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, quizz_machine
  17
  18 # world quizzes vs. culture quizzes
  19
  20 ######################################################################
  21
  22 accuracy_to_make_c_quizzes = 0.975
  23 nb_new_c_quizzes_for_train = 1000
  24 nb_new_c_quizzes_for_test = 100
  25
  26 ######################################################################
  27
  28 if torch.cuda.is_available():
  29     device = torch.device("cuda")
  30     torch.backends.cuda.matmul.allow_tf32 = True
  31 else:
  32     device = torch.device("cpu")
  33
  34 ######################################################################
  35
  36 parser = argparse.ArgumentParser(
  37     description="An implementation of GPT with cache.",
  38     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
  39 )
  40
  41 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
  42
  43 parser.add_argument("--result_dir", type=str, default=None)
  44
  45 parser.add_argument("--seed", type=int, default=0)
  46
  47 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
  48
  49 ########################################
  50
  51 parser.add_argument("--nb_epochs", type=int, default=10000)
  52
  53 parser.add_argument("--batch_size", type=int, default=None)
  54
  55 parser.add_argument("--physical_batch_size", type=int, default=None)
  56
  57 parser.add_argument("--nb_train_samples", type=int, default=None)
  58
  59 parser.add_argument("--nb_test_samples", type=int, default=None)
  60
  61 parser.add_argument("--learning_rate", type=float, default=1e-3)
  62
  63 ########################################
  64
  65 parser.add_argument("--model", type=str, default=None)
  66
  67 parser.add_argument("--dim_model", type=int, default=None)
  68
  69 parser.add_argument("--dim_keys", type=int, default=None)
  70
  71 parser.add_argument("--dim_hidden", type=int, default=None)
  72
  73 parser.add_argument("--nb_heads", type=int, default=None)
  74
  75 parser.add_argument("--nb_blocks", type=int, default=None)
  76
  77 parser.add_argument("--dropout", type=float, default=0.1)
  78
  79 ########################################
  80
  81 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  82
  83 parser.add_argument("--nb_gpts", type=int, default=5)
  84
  85 parser.add_argument("--nb_correct_to_validate", type=int, default=4)
  86
  87 parser.add_argument("--dirty_debug", action="store_true", default=False)
  88
  89 ######################################################################
  90
  91 args = parser.parse_args()
  92
  93 if args.result_dir is None:
  94     args.result_dir = f"results_culture"
  95
  96 ######################################################################
  97
  98 if args.dirty_debug:
  99     accuracy_to_make_c_quizzes = 0.0
 100     nb_new_c_quizzes_for_train = 100
 101     nb_new_c_quizzes_for_test = 10
 102
 103 ######################################################################
 104
 105 default_args = {
 106     "model": "37M",
 107     "batch_size": 100,
 108     "nb_train_samples": 100000,
 109     "nb_test_samples": 10000,
 110 }
 111
 112 for k, v in default_args.items():
 113     if getattr(args, k) is None:
 114         setattr(args, k, v)
 115
 116 ######################################################################
 117
 118 default_model_args = {
 119     "17K": {
 120         "dim_model": 32,
 121         "dim_keys": 32,
 122         "dim_hidden": 32,
 123         "nb_heads": 2,
 124         "nb_blocks": 2,
 125     },
 126     "4M": {
 127         "dim_model": 256,
 128         "dim_keys": 32,
 129         "dim_hidden": 1024,
 130         "nb_heads": 4,
 131         "nb_blocks": 6,
 132     },
 133     "37M": {
 134         "dim_model": 512,
 135         "dim_keys": 64,
 136         "dim_hidden": 2048,
 137         "nb_heads": 8,
 138         "nb_blocks": 12,
 139     },
 140     "122M": {
 141         "dim_model": 768,
 142         "dim_keys": 64,
 143         "dim_hidden": 2048,
 144         "nb_heads": 8,
 145         "nb_blocks": 24,
 146     },
 147     "352M": {
 148         "dim_model": 1024,
 149         "dim_keys": 64,
 150         "dim_hidden": 2048,
 151         "nb_heads": 8,
