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-4)
  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("--dirty_debug", action="store_true", default=False)
  86
  87 ######################################################################
  88
  89 args = parser.parse_args()
  90
  91 if args.result_dir is None:
  92     args.result_dir = f"results_culture"
  93
  94 ######################################################################
  95
  96 if args.dirty_debug:
  97     accuracy_to_make_c_quizzes = 0.0
  98     nb_new_c_quizzes_for_train = 100
  99     nb_new_c_quizzes_for_test = 10
 100
 101 ######################################################################
 102
 103 default_args = {
 104     "model": "37M",
 105     "batch_size": 100,
 106     "nb_train_samples": 250000,
 107     "nb_test_samples": 10000,
 108 }
 109
 110 for k, v in default_args.items():
 111     if getattr(args, k) is None:
 112         setattr(args, k, v)
 113
 114 ######################################################################
 115
 116 default_model_args = {
 117     "17K": {
 118         "dim_model": 32,
 119         "dim_keys": 32,
 120         "dim_hidden": 32,
 121         "nb_heads": 2,
 122         "nb_blocks": 2,
 123     },
 124     "4M": {
 125         "dim_model": 256,
 126         "dim_keys": 32,
 127         "dim_hidden": 1024,
 128         "nb_heads": 4,
 129         "nb_blocks": 6,
 130     },
 131     "37M": {
 132         "dim_model": 512,
 133         "dim_keys": 64,
 134         "dim_hidden": 2048,
 135         "nb_heads": 8,
 136         "nb_blocks": 12,
 137     },
 138     "122M": {
 139         "dim_model": 768,
 140         "dim_keys": 64,
 141         "dim_hidden": 2048,
 142         "nb_heads": 8,
 143         "nb_blocks": 24,
 144     },
 145     "352M": {
 146         "dim_model": 1024,
 147         "dim_keys": 64,
 148         "dim_hidden": 2048,
 149         "nb_heads": 8,
 150         "nb_blocks": 48,
 151     },
 152 }
 153
 154 if args.model in default_model_args:
 155     for k, v in default_model_args[args.model].items():
 156         if getattr(args, k) is None:
 157             setattr(args, k, v)
 158 else:
 159     raise ValueError(f"Unknown model {args.model}")
 160
 161 ######################################################################
 162
 163 try:
 164     os.mkdir(args.result_dir)
 165 except FileExistsError:
 166     print(f"result directory {args.result_dir} already exists")
 167     exit(1)
 168
 169 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 170
 171 if args.seed >= 0:
 172     # torch.backends.cudnn.deterministic = True
 173     # torch.backends.cudnn.benchmark = False
 174     # torch.use_deterministic_algorithms(True)
 175     torch.manual_seed(args.seed)
 176     if torch.cuda.is_available():
 177         torch.cuda.manual_seed_all(args.seed)
 178
 179 ######################################################################
 180
 181
 182 def log_string(s):
 183     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 184
 185     if log_file is not None:
 186         log_file.write(t + s + "\n")
 187         log_file.flush()
 188
 189     print(t + s)
 190     sys.stdout.flush()
 191
 192
 193 log_string(f"argv {' '.join(sys.argv)}")
 194
 195 for n in vars(args):
 196     log_string(f"args.{n} {getattr(args, n)}")
 197
 198
 199 ######################################################################
 200
 201 if args.dirty_debug:
 202     args.nb_train_samples = 2500
 203     args.nb_test_samples = 100
 204
 205 if args.physical_batch_size is None:
 206     args.physical_batch_size = args.batch_size
 207 else:
 208     assert args.batch_size % args.physical_batch_size == 0
 209
 210 assert args.nb_train_samples % args.batch_size == 0
 211 assert args.nb_test_samples % args.batch_size == 0
 212
 213 quizz_machine = quizz_machine.QuizzMachine(
 214     sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2),
 215     nb_train_samples=args.nb_train_samples,
 216     nb_test_samples=args.nb_test_samples,
 217     batch_size=args.physical_batch_size,
 218     result_dir=args.result_dir,
 219     logger=log_string,
 220     device=device,
 221 )
 222
 223 ######################################################################
 224
 225 log_string(f"device {device}")
 226
 227 vocabulary_size = quizz_machine.vocabulary_size()
 228
