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("--problem", type=str, default="sky")
  83
  84 parser.add_argument("--nb_gpts", type=int, default=5)
  85
  86 parser.add_argument("--nb_models_for_generation", type=int, default=1)
  87
  88 parser.add_argument("--generation_mode", type=str, default="groupthink")
  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 ######################################################################
  99
 100 args = parser.parse_args()
 101
 102 if args.result_dir is None:
 103     args.result_dir = f"results_culture"
 104
 105 ######################################################################
 106
 107 if args.dirty_debug:
 108     args.accuracy_to_make_c_quizzes = 0.0
 109     nb_new_c_quizzes_for_train = 100
 110     nb_new_c_quizzes_for_test = 10
 111
 112 ######################################################################
 113
 114 default_args = {
 115     "model": "37M",
 116     "batch_size": 100,
 117     "nb_train_samples": 100000,
 118     "nb_test_samples": 10000,
 119 }
 120
 121 for k, v in default_args.items():
 122     if getattr(args, k) is None:
 123         setattr(args, k, v)
 124
 125 ######################################################################
 126
 127 default_model_args = {
 128     "17K": {
 129         "dim_model": 32,
 130         "dim_keys": 32,
 131         "dim_hidden": 32,
 132         "nb_heads": 2,
 133         "nb_blocks": 2,
 134     },
 135     "4M": {
 136         "dim_model": 256,
 137         "dim_keys": 32,
 138         "dim_hidden": 1024,
 139         "nb_heads": 4,
 140         "nb_blocks": 6,
 141     },
 142     "37M": {
 143         "dim_model": 512,
 144         "dim_keys": 64,
 145         "dim_hidden": 2048,
 146         "nb_heads": 8,
 147         "nb_blocks": 12,
 148     },
 149     "122M": {
 150         "dim_model": 768,
 151         "dim_keys": 64,
 152         "dim_hidden": 2048,
 153         "nb_heads": 8,
 154         "nb_blocks": 24,
 155     },
 156     "352M": {
 157         "dim_model": 1024,
 158         "dim_keys": 64,
 159         "dim_hidden": 2048,
 160         "nb_heads": 8,
 161         "nb_blocks": 48,
 162     },
 163 }
 164
 165 if args.model in default_model_args:
 166     for k, v in default_model_args[args.model].items():
 167         if getattr(args, k) is None:
 168             setattr(args, k, v)
 169 else:
 170     raise ValueError(f"Unknown model {args.model}")
 171
 172 ######################################################################
 173
 174 try:
 175     os.mkdir(args.result_dir)
 176 except FileExistsError:
 177     print(f"result directory {args.result_dir} already exists")
 178     exit(1)
 179
 180 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 181
 182 if args.seed >= 0:
 183     # torch.backends.cudnn.deterministic = True
 184     # torch.backends.cudnn.benchmark = False
 185     # torch.use_deterministic_algorithms(True)
 186     torch.manual_seed(args.seed)
 187     if torch.cuda.is_available():
 188         torch.cuda.manual_seed_all(args.seed)
 189
 190 ######################################################################
 191
 192
 193 def log_string(s):
 194     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 195
 196     if log_file is not None:
 197         log_file.write(t + s + "\n")
 198         log_file.flush()
 199
 200     print(t + s)
 201     sys.stdout.flush()
 202
 203
 204 log_string(f"argv {' '.join(sys.argv)}")
 205
 206 for n in vars(args):
 207     log_string(f"args.{n} {getattr(args, n)}")
 208
 209
 210 ######################################################################
 211
 212 if args.dirty_debug:
 213     args.nb_train_samples = 2500
 214     args.nb_test_samples = 100
 215
 216 if args.physical_batch_size is None:
 217     args.physical_batch_size = args.batch_size
 218 else:
 219     assert args.batch_size % args.physical_batch_size == 0
 220
 221 assert args.nb_train_samples % args.batch_size == 0
 222 assert args.nb_test_samples % args.batch_size == 0
 223
 224 if args.problem=="sky":
 225     problem=sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2, speed=2),
 226 elif args.problem="wireworld":
 227     problem=wireworld.Wireworld(height=10, width=15, nb_iterations=4)
 228 else:
 229     raise ValueError
 230
 231 quizz_machine = quizz_machine.QuizzMachine(
 232     problem=problem,
 233     nb_train_samples=args.nb_train_samples,
 234     nb_test_samples=args.nb_test_samples,
 235     batch_size=args.physical_batch_size,
 236     result_dir=args.result_dir,
 237     logger=log_string,
 238     device=device,
 239 )
 240
 241 ######################################################################
 242
 243 log_string(f"device {device}")
 244
 245 vocabulary_size = quizz_machine.vocabulary_size()
 246
 247 log_string(f"vocabulary_size {vocabulary_size}")
 248
 249 ######################################################################
 250