 152         "nb_blocks": 48,
 153     },
 154 }
 155
 156 if args.model in default_model_args:
 157     for k, v in default_model_args[args.model].items():
 158         if getattr(args, k) is None:
 159             setattr(args, k, v)
 160 else:
 161     raise ValueError(f"Unknown model {args.model}")
 162
 163 ######################################################################
 164
 165 try:
 166     os.mkdir(args.result_dir)
 167 except FileExistsError:
 168     print(f"result directory {args.result_dir} already exists")
 169     exit(1)
 170
 171 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 172
 173 if args.seed >= 0:
 174     # torch.backends.cudnn.deterministic = True
 175     # torch.backends.cudnn.benchmark = False
 176     # torch.use_deterministic_algorithms(True)
 177     torch.manual_seed(args.seed)
 178     if torch.cuda.is_available():
 179         torch.cuda.manual_seed_all(args.seed)
 180
 181 ######################################################################
 182
 183
 184 def log_string(s):
 185     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 186
 187     if log_file is not None:
 188         log_file.write(t + s + "\n")
 189         log_file.flush()
 190
 191     print(t + s)
 192     sys.stdout.flush()
 193
 194
 195 log_string(f"argv {' '.join(sys.argv)}")
 196
 197 for n in vars(args):
 198     log_string(f"args.{n} {getattr(args, n)}")
 199
 200
 201 ######################################################################
 202
 203 if args.dirty_debug:
 204     args.nb_train_samples = 2500
 205     args.nb_test_samples = 100
 206
 207 if args.physical_batch_size is None:
 208     args.physical_batch_size = args.batch_size
 209 else:
 210     assert args.batch_size % args.physical_batch_size == 0
 211
 212 assert args.nb_train_samples % args.batch_size == 0
 213 assert args.nb_test_samples % args.batch_size == 0
 214
 215 quizz_machine = quizz_machine.QuizzMachine(
 216     problem=sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2, speed=2),
 217     nb_train_samples=args.nb_train_samples,
 218     nb_test_samples=args.nb_test_samples,
 219     batch_size=args.physical_batch_size,
 220     result_dir=args.result_dir,
 221     logger=log_string,
 222     device=device,
 223 )
 224
 225 ######################################################################
 226
 227 log_string(f"device {device}")
 228
 229 vocabulary_size = quizz_machine.vocabulary_size()
 230
 231 log_string(f"vocabulary_size {vocabulary_size}")
 232
 233 ######################################################################
 234
 235 # Compute the entropy of the training tokens
 236
 237 token_count = 0
 238 for input in quizz_machine.batches(split="train", desc="train-entropy"):
 239     token_count += F.one_hot(input, num_classes=quizz_machine.vocabulary_size()).sum(
 240         (0, 1)
 241     )
 242 token_probas = token_count / token_count.sum()
 243 entropy = -torch.xlogy(token_probas, token_probas).sum()
 244 train_set_perplexity = math.exp(entropy)
 245
 246 ######################################################################
 247 # A bit of paranoia never hurts
 248
 249 if args.max_percents_of_test_in_train >= 0:
 250
 251     def subsets_as_tuples(batches, cs):
 252         s = set()
 253         for batch in batches:
 254             for x in batch:
 255                 s.add(tuple([v.item() for v in x]))
 256                 if len(s) == cs:
 257                     yield s
 258                     s = set()
 259         yield s
 260
 261     nb_test, nb_in_train = 0, 0
 262     for test_subset in subsets_as_tuples(
 263         quizz_machine.batches(split="test", desc="test-check"), 25000
 264     ):
 265         in_train = set()
 266         for train_subset in subsets_as_tuples(
 267             quizz_machine.batches(split="train", desc="train-check"), 25000
 268         ):
 269             in_train.update(test_subset.intersection(train_subset))
 270         nb_in_train += len(in_train)
 271         nb_test += len(test_subset)
 272
 273     log_string(
 274         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 275     )
 276
 277     assert (
