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
  16 import mygpt
  17 import sky, grids, quiz_machine
  18
  19 import torch.multiprocessing as mp
  20
  21 # mp.set_start_method('spawn')
  22
  23 # world quizzes vs. culture quizzes
  24
  25 ######################################################################
  26
  27
  28 def log_string(s):
  29     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
  30
  31     if log_file is not None:
  32         log_file.write(t + s + "\n")
  33         log_file.flush()
  34
  35     print(t + s)
  36     sys.stdout.flush()
  37
  38
  39 ######################################################################
  40
  41
  42 def run_tests(model, quiz_machine, deterministic_synthesis, local_device=None):
  43     if local_device is None:
  44         local_device = device
  45
  46     with torch.autograd.no_grad():
  47         model.eval().to(local_device)
  48
  49         nb_test_samples, acc_test_loss = 0, 0.0
  50         nb_samples_accumulated = 0
  51
  52         for input in quiz_machine.batches(model, split="test"):
  53             input = input.to(local_device)
  54
  55             bs = model(mygpt.BracketedSequence(input))
  56             output = bs.x
  57
  58             loss = F.cross_entropy(output.transpose(1, 2), input)
  59
  60             acc_test_loss += loss.item() * input.size(0)
  61
  62             nb_test_samples += input.size(0)
  63
  64         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
  65
  66         log_string(f"test_perplexity {n_epoch} model {model.id} {test_perplexity}")
  67
  68         model.main_test_accuracy = quiz_machine.produce_results(
  69             n_epoch=n_epoch,
  70             model=model,
  71             result_dir=args.result_dir,
  72             deterministic_synthesis=deterministic_synthesis,
  73         )
  74
  75
  76 def one_epoch(model, quiz_machine, local_device=None):
  77     if local_device is None:
  78         local_device = device
  79
  80     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
  81
  82     model.to(local_device).train()
  83
  84     nb_train_samples, acc_train_loss = 0, 0.0
  85
  86     for input in quiz_machine.batches(model, split="train"):
  87         input = input.to(local_device)
  88
  89         if nb_train_samples % args.batch_size == 0:
  90             optimizer.zero_grad()
  91
  92         output = model(mygpt.BracketedSequence(input)).x
  93         loss = F.cross_entropy(output.transpose(1, 2), input)
  94         acc_train_loss += loss.item() * input.size(0)
  95
  96         nb_train_samples += input.size(0)
  97
  98         loss.backward()
  99
 100         if nb_train_samples % args.batch_size == 0:
 101             optimizer.step()
 102
 103     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 104
 105     log_string(f"train_perplexity {n_epoch} model {model.id} {train_perplexity}")
 106
 107     run_tests(model, quiz_machine, deterministic_synthesis=False)
 108
 109
 110 ######################################################################
 111
 112
 113 def standard_validity(logproba):
 114     l = logproba.sort(dim=-1).values
 115     return (l[:, 0] < math.log(0.5)) & (l[:, 1] > math.log(0.99))
 116     # warnings.warn("TEST!!!", RuntimeWarning)
 117     # print(l.exp())
 118     # return (l[:, 0] < math.log(0.99))
 119
 120
 121 def valid_c_quizzes(recorded, criteria):
 122     result = [q[criteria(lp)] for q, lp in recorded]
 123     return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
 124
 125
 126 ######################################################################
 127
 128
 129 def create_c_quizzes(
 130     models,
 131     quiz_machine,
 132     nb_for_train=1000,
 133     nb_for_test=100,
 134 ):
 135     quizzes_and_logproba_records = []
 136
 137     nb_to_create = nb_for_train + nb_for_test
 138
 139     # ------------------------------------------------------------
 140
 141     file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
 142
 143     with open(file_name, "w") as logp_file:
 144         while (
 145             valid_c_quizzes(quizzes_and_logproba_records, standard_validity).size(0)
 146             < nb_to_create
 147         ):
 148             # Select a model at random to generate the new quizzes
 149
 150             model_for_generation = models[torch.randint(len(models), (1,))]
 151
 152             c_quizzes = quiz_machine.generate_quizzes(
 153                 nb_to_create,
 154                 model_for_generation=model_for_generation,
 155                 temperature=args.generation_temperature,
 156             )
 157
 158             c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
 159
 160             if c_quizzes.size(0) > 0:
 161                 logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