 229 log_string(f"vocabulary_size {vocabulary_size}")
 230
 231 ######################################################################
 232
 233 # Compute the entropy of the training tokens
 234
 235 token_count = 0
 236 for input in quizz_machine.batches(split="train", desc="train-entropy"):
 237     token_count += F.one_hot(input, num_classes=quizz_machine.vocabulary_size()).sum(
 238         (0, 1)
 239     )
 240 token_probas = token_count / token_count.sum()
 241 entropy = -torch.xlogy(token_probas, token_probas).sum()
 242 train_set_perplexity = math.exp(entropy)
 243
 244 ######################################################################
 245 # A bit of paranoia never hurts
 246
 247 if args.max_percents_of_test_in_train >= 0:
 248
 249     def subsets_as_tuples(batches, cs):
 250         s = set()
 251         for batch in batches:
 252             for x in batch:
 253                 s.add(tuple([v.item() for v in x]))
 254                 if len(s) == cs:
 255                     yield s
 256                     s = set()
 257         yield s
 258
 259     nb_test, nb_in_train = 0, 0
 260     for test_subset in subsets_as_tuples(
 261         quizz_machine.batches(split="test", desc="test-check"), 25000
 262     ):
 263         in_train = set()
 264         for train_subset in subsets_as_tuples(
 265             quizz_machine.batches(split="train", desc="train-check"), 25000
 266         ):
 267             in_train.update(test_subset.intersection(train_subset))
 268         nb_in_train += len(in_train)
 269         nb_test += len(test_subset)
 270
 271     log_string(
 272         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 273     )
 274
 275     assert (
 276         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 277     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 278
 279 ##############################
 280
 281
 282 def one_epoch(model, quizz_machine):
 283     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 284
 285     model.train()
 286
 287     nb_train_samples, acc_train_loss = 0, 0.0
 288
 289     for input in quizz_machine.batches(split="train"):
 290         input = input.to(device)
 291
 292         if nb_train_samples % args.batch_size == 0:
 293             optimizer.zero_grad()
 294
 295         output = model(mygpt.BracketedSequence(input)).x
 296         loss = F.cross_entropy(output.transpose(1, 2), input)
 297         acc_train_loss += loss.item() * input.size(0)
 298
 299         nb_train_samples += input.size(0)
 300
 301         loss.backward()
 302
 303         if nb_train_samples % args.batch_size == 0:
 304             optimizer.step()
 305
 306     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 307
 308     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 309
 310
 311 ######################################################################
 312
 313
 314 def run_tests(model, quizz_machine, deterministic_synthesis):
 315     with torch.autograd.no_grad():
 316         model.eval()
 317
 318         nb_test_samples, acc_test_loss = 0, 0.0
 319         nb_samples_accumulated = 0
 320
 321         for input in quizz_machine.batches(split="test"):
 322             input = input.to(device)
 323
 324             bs = model(mygpt.BracketedSequence(input))
 325             output = bs.x
 326
 327             loss = F.cross_entropy(output.transpose(1, 2), input)
 328
 329             acc_test_loss += loss.item() * input.size(0)
 330
 331             nb_test_samples += input.size(0)
 332
 333         main_test_accuracy = quizz_machine.produce_results(
 334             n_epoch=n_epoch,
 335             model=model,
 336             result_dir=args.result_dir,
 337             logger=log_string,
 338             deterministic_synthesis=deterministic_synthesis,
 339         )
 340
 341         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 342
 343         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 344
 345     model.main_test_accuracy = main_test_accuracy
 346
 347
 348 ######################################################################
 349
 350
 351 def create_c_quizzes(
 352     model,
 353     other_models,
 354     quizz_machine,
 355     nb_for_train=1000,
 356     nb_for_test=100,
 357     min_ave_seq_logproba=None,
 358 ):
 359     kept = []
 360
 361     sum_logits, sum_nb_c_quizzes = 0, 0
 362