 251 # Compute the entropy of the training tokens
 252
 253 token_count = 0
 254 for input in quizz_machine.batches(split="train", desc="train-entropy"):
 255     token_count += F.one_hot(input, num_classes=quizz_machine.vocabulary_size()).sum(
 256         (0, 1)
 257     )
 258 token_probas = token_count / token_count.sum()
 259 entropy = -torch.xlogy(token_probas, token_probas).sum()
 260 train_set_perplexity = math.exp(entropy)
 261
 262 ######################################################################
 263 # A bit of paranoia never hurts
 264
 265 if args.max_percents_of_test_in_train >= 0:
 266
 267     def subsets_as_tuples(batches, cs):
 268         s = set()
 269         for batch in batches:
 270             for x in batch:
 271                 s.add(tuple([v.item() for v in x]))
 272                 if len(s) == cs:
 273                     yield s
 274                     s = set()
 275         yield s
 276
 277     nb_test, nb_in_train = 0, 0
 278     for test_subset in subsets_as_tuples(
 279         quizz_machine.batches(split="test", desc="test-check"), 25000
 280     ):
 281         in_train = set()
 282         for train_subset in subsets_as_tuples(
 283             quizz_machine.batches(split="train", desc="train-check"), 25000
 284         ):
 285             in_train.update(test_subset.intersection(train_subset))
 286         nb_in_train += len(in_train)
 287         nb_test += len(test_subset)
 288
 289     log_string(
 290         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 291     )
 292
 293     assert (
 294         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 295     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 296
 297 ##############################
 298
 299
 300 def one_epoch(model, quizz_machine):
 301     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 302
 303     model.train()
 304
 305     nb_train_samples, acc_train_loss = 0, 0.0
 306
 307     for input in quizz_machine.batches(split="train"):
 308         input = input.to(device)
 309
 310         if nb_train_samples % args.batch_size == 0:
 311             optimizer.zero_grad()
 312
 313         output = model(mygpt.BracketedSequence(input)).x
 314         loss = F.cross_entropy(output.transpose(1, 2), input)
 315         acc_train_loss += loss.item() * input.size(0)
 316
 317         nb_train_samples += input.size(0)
 318
 319         loss.backward()
 320
 321         if nb_train_samples % args.batch_size == 0:
 322             optimizer.step()
 323
 324     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 325
 326     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 327
 328
 329 ######################################################################
 330
 331
 332 def run_tests(model, quizz_machine, deterministic_synthesis):
 333     with torch.autograd.no_grad():
 334         model.eval()
 335
 336         nb_test_samples, acc_test_loss = 0, 0.0
 337         nb_samples_accumulated = 0
 338
 339         for input in quizz_machine.batches(split="test"):
 340             input = input.to(device)
 341
 342             bs = model(mygpt.BracketedSequence(input))
 343             output = bs.x
 344
 345             loss = F.cross_entropy(output.transpose(1, 2), input)
 346
 347             acc_test_loss += loss.item() * input.size(0)
 348
 349             nb_test_samples += input.size(0)
 350
 351         main_test_accuracy = quizz_machine.produce_results(
 352             n_epoch=n_epoch,
 353             model=model,
 354             result_dir=args.result_dir,
 355             deterministic_synthesis=deterministic_synthesis,
 356         )
 357
 358         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 359
 360         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 361
 362     model.main_test_accuracy = main_test_accuracy
 363
 364
 365 ######################################################################
 366
 367
 368 def create_c_quizzes(
 369     models,
 370     quizz_machine,
 371     nb_for_train=1000,
 372     nb_for_test=100,
 373     min_ave_seq_logproba=None,
 374 ):
 375     # We will store the generated quizzes for each number of
 376     # correct prediction
 377     recorded = dict([(n, []) for n in range(len(models) + 1)])
 378
 379     model_indexes = []
 380     sum_logits, sum_nb_c_quizzes = 0, 0
 381
 382     def nb_generated():
 383         return sum([sum([x.size(0) for x in recorded[n]]) for n in recorded.keys()])
 384
 385     def nb_validated():
 386         return sum(
 387             [
 388                 sum([x.size(0) for x in recorded[n]])
 389                 for n in range(args.min_to_validate, args.max_to_validate + 1)
 390             ]
 391         )
 392
 393     nb_to_create = nb_for_train + nb_for_test
 394
 395     while nb_validated() < nb_to_create:
 396         (
 397             new_c_quizzes,
 398             nb_correct,
 399             ave_seq_logproba,