 162                 for l in logproba:
 163                     s = " ".join([str(x.item()) for x in l])
 164                     logp_file.write(s + "\n")
 165                 quizzes_and_logproba_records.append((c_quizzes, logproba))
 166
 167             nb_validated = valid_c_quizzes(
 168                 quizzes_and_logproba_records, standard_validity
 169             ).size(0)
 170
 171             log_string(
 172                 f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
 173             )
 174
 175     # store the new c_quizzes which have been validated
 176
 177     new_c_quizzes = valid_c_quizzes(quizzes_and_logproba_records, standard_validity)
 178
 179     quiz_machine.reverse_random_half_in_place(new_c_quizzes)
 180
 181     quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 182     quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 183
 184     # save a bunch of images to investigate what quizzes with a
 185     # certain nb of correct predictions look like
 186
 187     q = new_c_quizzes[:72]
 188
 189     if q.size(0) > 0:
 190         quiz_machine.save_quizzes(args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q)
 191
 192
 193 ######################################################################
 194
 195 if torch.cuda.is_available():
 196     device = torch.device("cuda")
 197     torch.backends.cuda.matmul.allow_tf32 = True
 198 else:
 199     device = torch.device("cpu")
 200
 201 ######################################################################
 202
 203 parser = argparse.ArgumentParser(
 204     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
 205 )
 206
 207 parser.add_argument("--log_filename", type=str, default="train.log")
 208
 209 parser.add_argument("--result_dir", type=str, default=None)
 210
 211 parser.add_argument("--seed", type=int, default=0)
 212
 213 parser.add_argument("--max_percents_of_test_in_train", type=int, default=-1)
 214
 215 ########################################
 216
 217 parser.add_argument("--nb_epochs", type=int, default=10000)
 218
 219 parser.add_argument("--batch_size", type=int, default=None)
 220
 221 parser.add_argument("--physical_batch_size", type=int, default=None)
 222
 223 parser.add_argument("--nb_train_samples", type=int, default=None)
 224
 225 parser.add_argument("--nb_test_samples", type=int, default=None)
 226
 227 parser.add_argument("--learning_rate", type=float, default=5e-4)
 228
 229 ########################################
 230
 231 parser.add_argument("--model", type=str, default=None)
 232
 233 parser.add_argument("--dim_model", type=int, default=None)
 234
 235 parser.add_argument("--dim_keys", type=int, default=None)
 236
 237 parser.add_argument("--dim_hidden", type=int, default=None)
 238
 239 parser.add_argument("--nb_heads", type=int, default=None)
 240
 241 parser.add_argument("--nb_blocks", type=int, default=None)
 242
 243 parser.add_argument("--dropout", type=float, default=0.1)
 244
 245 ########################################
 246
 247 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
 248
 249 parser.add_argument("--problem", type=str, default="grids")
 250
 251 parser.add_argument("--nb_threads", type=int, default=1)
 252
 253 parser.add_argument("--nb_gpus", type=int, default=1)
 254
 255 parser.add_argument("--nb_gpts", type=int, default=5)
 256
 257 parser.add_argument("--min_to_validate", type=int, default=None)
 258
 259 parser.add_argument("--max_to_validate", type=int, default=None)
 260
 261 parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.9)
 262
 263 parser.add_argument("--generation_temperature", type=float, default=2.0)
 264
 265 parser.add_argument("--dirty_debug", action="store_true", default=False)
 266
 267 ######################################################################
 268
 269 parser.add_argument("--sky_height", type=int, default=6)
 270
 271 parser.add_argument("--sky_width", type=int, default=8)
 272
 273 parser.add_argument("--sky_nb_birds", type=int, default=3)
 274
 275 parser.add_argument("--sky_nb_iterations", type=int, default=2)
 276
 277 parser.add_argument("--sky_speed", type=int, default=3)
 278
 279 ######################################################################
 280
 281 args = parser.parse_args()
 282
 283 if args.min_to_validate is None:
 284     args.min_to_validate = args.nb_gpts - 1
 285
 286 if args.max_to_validate is None:
 287     args.max_to_validate = args.nb_gpts - 1
 288
 289 if args.result_dir is None:
 290     args.result_dir = f"results_culture"
 291
 292 ######################################################################
 293
 294 default_args = {
 295     "model": "37M",
 296     "batch_size": 25,
 297     "nb_train_samples": 100000,
 298     "nb_test_samples": 10000,
 299 }
 300
