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("--problem", type=str, default="sky")
  85
  86 parser.add_argument("--nb_gpts", type=int, default=5)
  87
  88 parser.add_argument("--nb_models_for_generation", type=int, default=1)
  89
  90 parser.add_argument("--generation_mode", type=str, default="groupthink")
  91
  92 parser.add_argument("--min_to_validate", type=int, default=4)
  93
  94 parser.add_argument("--max_to_validate", type=int, default=4)
  95
  96 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
  97
  98 parser.add_argument("--dirty_debug", action="store_true", default=False)
  99
 100 parser.add_argument("--sky_height", type=int, default=6)
 101
 102 parser.add_argument("--sky_width", type=int, default=8)
 103
 104 parser.add_argument("--sky_nb_birds", type=int, default=3)
 105
 106 parser.add_argument("--sky_nb_iterations", type=int, default=2)
 107
 108 parser.add_argument("--sky_speed", type=int, default=3)
 109
 110 ######################################################################
 111
 112 args = parser.parse_args()
 113
 114 if args.result_dir is None:
 115     args.result_dir = f"results_culture"
 116
 117 ######################################################################
 118
 119 if args.dirty_debug:
 120     args.accuracy_to_make_c_quizzes = 0.0
 121     nb_new_c_quizzes_for_train = 100
 122     nb_new_c_quizzes_for_test = 10
 123
 124 ######################################################################
 125
 126 default_args = {
 127     "model": "37M",
 128     "batch_size": 100,
 129     "nb_train_samples": 100000,
 130     "nb_test_samples": 10000,
 131 }
 132
 133 for k, v in default_args.items():
 134     if getattr(args, k) is None:
 135         setattr(args, k, v)
 136
 137 ######################################################################
 138
 139 default_model_args = {
 140     "17K": {
 141         "dim_model": 32,
 142         "dim_keys": 32,
 143         "dim_hidden": 32,
 144         "nb_heads": 2,
 145         "nb_blocks": 2,
 146     },
 147     "4M": {
 148         "dim_model": 256,
 149         "dim_keys": 32,
 150         "dim_hidden": 1024,
 151         "nb_heads": 4,
 152         "nb_blocks": 6,
 153     },
 154     "37M": {
 155         "dim_model": 512,
 156         "dim_keys": 64,
 157         "dim_hidden": 2048,
 158         "nb_heads": 8,
 159         "nb_blocks": 12,
 160     },
 161     "122M": {
 162         "dim_model": 768,
 163         "dim_keys": 64,
 164         "dim_hidden": 2048,
 165         "nb_heads": 8,
 166         "nb_blocks": 24,
 167     },
 168     "352M": {
 169         "dim_model": 1024,
 170         "dim_keys": 64,
 171         "dim_hidden": 2048,
 172         "nb_heads": 8,
 173         "nb_blocks": 48,
 174     },
 175 }
 176
 177 if args.model in default_model_args:
 178     for k, v in default_model_args[args.model].items():
 179         if getattr(args, k) is None:
 180             setattr(args, k, v)
 181 else:
 182     raise ValueError(f"Unknown model {args.model}")
 183
 184 ######################################################################
 185
 186 try:
 187     os.mkdir(args.result_dir)
 188 except FileExistsError:
 189     print(f"result directory {args.result_dir} already exists")
 190     exit(1)
 191
 192 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 193
 194 if args.seed >= 0:
 195     # torch.backends.cudnn.deterministic = True
 196     # torch.backends.cudnn.benchmark = False
 197     # torch.use_deterministic_algorithms(True)
 198     torch.manual_seed(args.seed)
 199     if torch.cuda.is_available():
 200         torch.cuda.manual_seed_all(args.seed)
 201
 202 ######################################################################
 203
 204
 205 def log_string(s):
 206     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 207
 208     if log_file is not None:
 209         log_file.write(t + s + "\n")
 210         log_file.flush()
 211
 212     print(t + s)
 213     sys.stdout.flush()
 214
 215
 216 log_string(f"argv {' '.join(sys.argv)}")
 217
 218 for n in vars(args):
 219     log_string(f"args.{n} {getattr(args, n)}")
 220
 221
 222 ######################################################################
 223
 224 if args.dirty_debug:
 225     args.nb_train_samples = 2500
 226     args.nb_test_samples = 100
 227
 228 if args.physical_batch_size is None:
 229     args.physical_batch_size = args.batch_size
 230 else:
 231     assert args.batch_size % args.physical_batch_size == 0
 232
 233 assert args.nb_train_samples % args.batch_size == 0
 234 assert args.nb_test_samples % args.batch_size == 0
 235
 236 if args.problem == "sky":
 237     problem = sky.Sky(
 238         height=args.sky_height,
 239         width=args.sky_width,
 240         nb_birds=args.sky_nb_birds,
 241         nb_iterations=args.sky_nb_iterations,
 242         speed=args.sky_speed,
 243     )
 244 elif args.problem == "wireworld":
 245     problem = wireworld.Wireworld(height=8, width=10, nb_iterations=2, speed=5)
 246 else:
 247     raise ValueError
 248
 249 quizz_machine = quizz_machine.QuizzMachine(
 250     problem=problem,
 251     nb_train_samples=args.nb_train_samples,
 252     nb_test_samples=args.nb_test_samples,
