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 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", help=" ")
  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("--nb_gpts", type=int, default=5)
  83
  84 parser.add_argument("--nb_models_for_generation", type=int, default=1)
  85
  86 parser.add_argument("--generation_mode", type=str, default="groupthink")
  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 ######################################################################
  97
  98 args = parser.parse_args()
  99
 100 if args.result_dir is None:
 101     args.result_dir = f"results_culture"
 102
 103 ######################################################################
 104
 105 if args.dirty_debug:
 106     args.accuracy_to_make_c_quizzes = 0.0
 107     nb_new_c_quizzes_for_train = 100
 108     nb_new_c_quizzes_for_test = 10
 109
 110 ######################################################################
 111
 112 default_args = {
 113     "model": "37M",
 114     "batch_size": 100,
 115     "nb_train_samples": 100000,
 116     "nb_test_samples": 10000,
 117 }
 118
 119 for k, v in default_args.items():
 120     if getattr(args, k) is None:
 121         setattr(args, k, v)
 122
 123 ######################################################################
 124
 125 default_model_args = {
 126     "17K": {
 127         "dim_model": 32,
 128         "dim_keys": 32,
 129         "dim_hidden": 32,
 130         "nb_heads": 2,
 131         "nb_blocks": 2,
 132     },
 133     "4M": {
 134         "dim_model": 256,
 135         "dim_keys": 32,
 136         "dim_hidden": 1024,
 137         "nb_heads": 4,
 138         "nb_blocks": 6,
 139     },
 140     "37M": {
 141         "dim_model": 512,
 142         "dim_keys": 64,
 143         "dim_hidden": 2048,
 144         "nb_heads": 8,
 145         "nb_blocks": 12,
 146     },
 147     "122M": {
 148         "dim_model": 768,
 149         "dim_keys": 64,
 150         "dim_hidden": 2048,
 151         "nb_heads": 8,
 152         "nb_blocks": 24,
 153     },
 154     "352M": {
 155         "dim_model": 1024,
 156         "dim_keys": 64,
 157         "dim_hidden": 2048,
 158         "nb_heads": 8,
 159         "nb_blocks": 48,
 160     },
 161 }
 162
 163 if args.model in default_model_args:
 164     for k, v in default_model_args[args.model].items():
 165         if getattr(args, k) is None:
 166             setattr(args, k, v)
 167 else:
 168     raise ValueError(f"Unknown model {args.model}")
 169
 170 ######################################################################
 171
 172 try:
 173     os.mkdir(args.result_dir)
 174 except FileExistsError:
 175     print(f"result directory {args.result_dir} already exists")
 176     exit(1)
 177
 178 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 179
 180 if args.seed >= 0:
 181     # torch.backends.cudnn.deterministic = True
 182     # torch.backends.cudnn.benchmark = False
 183     # torch.use_deterministic_algorithms(True)
 184     torch.manual_seed(args.seed)
 185     if torch.cuda.is_available():
 186         torch.cuda.manual_seed_all(args.seed)
 187
 188 ######################################################################
 189
 190
 191 def log_string(s):
 192     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 193
 194     if log_file is not None:
 195         log_file.write(t + s + "\n")
 196         log_file.flush()
 197
 198     print(t + s)
 199     sys.stdout.flush()
 200
 201
 202 log_string(f"argv {' '.join(sys.argv)}")
 203
 204 for n in vars(args):
 205     log_string(f"args.{n} {getattr(args, n)}")
 206
 207
 208 ######################################################################
 209
 210 if args.dirty_debug:
 211     args.nb_train_samples = 2500
 212     args.nb_test_samples = 100
 213
 214 if args.physical_batch_size is None:
 215     args.physical_batch_size = args.batch_size
 216 else:
 217     assert args.batch_size % args.physical_batch_size == 0
 218
 219 assert args.nb_train_samples % args.batch_size == 0
 220 assert args.nb_test_samples % args.batch_size == 0
 221
 222 quizz_machine = quizz_machine.QuizzMachine(
 223     problem=sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2, speed=2),
 224     nb_train_samples=args.nb_train_samples,
 225     nb_test_samples=args.nb_test_samples,
 226     batch_size=args.physical_batch_size,
 227     result_dir=args.result_dir,
 228     logger=log_string,
 229     device=device,
 230 )
 231
 232 ######################################################################
 233
 234 log_string(f"device {device}")
 235
 236 vocabulary_size = quizz_machine.vocabulary_size()
 237
 238 log_string(f"vocabulary_size {vocabulary_size}")
 239
 240 ######################################################################
 241
 242 # Compute the entropy of the training tokens
 243
 244 token_count = 0
 245 for input in quizz_machine.batches(split="train", desc="train-entropy"):