 301 for k, v in default_args.items():
 302     if getattr(args, k) is None:
 303         setattr(args, k, v)
 304
 305 ######################################################################
 306
 307 default_model_args = {
 308     "17K": {
 309         "dim_model": 32,
 310         "dim_keys": 32,
 311         "dim_hidden": 32,
 312         "nb_heads": 2,
 313         "nb_blocks": 2,
 314     },
 315     "4M": {
 316         "dim_model": 256,
 317         "dim_keys": 32,
 318         "dim_hidden": 1024,
 319         "nb_heads": 4,
 320         "nb_blocks": 6,
 321     },
 322     "37M": {
 323         "dim_model": 512,
 324         "dim_keys": 64,
 325         "dim_hidden": 2048,
 326         "nb_heads": 8,
 327         "nb_blocks": 12,
 328     },
 329     "122M": {
 330         "dim_model": 768,
 331         "dim_keys": 64,
 332         "dim_hidden": 2048,
 333         "nb_heads": 8,
 334         "nb_blocks": 24,
 335     },
 336     "352M": {
 337         "dim_model": 1024,
 338         "dim_keys": 64,
 339         "dim_hidden": 2048,
 340         "nb_heads": 8,
 341         "nb_blocks": 48,
 342     },
 343 }
 344
 345 if args.model in default_model_args:
 346     for k, v in default_model_args[args.model].items():
 347         if getattr(args, k) is None:
 348             setattr(args, k, v)
 349 else:
 350     raise ValueError(f"Unknown model {args.model}")
 351
 352 ######################################################################
 353
 354 try:
 355     os.mkdir(args.result_dir)
 356 except FileExistsError:
 357     print(f"result directory {args.result_dir} already exists")
 358     exit(1)
 359
 360 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 361
 362 log_string(f"argv {' '.join(sys.argv)}")
 363
 364 for n in vars(args):
 365     log_string(f"args.{n} {getattr(args, n)}")
 366
 367 if args.seed >= 0:
 368     # torch.backends.cudnn.deterministic = True
 369     # torch.backends.cudnn.benchmark = False
 370     # torch.use_deterministic_algorithms(True)
 371     torch.manual_seed(args.seed)
 372     if torch.cuda.is_available():
 373         torch.cuda.manual_seed_all(args.seed)
 374
 375 ######################################################################
 376
 377 if args.dirty_debug:
 378     args.nb_train_samples = 2500
 379     args.nb_test_samples = 100
 380
 381 if args.physical_batch_size is None:
 382     args.physical_batch_size = args.batch_size
 383 else:
 384     assert args.batch_size % args.physical_batch_size == 0
 385
 386 assert args.nb_train_samples % args.batch_size == 0
 387 assert args.nb_test_samples % args.batch_size == 0
 388
 389 if args.problem == "sky":
 390     problem = sky.Sky(
 391         height=args.sky_height,
 392         width=args.sky_width,
 393         nb_birds=args.sky_nb_birds,
 394         nb_iterations=args.sky_nb_iterations,
 395         speed=args.sky_speed,
 396         max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
 397         chunk_size=100,
 398         nb_threads=args.nb_threads,
 399     )
 400     back_accuracy = False
 401 elif args.problem == "grids":
 402     problem = grids.Grids(
 403         max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
 404         chunk_size=100,
 405         nb_threads=args.nb_threads,
 406     )
 407     back_accuracy = True
 408 else:
 409     raise ValueError
 410
 411 problem.save_some_examples(args.result_dir)
 412
 413 quiz_machine = quiz_machine.QuizMachine(
 414     problem=problem,
 415     nb_train_samples=args.nb_train_samples,
 416     nb_test_samples=args.nb_test_samples,
 417     back_accuracy=back_accuracy,
 418     batch_size=args.physical_batch_size,
 419     result_dir=args.result_dir,
 420     logger=log_string,
 421     device=device,
 422 )
 423
 424 ######################################################################
 425
 426 log_string(f"device {device}")
 427
 428 vocabulary_size = quiz_machine.vocabulary_size()
 429
 430 log_string(f"vocabulary_size {vocabulary_size}")
 431
 432 ######################################################################
 433
 434 models = []
 435
 436 for k in range(args.nb_gpts):
 437     log_string(f"creating model {k} and its w_quizzes")
 438     model = mygpt.MyGPT(
 439         vocabulary_size=vocabulary_size,
 440         dim_model=args.dim_model,
 441         dim_keys=args.dim_keys,
 442         dim_hidden=args.dim_hidden,
 443         nb_heads=args.nb_heads,
 444         nb_blocks=args.nb_blocks,
 445         causal=True,
 446         dropout=args.dropout,
 447     ).to(device)
 448
 449     model.main_test_accuracy = 0.0
 450     model.id = k
 451
 452     model.train_w_quizzes = quiz_machine.generate_token_sequences(
 453         args.nb_train_samples
 454     ).to(device)
 455     quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