 400         ) = quizz_machine.gang_create_c_quizzes(
 401             nb=nb_to_create,
 402             nb_models_for_generation=args.nb_models_for_generation,
 403             models=models,
 404             mode=args.generation_mode,
 405             min_ave_seq_logproba=min_ave_seq_logproba,
 406             n_epoch=n_epoch,
 407             result_dir=args.result_dir,
 408         )
 409
 410         sum_logits += new_c_quizzes.size(0) * ave_seq_logproba
 411         sum_nb_c_quizzes += new_c_quizzes.size(0)
 412
 413         if args.dirty_debug:
 414             nb_correct = torch.randint(
 415                 len(models) + 1, nb_correct.size(), device=new_c_quizzes.device
 416             )
 417
 418         for n in range(nb_correct.max() + 1):
 419             recorded[n].append(new_c_quizzes[nb_correct == n].clone())
 420
 421         log_string(
 422             f"keep c_quizzes {nb_validated()*100/nb_generated():.02f}% kept total {nb_validated()} / {nb_to_create}"
 423         )
 424
 425     # concatenate and shuffle
 426     for n in recorded.keys():
 427         if len(recorded[n]) > 0:
 428             q = torch.cat(recorded[n], dim=0)
 429             q = q[torch.randperm(q.size(0), device=q.device)]
 430             recorded[n] = q
 431         else:
 432             del recorded[n]
 433
 434     new_c_quizzes = torch.cat(
 435         [recorded[n] for n in range(args.min_to_validate, args.max_to_validate + 1)],
 436         dim=0,
 437     )
 438
 439     new_c_quizzes = new_c_quizzes[
 440         torch.randperm(new_c_quizzes.size(0), device=new_c_quizzes.device)[
 441             : nb_for_train + nb_for_test
 442         ]
 443     ]
 444
 445     quizz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 446     quizz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 447
 448     for n in recorded.keys():
 449         s = (
 450             "_validated"
 451             if n >= args.min_to_validate and n <= args.max_to_validate
 452             else ""
 453         )
 454         quizz_machine.problem.save_quizzes(
 455             recorded[n][:72],
 456             args.result_dir,
 457             f"culture_c_quiz_{n_epoch:04d}_N{n}{s}",
 458         )
 459
 460     return sum_logits / sum_nb_c_quizzes
 461
 462
 463 ######################################################################
 464
 465 models = []
 466
 467 for k in range(args.nb_gpts):
 468     model = mygpt.MyGPT(
 469         vocabulary_size=vocabulary_size,
 470         dim_model=args.dim_model,
 471         dim_keys=args.dim_keys,
 472         dim_hidden=args.dim_hidden,
 473         nb_heads=args.nb_heads,
 474         nb_blocks=args.nb_blocks,
 475         causal=True,
 476         dropout=args.dropout,
 477     ).to(device)
 478
 479     model.main_test_accuracy = 0.0
 480     model.id = k
 481
 482     models.append(model)
 483
 484
 485 nb_parameters = sum(p.numel() for p in models[0].parameters())
 486 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 487
 488 ######################################################################
 489
 490 min_ave_seq_logproba = None
 491
 492 for n_epoch in range(args.nb_epochs):
 493     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 494
 495     a = [(model.id, float(model.main_test_accuracy)) for model in models]
 496     a.sort(key=lambda p: p[0])
 497     s = " ".join([f"{p[1]*100:.02f}%" for p in a])
 498     log_string(f"current accuracies {s}")
 499
 500     # select the model with lowest accuracy
 501     models.sort(key=lambda model: model.main_test_accuracy)
 502     model = models[0]
 503
 504     log_string(
 505         f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
 506     )
 507
 508     # improve it
 509     one_epoch(model, quizz_machine)
 510
 511     quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
 512
 513     log_string(
 514         f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
 515     )
 516
 517     # test it
 518     run_tests(model, quizz_machine, deterministic_synthesis=False)
 519
 520     log_string(
 521         f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
 522     )
 523
 524     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
 525         ave_seq_logproba = 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             min_ave_seq_logproba=min_ave_seq_logproba,
 531         )
 532
 533         # We keep the first average logits as a reference
 534         # if min_ave_seq_logproba is None:
 535         # min_ave_seq_logproba = ave_seq_logproba
 536         # else:
 537         # log_string(
 538         # f"min_ave_seq_logproba {min_ave_seq_logproba} ave_seq_logproba {ave_seq_logproba}"
 539         # )
 540
 541         # We update everyone
 542         for model in models:
 543             run_tests(model, quizz_machine, deterministic_synthesis=False)
 544
 545
 546 ######################################################################