 278         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 279     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 280
 281 ##############################
 282
 283
 284 def one_epoch(model, quizz_machine):
 285     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 286
 287     model.train()
 288
 289     nb_train_samples, acc_train_loss = 0, 0.0
 290
 291     for input in quizz_machine.batches(split="train"):
 292         input = input.to(device)
 293
 294         if nb_train_samples % args.batch_size == 0:
 295             optimizer.zero_grad()
 296
 297         output = model(mygpt.BracketedSequence(input)).x
 298         loss = F.cross_entropy(output.transpose(1, 2), input)
 299         acc_train_loss += loss.item() * input.size(0)
 300
 301         nb_train_samples += input.size(0)
 302
 303         loss.backward()
 304
 305         if nb_train_samples % args.batch_size == 0:
 306             optimizer.step()
 307
 308     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 309
 310     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 311
 312
 313 ######################################################################
 314
 315
 316 def run_tests(model, quizz_machine, deterministic_synthesis):
 317     with torch.autograd.no_grad():
 318         model.eval()
 319
 320         nb_test_samples, acc_test_loss = 0, 0.0
 321         nb_samples_accumulated = 0
 322
 323         for input in quizz_machine.batches(split="test"):
 324             input = input.to(device)
 325
 326             bs = model(mygpt.BracketedSequence(input))
 327             output = bs.x
 328
 329             loss = F.cross_entropy(output.transpose(1, 2), input)
 330
 331             acc_test_loss += loss.item() * input.size(0)
 332
 333             nb_test_samples += input.size(0)
 334
 335         main_test_accuracy = quizz_machine.produce_results(
 336             n_epoch=n_epoch,
 337             model=model,
 338             result_dir=args.result_dir,
 339             logger=log_string,
 340             deterministic_synthesis=deterministic_synthesis,
 341         )
 342
 343         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 344
 345         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 346
 347     model.main_test_accuracy = main_test_accuracy
 348
 349
 350 ######################################################################
 351
 352
 353 def create_c_quizzes(
 354     models,
 355     quizz_machine,
 356     nb_for_train=1000,
 357     nb_for_test=100,
 358     min_ave_seq_logproba=None,
 359 ):
 360     # We will store the generated quizzes for each number of
 361     # correct prediction
 362     recorded = dict([(n, []) for n in range(len(models) + 1)])
 363
 364     model_indexes = []
 365     sum_logits, sum_nb_c_quizzes = 0, 0
 366
 367     def nb_generated():
 368         return sum([sum([x.size(0) for x in recorded[n]]) for n in recorded.keys()])
 369
 370     def nb_validated():
 371         return sum(
 372             [
 373                 sum([x.size(0) for x in recorded[n]])
 374                 for n in range(args.nb_correct_to_validate, len(models))
 375             ]
 376         )
 377
 378     while nb_validated() < nb_for_train + nb_for_test:
 379         nb_to_validate = nb_for_train + nb_for_test
 380
 381         if len(model_indexes) == 0:
 382             model_indexes = [i.item() for i in torch.randperm(len(models))]
 383
 384         model = models[model_indexes.pop()]
 385
 386         new_c_quizzes, nb_correct, ave_seq_logproba = quizz_machine.create_c_quizzes(
 387             nb=nb_to_validate,
 388             model_for_generation=model,
 389             models_for_validation=models,
 390             min_ave_seq_logproba=min_ave_seq_logproba,
 391             n_epoch=n_epoch,
 392             result_dir=args.result_dir,
 393             logger=log_string,
 394         )
 395
 396         sum_logits += new_c_quizzes.size(0) * ave_seq_logproba
 397         sum_nb_c_quizzes += new_c_quizzes.size(0)
 398
 399         if args.dirty_debug:
 400             nb_correct = torch.randint(
 401                 len(models) + 1, nb_correct.size(), device=new_c_quizzes.device
 402             )
 403
 404         for n in range(nb_correct.max() + 1):