 246     token_count += F.one_hot(input, num_classes=quizz_machine.vocabulary_size()).sum(
 247         (0, 1)
 248     )
 249 token_probas = token_count / token_count.sum()
 250 entropy = -torch.xlogy(token_probas, token_probas).sum()
 251 train_set_perplexity = math.exp(entropy)
 252
 253 ######################################################################
 254 # A bit of paranoia never hurts
 255
 256 if args.max_percents_of_test_in_train >= 0:
 257
 258     def subsets_as_tuples(batches, cs):
 259         s = set()
 260         for batch in batches:
 261             for x in batch:
 262                 s.add(tuple([v.item() for v in x]))
 263                 if len(s) == cs:
 264                     yield s
 265                     s = set()
 266         yield s
 267
 268     nb_test, nb_in_train = 0, 0
 269     for test_subset in subsets_as_tuples(
 270         quizz_machine.batches(split="test", desc="test-check"), 25000
 271     ):
 272         in_train = set()
 273         for train_subset in subsets_as_tuples(
 274             quizz_machine.batches(split="train", desc="train-check"), 25000
 275         ):
 276             in_train.update(test_subset.intersection(train_subset))
 277         nb_in_train += len(in_train)
 278         nb_test += len(test_subset)
 279
 280     log_string(
 281         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 282     )
 283
 284     assert (
 285         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 286     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 287
 288 ##############################
 289
 290
 291 def one_epoch(model, quizz_machine):
 292     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 293
 294     model.train()
 295
 296     nb_train_samples, acc_train_loss = 0, 0.0
 297
 298     for input in quizz_machine.batches(split="train"):
 299         input = input.to(device)
 300
 301         if nb_train_samples % args.batch_size == 0:
 302             optimizer.zero_grad()
 303
 304         output = model(mygpt.BracketedSequence(input)).x
 305         loss = F.cross_entropy(output.transpose(1, 2), input)
 306         acc_train_loss += loss.item() * input.size(0)
 307
 308         nb_train_samples += input.size(0)
 309
 310         loss.backward()
 311
 312         if nb_train_samples % args.batch_size == 0:
 313             optimizer.step()
 314
 315     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 316
 317     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 318
 319
 320 ######################################################################
 321
 322
 323 def run_tests(model, quizz_machine, deterministic_synthesis):
 324     with torch.autograd.no_grad():
 325         model.eval()
 326
 327         nb_test_samples, acc_test_loss = 0, 0.0
 328         nb_samples_accumulated = 0
 329
 330         for input in quizz_machine.batches(split="test"):
 331             input = input.to(device)
 332
 333             bs = model(mygpt.BracketedSequence(input))
 334             output = bs.x
 335
 336             loss = F.cross_entropy(output.transpose(1, 2), input)
 337
 338             acc_test_loss += loss.item() * input.size(0)
 339
 340             nb_test_samples += input.size(0)
 341
 342         main_test_accuracy = quizz_machine.produce_results(
 343             n_epoch=n_epoch,
 344             model=model,
 345             result_dir=args.result_dir,
 346             deterministic_synthesis=deterministic_synthesis,
 347         )
 348
 349         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 350
 351         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 352
 353     model.main_test_accuracy = main_test_accuracy
 354
 355
 356 ######################################################################
 357
 358
 359 def create_c_quizzes(
 360     models,
 361     quizz_machine,
 362     nb_for_train=1000,
 363     nb_for_test=100,
 364     min_ave_seq_logproba=None,
 365 ):
 366     # We will store the generated quizzes for each number of
 367     # correct prediction
 368     recorded = dict([(n, []) for n in range(len(models) + 1)])
 369
 370     model_indexes = []
 371     sum_logits, sum_nb_c_quizzes = 0, 0
 372
 373     def nb_generated():
 374         return sum([sum([x.size(0) for x in recorded[n]]) for n in recorded.keys()])
 375
 376     def nb_validated():
 377         return sum(
 378             [
 379                 sum([x.size(0) for x in recorded[n]])
 380                 for n in range(args.min_to_validate, args.max_to_validate + 1)
 381             ]
 382         )
 383
 384     nb_to_create = nb_for_train + nb_for_test
 385
 386     while nb_validated() < nb_to_create:
 387         (
 388             new_c_quizzes,
 389             nb_correct,
 390             ave_seq_logproba,
 391         ) = quizz_machine.gang_create_c_quizzes(
 392             nb=nb_to_create,