 253     batch_size=args.physical_batch_size,
 254     result_dir=args.result_dir,
 255     logger=log_string,
 256     device=device,
 257 )
 258
 259 ######################################################################
 260
 261 log_string(f"device {device}")
 262
 263 vocabulary_size = quizz_machine.vocabulary_size()
 264
 265 log_string(f"vocabulary_size {vocabulary_size}")
 266
 267 ######################################################################
 268
 269 # Compute the entropy of the training tokens
 270
 271 token_count = 0
 272 for input in quizz_machine.batches(split="train", desc="train-entropy"):
 273     token_count += F.one_hot(input, num_classes=quizz_machine.vocabulary_size()).sum(
 274         (0, 1)
 275     )
 276 token_probas = token_count / token_count.sum()
 277 entropy = -torch.xlogy(token_probas, token_probas).sum()
 278 train_set_perplexity = math.exp(entropy)
 279
 280 ######################################################################
 281 # A bit of paranoia never hurts
 282
 283 if args.max_percents_of_test_in_train >= 0:
 284
 285     def subsets_as_tuples(batches, cs):
 286         s = set()
 287         for batch in batches:
 288             for x in batch:
 289                 s.add(tuple([v.item() for v in x]))
 290                 if len(s) == cs:
 291                     yield s
 292                     s = set()
 293         yield s
 294
 295     nb_test, nb_in_train = 0, 0
 296     for test_subset in subsets_as_tuples(
 297         quizz_machine.batches(split="test", desc="test-check"), 25000
 298     ):
 299         in_train = set()
 300         for train_subset in subsets_as_tuples(
 301             quizz_machine.batches(split="train", desc="train-check"), 25000
 302         ):
 303             in_train.update(test_subset.intersection(train_subset))
 304         nb_in_train += len(in_train)
 305         nb_test += len(test_subset)
 306
 307     log_string(
 308         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 309     )
 310
 311     assert (
 312         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 313     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 314
 315 ##############################
 316
 317
 318 def one_epoch(model, quizz_machine):
 319     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 320
 321     model.train()
 322
 323     nb_train_samples, acc_train_loss = 0, 0.0
 324
 325     for input in quizz_machine.batches(split="train"):
 326         input = input.to(device)
 327
 328         if nb_train_samples % args.batch_size == 0:
 329             optimizer.zero_grad()
 330
 331         output = model(mygpt.BracketedSequence(input)).x
 332         loss = F.cross_entropy(output.transpose(1, 2), input)
 333         acc_train_loss += loss.item() * input.size(0)
 334
 335         nb_train_samples += input.size(0)
 336
 337         loss.backward()
 338
 339         if nb_train_samples % args.batch_size == 0:
 340             optimizer.step()
 341
 342     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 343
 344     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 345
 346
 347 ######################################################################
 348
 349
 350 def run_tests(model, quizz_machine, deterministic_synthesis):
 351     with torch.autograd.no_grad():
 352         model.eval()
 353
 354         nb_test_samples, acc_test_loss = 0, 0.0
 355         nb_samples_accumulated = 0
 356
 357         for input in quizz_machine.batches(split="test"):
 358             input = input.to(device)
 359
 360             bs = model(mygpt.BracketedSequence(input))
 361             output = bs.x
 362
 363             loss = F.cross_entropy(output.transpose(1, 2), input)
 364
 365             acc_test_loss += loss.item() * input.size(0)
 366
 367             nb_test_samples += input.size(0)
 368
 369         main_test_accuracy = quizz_machine.produce_results(
 370             n_epoch=n_epoch,
 371             model=model,
 372             result_dir=args.result_dir,
 373             deterministic_synthesis=deterministic_synthesis,
 374         )
 375
 376         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 377
 378         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 379
 380     model.main_test_accuracy = main_test_accuracy
 381
 382
 383 ######################################################################
 384
 385
 386 def create_c_quizzes(
 387     models,
 388     quizz_machine,
 389     nb_for_train=1000,
 390     nb_for_test=100,
 391     min_ave_seq_logproba=None,
 392 ):
 393     # We will store the generated quizzes for each number of
 394     # correct prediction
 395     recorded = dict([(n, []) for n in range(len(models) + 1)])
 396
 397     model_indexes = []
 398     sum_logits, sum_nb_c_quizzes = 0, 0
 399
 400     def nb_generated():
 401         return sum([sum([x.size(0) for x in recorded[n]]) for n in recorded.keys()])
 402
 403     def nb_validated():
 404         return sum(
 405             [
 406                 sum([x.size(0) for x in recorded[n]])
 407                 for n in range(args.min_to_validate, args.max_to_validate + 1)
 408             ]
 409         )
 410
 411     nb_to_create = nb_for_train + nb_for_test