 363     while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
 364         nb_to_generate = 4 * (nb_for_train + nb_for_test)
 365
 366         new_c_quizzes, nb_correct, ave_seq_logproba = quizz_machine.create_c_quizzes(
 367             n_epoch=n_epoch,
 368             result_dir=args.result_dir,
 369             logger=log_string,
 370             nb=nb_to_generate,
 371             model=model,
 372             other_models=other_models,
 373             min_ave_seq_logproba=min_ave_seq_logproba,
 374         )
 375
 376         sum_logits += new_c_quizzes.size(0) * ave_seq_logproba
 377         sum_nb_c_quizzes += new_c_quizzes.size(0)
 378
 379         to_keep = new_c_quizzes[nb_correct == len(other_models) - 1]
 380
 381         if args.dirty_debug:
 382             to_keep = new_c_quizzes
 383
 384         log_string(
 385             f"keep {to_keep.size(0)}/{new_c_quizzes.size(0)} c_quizzes ({to_keep.size(0)*100/new_c_quizzes.size(0):.02f}%)"
 386         )
 387
 388         kept.append(to_keep)
 389
 390     new_c_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
 391
 392     quizz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 393     quizz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 394
 395     quizz_machine.problem.save_quizzes(
 396         new_c_quizzes[:72],
 397         args.result_dir,
 398         f"culture_c_quiz_{n_epoch:04d}_{model.id:02d}",
 399         log_string,
 400     )
 401
 402     return sum_logits / sum_nb_c_quizzes
 403
 404
 405 ######################################################################
 406
 407 models = []
 408
 409 for k in range(args.nb_gpts):
 410     model = mygpt.MyGPT(
 411         vocabulary_size=vocabulary_size,
 412         dim_model=args.dim_model,
 413         dim_keys=args.dim_keys,
 414         dim_hidden=args.dim_hidden,
 415         nb_heads=args.nb_heads,
 416         nb_blocks=args.nb_blocks,
 417         causal=True,
 418         dropout=args.dropout,
 419     ).to(device)
 420
 421     model.main_test_accuracy = 0.0
 422     model.id = k
 423
 424     models.append(model)
 425
 426
 427 nb_parameters = sum(p.numel() for p in models[0].parameters())
 428 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 429
 430 ######################################################################
 431
 432 min_ave_seq_logproba = None
 433
 434 for n_epoch in range(args.nb_epochs):
 435     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 436
 437     a = [(model.id, float(model.main_test_accuracy)) for model in models]
 438     a.sort(key=lambda p: p[0])
 439     log_string(f"current accuracies {a}")
 440
 441     # select the model with lowest accuracy
 442     models.sort(key=lambda model: model.main_test_accuracy)
 443     model = models[0]
 444
 445     log_string(
 446         f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
 447     )
 448
 449     # improve it
 450     one_epoch(model, quizz_machine)
 451
 452     quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
 453
 454     log_string(
 455         f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
 456     )
 457
 458     # test it
 459     run_tests(model, quizz_machine, deterministic_synthesis=False)
 460
 461     log_string(
 462         f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
 463     )
 464
 465     if min([m.main_test_accuracy for m in models]) >= accuracy_to_make_c_quizzes:
 466         other_models = models.copy()
 467         other_models.remove(model)
 468
 469         ave_seq_logproba = create_c_quizzes(
 470             model,
 471             other_models,
 472             quizz_machine,
 473             nb_for_train=nb_new_c_quizzes_for_train,
 474             nb_for_test=nb_new_c_quizzes_for_test,
 475             min_ave_seq_logproba=min_ave_seq_logproba,
 476         )
 477
 478         # We keep the first average logits as a reference
 479         if min_ave_seq_logproba is None:
 480             min_ave_seq_logproba = ave_seq_logproba
 481         else:
 482             log_string(
 483                 f"min_ave_seq_logproba {min_ave_seq_logproba} ave_seq_logproba {ave_seq_logproba}"
 484             )
 485
 486         # We update everyone
 487         for model in models:
 488             run_tests(model, quizz_machine, deterministic_synthesis=False)
 489
 490
 491 ######################################################################