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