 412
 413     while nb_validated() < nb_to_create:
 414         (
 415             new_c_quizzes,
 416             nb_correct,
 417             ave_seq_logproba,
 418         ) = quizz_machine.gang_create_c_quizzes(
 419             nb=nb_to_create,
 420             nb_models_for_generation=args.nb_models_for_generation,
 421             models=models,
 422             mode=args.generation_mode,
 423             reverse_cleanup=args.reverse_cleanup,
 424             min_ave_seq_logproba=min_ave_seq_logproba,
 425             n_epoch=n_epoch,
 426             result_dir=args.result_dir,
 427         )
 428
 429         sum_logits += new_c_quizzes.size(0) * ave_seq_logproba
 430         sum_nb_c_quizzes += new_c_quizzes.size(0)
 431
 432         if args.dirty_debug:
 433             nb_correct = torch.randint(
 434                 len(models) + 1, nb_correct.size(), device=new_c_quizzes.device
 435             )
 436
 437         for n in range(nb_correct.max() + 1):
 438             recorded[n].append(new_c_quizzes[nb_correct == n].clone())
 439
 440         nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0)
 441         nv = " ".join([str(x.item()) for x in nv])
 442
 443         log_string(f"keep c_quizzes kept {nv} total {nb_validated()} / {nb_to_create}")
 444
 445     # concatenate and shuffle
 446     for n in recorded.keys():
 447         if len(recorded[n]) > 0:
 448             q = torch.cat(recorded[n], dim=0)
 449             q = q[torch.randperm(q.size(0), device=q.device)]
 450             recorded[n] = q
 451         else:
 452             del recorded[n]
 453
 454     new_c_quizzes = torch.cat(
 455         [recorded[n] for n in range(args.min_to_validate, args.max_to_validate + 1)],
 456         dim=0,
 457     )
 458
 459     new_c_quizzes = new_c_quizzes[
 460         torch.randperm(new_c_quizzes.size(0), device=new_c_quizzes.device)[
 461             : nb_for_train + nb_for_test
 462         ]
 463     ]
 464
 465     quizz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 466     quizz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 467
 468     for n in recorded.keys():
 469         s = (
 470             "_validated"
 471             if n >= args.min_to_validate and n <= args.max_to_validate
 472             else ""
 473         )
 474         quizz_machine.problem.save_quizzes(
 475             recorded[n][:72],
 476             args.result_dir,
 477             f"culture_c_quiz_{n_epoch:04d}_N{n}{s}",
 478         )
 479
 480     return sum_logits / sum_nb_c_quizzes
 481
 482
 483 ######################################################################
 484
 485 models = []
 486
 487 for k in range(args.nb_gpts):
 488     model = mygpt.MyGPT(
 489         vocabulary_size=vocabulary_size,
 490         dim_model=args.dim_model,
 491         dim_keys=args.dim_keys,
 492         dim_hidden=args.dim_hidden,
 493         nb_heads=args.nb_heads,
 494         nb_blocks=args.nb_blocks,
 495         causal=True,
 496         dropout=args.dropout,
 497     ).to(device)
 498
 499     model.main_test_accuracy = 0.0
 500     model.id = k
 501
 502     models.append(model)
 503
 504
 505 nb_parameters = sum(p.numel() for p in models[0].parameters())
 506 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 507
 508 ######################################################################
 509
 510 min_ave_seq_logproba = None
 511
 512 for n_epoch in range(args.nb_epochs):
 513     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 514
 515     a = [(model.id, float(model.main_test_accuracy)) for model in models]
 516     a.sort(key=lambda p: p[0])
 517     s = " ".join([f"{p[1]*100:.02f}%" for p in a])
 518     log_string(f"current accuracies {s}")
 519
 520     # select the model with lowest accuracy
 521     models.sort(key=lambda model: model.main_test_accuracy)
 522     model = models[0]
 523
 524     log_string(
 525         f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
 526     )
 527
 528     # improve it
 529     one_epoch(model, quizz_machine)
 530
 531     quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
 532
 533     log_string(
 534         f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
 535     )
 536
 537     # test it
 538     run_tests(model, quizz_machine, deterministic_synthesis=False)
 539
 540     log_string(
 541         f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
 542     )
 543
 544     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
 545         ave_seq_logproba = 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             min_ave_seq_logproba=min_ave_seq_logproba,
 551         )
 552
 553         # We keep the first average logits as a reference
 554         # if min_ave_seq_logproba is None:
 555         # min_ave_seq_logproba = ave_seq_logproba
 556         # else:
 557         # log_string(
 558         # f"min_ave_seq_logproba {min_ave_seq_logproba} ave_seq_logproba {ave_seq_logproba}"
 559         # )
 560
 561         # We update everyone
 562         for model in models:
 563             run_tests(model, quizz_machine, deterministic_synthesis=False)
 564
 565
 566 ######################################################################