 405             recorded[n].append(new_c_quizzes[nb_correct == n].clone())
 406
 407         log_string(
 408             f"keep c_quizzes {nb_validated()*100/nb_generated():.02f}% kept total {nb_validated()} / {nb_to_validate}"
 409         )
 410
 411     # concatenate and shuffle
 412     for n in recorded.keys():
 413         if len(recorded[n]) > 0:
 414             q = torch.cat(recorded[n], dim=0)
 415             q = q[torch.randperm(q.size(0), device=q.device)]
 416             recorded[n] = q
 417         else:
 418             del recorded[n]
 419
 420     new_c_quizzes = torch.cat(
 421         [recorded[n] for n in range(args.nb_correct_to_validate, len(models))], dim=0
 422     )
 423
 424     new_c_quizzes = new_c_quizzes[
 425         torch.randperm(new_c_quizzes.size(0), device=new_c_quizzes.device)[
 426             : nb_for_train + nb_for_test
 427         ]
 428     ]
 429
 430     quizz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 431     quizz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 432
 433     for n in recorded.keys():
 434         s = "_validated" if n >= args.nb_correct_to_validate and n < len(models) else ""
 435         quizz_machine.problem.save_quizzes(
 436             recorded[n][:72],
 437             args.result_dir,
 438             f"culture_c_quiz_{n_epoch:04d}_N{n}{s}",
 439         )
 440
 441     return sum_logits / sum_nb_c_quizzes
 442
 443
 444 ######################################################################
 445
 446 models = []
 447
 448 for k in range(args.nb_gpts):
 449     model = mygpt.MyGPT(
 450         vocabulary_size=vocabulary_size,
 451         dim_model=args.dim_model,
 452         dim_keys=args.dim_keys,
 453         dim_hidden=args.dim_hidden,
 454         nb_heads=args.nb_heads,
 455         nb_blocks=args.nb_blocks,
 456         causal=True,
 457         dropout=args.dropout,
 458     ).to(device)
 459
 460     model.main_test_accuracy = 0.0
 461     model.id = k
 462
 463     models.append(model)
 464
 465
 466 nb_parameters = sum(p.numel() for p in models[0].parameters())
 467 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 468
 469 ######################################################################
 470
 471 min_ave_seq_logproba = None
 472
 473 for n_epoch in range(args.nb_epochs):
 474     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 475
 476     a = [(model.id, float(model.main_test_accuracy)) for model in models]
 477     a.sort(key=lambda p: p[0])
 478     log_string(f"current accuracies {a}")
 479
 480     # select the model with lowest accuracy
 481     models.sort(key=lambda model: model.main_test_accuracy)
 482     model = models[0]
 483
 484     log_string(
 485         f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
 486     )
 487
 488     # improve it
 489     one_epoch(model, quizz_machine)
 490
 491     quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
 492
 493     log_string(
 494         f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
 495     )
 496
 497     # test it
 498     run_tests(model, quizz_machine, deterministic_synthesis=False)
 499
 500     log_string(
 501         f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
 502     )
 503
 504     if min([m.main_test_accuracy for m in models]) >= accuracy_to_make_c_quizzes:
 505         ave_seq_logproba = create_c_quizzes(
 506             models,
 507             quizz_machine,
 508             nb_for_train=nb_new_c_quizzes_for_train,
 509             nb_for_test=nb_new_c_quizzes_for_test,
 510             min_ave_seq_logproba=min_ave_seq_logproba,
 511         )
 512
 513         # We keep the first average logits as a reference
 514         # if min_ave_seq_logproba is None:
 515         # min_ave_seq_logproba = ave_seq_logproba
 516         # else:
 517         # log_string(
 518         # f"min_ave_seq_logproba {min_ave_seq_logproba} ave_seq_logproba {ave_seq_logproba}"
 519         # )
 520
 521         # We update everyone
 522         for model in models:
 523             run_tests(model, quizz_machine, deterministic_synthesis=False)
 524
 525
 526 ######################################################################