 456     model.test_w_quizzes = quiz_machine.generate_token_sequences(
 457         args.nb_test_samples
 458     ).to(device)
 459     quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
 460
 461     models.append(model)
 462
 463
 464 nb_parameters = sum(p.numel() for p in models[0].parameters())
 465 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 466
 467 ######################################################################
 468
 469 # Compute the entropy of the training tokens
 470
 471 token_count = 0
 472 for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
 473     token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum(
 474         (0, 1)
 475     )
 476 token_probas = token_count / token_count.sum()
 477 entropy = -torch.xlogy(token_probas, token_probas).sum()
 478 train_set_perplexity = math.exp(entropy)
 479
 480 ######################################################################
 481 # A bit of paranoia never hurts
 482
 483 if args.max_percents_of_test_in_train >= 0:
 484
 485     def subsets_as_tuples(batches, cs):
 486         s = set()
 487         for batch in batches:
 488             for x in batch:
 489                 s.add(tuple([v.item() for v in x]))
 490                 if len(s) == cs:
 491                     yield s
 492                     s = set()
 493         yield s
 494
 495     nb_test, nb_in_train = 0, 0
 496     for test_subset in subsets_as_tuples(
 497         quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
 498     ):
 499         in_train = set()
 500         for train_subset in subsets_as_tuples(
 501             quiz_machine.batches(models[0], split="train", desc="train-check"), 25000
 502         ):
 503             in_train.update(test_subset.intersection(train_subset))
 504         nb_in_train += len(in_train)
 505         nb_test += len(test_subset)
 506
 507     log_string(
 508         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 509     )
 510
 511     assert (
 512         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 513     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 514
 515 ######################################################################
 516
 517 nb_new_c_quizzes_for_train = args.nb_train_samples // 50
 518 nb_new_c_quizzes_for_test = args.nb_test_samples // 50
 519
 520 log_string(
 521     f"nb_new_c_quizzes_for_train {nb_new_c_quizzes_for_train} nb_new_c_quizzes_for_test {nb_new_c_quizzes_for_test}"
 522 )
 523
 524 ######################################################################
 525
 526 if args.dirty_debug:
 527     args.accuracy_to_make_c_quizzes = 0.0
 528     args.nb_gpts = 2
 529     nb_new_c_quizzes_for_train = 100
 530     nb_new_c_quizzes_for_test = 10
 531
 532     def standard_validity(logproba):
 533         l = logproba.sort(dim=-1).values
 534         return l[:, 0] < math.log(0.99)
 535
 536
 537 ######################################################################
 538
 539 for n_epoch in range(args.nb_epochs):
 540     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 541
 542     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
 543     log_string(f"current_test_accuracies {cta}")
 544
 545     ##################################################
 546     # Select, improve, and eval the worst models
 547
 548     ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
 549
 550     weakest_models = ranked_models[: args.nb_gpus]
 551
 552     processes = []
 553
 554     for gpu_id, model in enumerate(weakest_models):
 555         log_string(
 556             f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
 557         )
 558
 559         process = mp.Process(
 560             target=one_epoch, args=(model, quiz_machine, f"cuda:{gpu_id}")
 561         )
 562
 563         processes.append(process)
 564
 565     for process in processes:
 566         process.start()
 567
 568     for process in processes:
 569         process.join()
 570
 571     ##################################################
 572     # Renew the train sets
 573
 574     log_string(
 575         f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
 576     )
 577
 578     for model in weakest_models:
 579         quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
 580
 581     ##################################################
 582     # If all the models are good enough, generate new quizzes and
 583     # re-compute the test errors
 584
 585     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
 586         create_c_quizzes(
 587             models,
 588             quiz_machine,
 589             nb_for_train=nb_new_c_quizzes_for_train,
 590             nb_for_test=nb_new_c_quizzes_for_test,
 591         )
 592
 593 ######################################################################