 393             nb_models_for_generation=args.nb_models_for_generation,
 394             models=models,
 395             mode=args.generation_mode,
 396             min_ave_seq_logproba=min_ave_seq_logproba,
 397             n_epoch=n_epoch,
 398             result_dir=args.result_dir,
 399         )
 400
 401         sum_logits += new_c_quizzes.size(0) * ave_seq_logproba
 402         sum_nb_c_quizzes += new_c_quizzes.size(0)
 403
 404         if args.dirty_debug:
 405             nb_correct = torch.randint(
 406                 len(models) + 1, nb_correct.size(), device=new_c_quizzes.device
 407             )
 408
 409         for n in range(nb_correct.max() + 1):
 410             recorded[n].append(new_c_quizzes[nb_correct == n].clone())
 411
 412         log_string(
 413             f"keep c_quizzes {nb_validated()*100/nb_generated():.02f}% kept total {nb_validated()} / {nb_to_create}"
 414         )
 415
 416     # concatenate and shuffle
 417     for n in recorded.keys():
 418         if len(recorded[n]) > 0:
 419             q = torch.cat(recorded[n], dim=0)
 420             q = q[torch.randperm(q.size(0), device=q.device)]
 421             recorded[n] = q
 422         else:
 423             del recorded[n]
 424
 425     new_c_quizzes = torch.cat(
 426         [recorded[n] for n in range(args.min_to_validate, args.max_to_validate + 1)],
 427         dim=0,
 428     )
 429
 430     new_c_quizzes = new_c_quizzes[
 431         torch.randperm(new_c_quizzes.size(0), device=new_c_quizzes.device)[
 432             : nb_for_train + nb_for_test
 433         ]
 434     ]
 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     for n in recorded.keys():
 440         s = (
 441             "_validated"
 442             if n >= args.min_to_validate and n <= args.max_to_validate
 443             else ""
 444         )
 445         quizz_machine.problem.save_quizzes(
 446             recorded[n][:72],
 447             args.result_dir,
 448             f"culture_c_quiz_{n_epoch:04d}_N{n}{s}",
 449         )
 450
 451     return sum_logits / sum_nb_c_quizzes
 452
 453
 454 ######################################################################
 455
 456 models = []
 457
 458 for k in range(args.nb_gpts):
 459     model = mygpt.MyGPT(
 460         vocabulary_size=vocabulary_size,
 461         dim_model=args.dim_model,
 462         dim_keys=args.dim_keys,
 463         dim_hidden=args.dim_hidden,
 464         nb_heads=args.nb_heads,
 465         nb_blocks=args.nb_blocks,
 466         causal=True,
 467         dropout=args.dropout,
 468     ).to(device)
 469
 470     model.main_test_accuracy = 0.0
 471     model.id = k
 472
 473     models.append(model)
 474
 475
 476 nb_parameters = sum(p.numel() for p in models[0].parameters())
 477 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 478
 479 ######################################################################
 480
 481 min_ave_seq_logproba = None
 482
 483 for n_epoch in range(args.nb_epochs):
 484     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 485
 486     a = [(model.id, float(model.main_test_accuracy)) for model in models]
 487     a.sort(key=lambda p: p[0])
 488     s = " ".join([f"{p[1]*100:.02f}%" for p in a])
 489     log_string(f"current accuracies {s}")
 490
 491     # select the model with lowest accuracy
 492     models.sort(key=lambda model: model.main_test_accuracy)
 493     model = models[0]
 494
 495     log_string(
 496         f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
 497     )
 498
 499     # improve it
 500     one_epoch(model, quizz_machine)
 501
 502     quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
 503
 504     log_string(
 505         f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
 506     )
 507
 508     # test it
 509     run_tests(model, quizz_machine, deterministic_synthesis=False)
 510
 511     log_string(
 512         f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
 513     )
 514
 515     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
 516         ave_seq_logproba = create_c_quizzes(
 517             models,
 518             quizz_machine,
 519             nb_for_train=nb_new_c_quizzes_for_train,
 520             nb_for_test=nb_new_c_quizzes_for_test,
 521             min_ave_seq_logproba=min_ave_seq_logproba,
 522         )
 523
 524         # We keep the first average logits as a reference
 525         # if min_ave_seq_logproba is None:
 526         # min_ave_seq_logproba = ave_seq_logproba
 527         # else:
 528         # log_string(
 529         # f"min_ave_seq_logproba {min_ave_seq_logproba} ave_seq_logproba {ave_seq_logproba}"
 530         # )
 531
 532         # We update everyone
 533         for model in models:
 534             run_tests(model, quizz_machine, deterministic_synthesis=False)
 535
 536
 537 ######################################################################