 370
 371             bs = model(mygpt.BracketedSequence(input))
 372             output = bs.x
 373
 374             loss = F.cross_entropy(output.transpose(1, 2), input)
 375
 376             acc_test_loss += loss.item() * input.size(0)
 377
 378             nb_test_samples += input.size(0)
 379
 380         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 381
 382         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 383
 384         model.main_test_accuracy = quizz_machine.produce_results(
 385             n_epoch=n_epoch,
 386             model=model,
 387             result_dir=args.result_dir,
 388             deterministic_synthesis=deterministic_synthesis,
 389         )
 390
 391
 392 ######################################################################
 393
 394
 395 def valid_c_quizzes(recorded, criteria):
 396     result = [q[criteria(c)] for q, c in recorded]
 397     return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
 398
 399
 400 ######################################################################
 401
 402
 403 def create_c_quizzes(
 404     models,
 405     quizz_machine,
 406     nb_for_train=1000,
 407     nb_for_test=100,
 408 ):
 409     recorded = []
 410
 411     nb_to_create = nb_for_train + nb_for_test
 412
 413     # ------------------------------------------------------------
 414
 415     standard_validity = lambda nb_correct: torch.logical_and(
 416         nb_correct >= args.min_to_validate, nb_correct <= args.max_to_validate
 417     )
 418
 419     file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
 420     with open(file_name, "w") as logp_file:
 421         while valid_c_quizzes(recorded, standard_validity).size(0) < nb_to_create:
 422             # Select a model at random to generate the new quizzes
 423
 424             model_for_generation = models[torch.randint(len(models), (1,))]
 425
 426             c_quizzes = quizz_machine.generate_quizzes(
 427                 nb_to_create,
 428                 model_for_generation=model_for_generation,
 429                 temperature=args.generation_temperature,
 430             )
 431
 432             nb_correct, seq_logproba = quizz_machine.compute_correctness(
 433                 c_quizzes,
 434                 models,
 435                 bidirectional_validation=args.bidirectional_validation,
 436                 deterministic_validation=args.deterministic_validation,
 437             )
 438
 439             for n, l in zip(nb_correct, seq_logproba):
 440                 s = " ".join([str(x.item()) for x in l])
 441                 logp_file.write(f"{n} {s}\n")
 442
 443             if args.dirty_debug:
 444                 nb_correct = torch.randint(
 445                     len(models) + 1, nb_correct.size(), device=c_quizzes.device
 446                 )
 447
 448             recorded.append((c_quizzes, nb_correct))
 449
 450             nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0)
 451             nv = " ".join([str(x.item()) for x in nv])
 452
 453             nb_validated = valid_c_quizzes(recorded, standard_validity).size(0)
 454
 455             log_string(
 456                 f"keep c_quizzes model {model_for_generation.id} kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
 457             )
 458
 459     # store the new c_quizzes which have been validated
 460
 461     new_c_quizzes = valid_c_quizzes(recorded, standard_validity)
 462
 463     quizz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 464     quizz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 465
 466     # save a bunch of images to investigate what quizzes with a
 467     # certain nb of correct predictions look like
 468
 469     for n in range(len(models) + 1):
 470         s = (
 471             "_validated"
 472             if n >= args.min_to_validate and n <= args.max_to_validate
 473             else ""
 474         )
 475
 476         q = valid_c_quizzes(recorded, criteria=lambda nb_correct: nb_correct == n)[:72]
 477
 478         if q.size(0) > 0:
 479             quizz_machine.save_quizzes(
 480                 args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
 481             )
 482
 483
 484 ######################################################################
 485
 486 models = []
 487
 488 for k in range(args.nb_gpts):
 489     model = mygpt.MyGPT(
 490         vocabulary_size=vocabulary_size,
 491         dim_model=args.dim_model,
 492         dim_keys=args.dim_keys,
 493         dim_hidden=args.dim_hidden,
 494         nb_heads=args.nb_heads,
 495         nb_blocks=args.nb_blocks,
 496         causal=True,
 497         dropout=args.dropout,
 498     ).to(device)
 499
 500     model.main_test_accuracy = 0.0
 501     model.id = k
 502
 503     models.append(model)
 504
 505
 506 nb_parameters = sum(p.numel() for p in models[0].parameters())
 507 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 508
 509 ######################################################################
 510
 511 for n_epoch in range(args.nb_epochs):
 512     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 513
 514     # Select, improve, and eval the worst model
 515
 516     weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))
 517
 518     log_string(
 519         f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
 520     )
 521
 522     one_epoch(weakest_model, quizz_machine)
 523
 524     log_string(
 525         f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
 526     )
 527
 528     run_tests(weakest_model, quizz_machine, deterministic_synthesis=False)
 529
 530     log_string(
 531         f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
 532     )
 533
 534     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
 535     log_string(f"current_test_accuracies {cta}")
 536
 537     # Replace a fraction of the w_quizzes with fresh ones
 538
 539     quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
 540
 541     # If all the models are good enough, generate new quizzes and
 542     # re-compute the test errors
 543
 544     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
 545         create_c_quizzes(
 546             models,
 547             quizz_machine,
 548             nb_for_train=nb_new_c_quizzes_for_train,
 549             nb_for_test=nb_new_c_quizzes_for_test,
 550         )
 551
 552         for model in models:
 553             run_tests(model, quizz_machine, deterministic_synthesis=False)
 554
 555
 556 ######################################################################