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, tasks
  16
  17 # world quizzes vs. culture quizzes
  18
  19 ######################################################################
  20
  21 accuracy_to_make_c_quizzes = 0.975
  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-4)
  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("--dirty_debug", action="store_true", default=False)
  85
  86 ######################################################################
  87
  88 args = parser.parse_args()
  89
  90 if args.result_dir is None:
  91     args.result_dir = f"results_culture"
  92
  93 ######################################################################
  94
  95 if args.dirty_debug:
  96     accuracy_to_make_c_quizzes = 0.0
  97     nb_new_c_quizzes_for_train = 100
  98     nb_new_c_quizzes_for_test = 10
  99
 100 ######################################################################
 101
 102 default_args = {
 103     "model": "37M",
 104     "batch_size": 100,
 105     "nb_train_samples": 250000,
 106     "nb_test_samples": 10000,
 107 }
 108
 109 for k, v in default_args.items():
 110     if getattr(args, k) is None:
 111         setattr(args, k, v)
 112
 113 ######################################################################
 114
 115 default_model_args = {
 116     "17K": {
 117         "dim_model": 32,
 118         "dim_keys": 32,
 119         "dim_hidden": 32,
 120         "nb_heads": 2,
 121         "nb_blocks": 2,
 122     },
 123     "4M": {
 124         "dim_model": 256,
 125         "dim_keys": 32,
 126         "dim_hidden": 1024,
 127         "nb_heads": 4,
 128         "nb_blocks": 6,
 129     },
 130     "37M": {
 131         "dim_model": 512,
 132         "dim_keys": 64,
 133         "dim_hidden": 2048,
 134         "nb_heads": 8,
 135         "nb_blocks": 12,
 136     },
 137     "122M": {
 138         "dim_model": 768,
 139         "dim_keys": 64,
 140         "dim_hidden": 2048,
 141         "nb_heads": 8,
 142         "nb_blocks": 24,
 143     },
 144     "352M": {
 145         "dim_model": 1024,
 146         "dim_keys": 64,
 147         "dim_hidden": 2048,
 148         "nb_heads": 8,
 149         "nb_blocks": 48,
 150     },
 151 }
 152
 153 if args.model in default_model_args:
 154     for k, v in default_model_args[args.model].items():
 155         if getattr(args, k) is None:
 156             setattr(args, k, v)
 157 else:
 158     raise ValueError(f"Unknown model {args.model}")
 159
 160 ######################################################################
 161
 162 try:
 163     os.mkdir(args.result_dir)
 164 except FileExistsError:
 165     print(f"result directory {args.result_dir} already exists")
 166     exit(1)
 167
 168 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 169
 170 if args.seed >= 0:
 171     # torch.backends.cudnn.deterministic = True
 172     # torch.backends.cudnn.benchmark = False
 173     # torch.use_deterministic_algorithms(True)
 174     torch.manual_seed(args.seed)
 175     if torch.cuda.is_available():
 176         torch.cuda.manual_seed_all(args.seed)
 177
 178 ######################################################################
 179
 180
 181 def log_string(s):
 182     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 183
 184     if log_file is not None:
 185         log_file.write(t + s + "\n")
 186         log_file.flush()
 187
 188     print(t + s)
 189     sys.stdout.flush()
 190
 191
 192 log_string(f"argv {' '.join(sys.argv)}")
 193
 194 for n in vars(args):
 195     log_string(f"args.{n} {getattr(args, n)}")
 196
 197
 198 ######################################################################
 199
 200 if args.dirty_debug:
 201     args.nb_train_samples = 2500
 202     args.nb_test_samples = 100
 203
 204 if args.physical_batch_size is None:
 205     args.physical_batch_size = args.batch_size
 206 else:
 207     assert args.batch_size % args.physical_batch_size == 0
 208
 209 assert args.nb_train_samples % args.batch_size == 0
 210 assert args.nb_test_samples % args.batch_size == 0
 211
 212 task = tasks.World(
 213     nb_train_samples=args.nb_train_samples,
 214     nb_test_samples=args.nb_test_samples,
 215     batch_size=args.physical_batch_size,
 216     result_dir=args.result_dir,
 217     logger=log_string,
 218     device=device,
 219 )
 220
 221 ######################################################################
 222
 223 log_string(f"device {device}")
 224
 225 vocabulary_size = task.vocabulary_size()
 226
 227 log_string(f"vocabulary_size {vocabulary_size}")
 228
 229 ######################################################################
 230
 231 # Compute the entropy of the training tokens
 232
 233 token_count = 0
 234 for input in task.batches(split="train", desc="train-entropy"):
 235     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 236 token_probas = token_count / token_count.sum()
 237 entropy = -torch.xlogy(token_probas, token_probas).sum()
 238 train_set_perplexity = math.exp(entropy)
 239
 240 ######################################################################
 241 # A bit of paranoia never hurts
 242
 243 if args.max_percents_of_test_in_train >= 0:
 244
 245     def subsets_as_tuples(batches, cs):
 246         s = set()
 247         for batch in batches:
 248             for x in batch:
 249                 s.add(tuple([v.item() for v in x]))
 250                 if len(s) == cs:
 251                     yield s
 252                     s = set()
 253         yield s
 254
 255     nb_test, nb_in_train = 0, 0
 256     for test_subset in subsets_as_tuples(
 257         task.batches(split="test", desc="test-check"), 25000
 258     ):
 259         in_train = set()
 260         for train_subset in subsets_as_tuples(
 261             task.batches(split="train", desc="train-check"), 25000
 262         ):
 263             in_train.update(test_subset.intersection(train_subset))
 264         nb_in_train += len(in_train)
 265         nb_test += len(test_subset)
 266
 267     log_string(
 268         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 269     )
 270
 271     assert (
 272         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 273     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 274
 275 ##############################
 276
 277
 278 def one_epoch(model, task):
 279     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 280
 281     model.train()
 282
 283     nb_train_samples, acc_train_loss = 0, 0.0
 284
 285     for input in task.batches(split="train"):
 286         input = input.to(device)
 287
 288         if nb_train_samples % args.batch_size == 0:
 289             optimizer.zero_grad()
 290
 291         output = model(mygpt.BracketedSequence(input)).x
 292         loss = F.cross_entropy(output.transpose(1, 2), input)
 293         acc_train_loss += loss.item() * input.size(0)
 294
 295         nb_train_samples += input.size(0)
 296
 297         loss.backward()
 298
 299         if nb_train_samples % args.batch_size == 0:
 300             optimizer.step()
 301
 302     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 303
 304     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 305
 306
 307 ######################################################################
 308
 309
 310 def run_tests(model, task, deterministic_synthesis):
 311     with torch.autograd.no_grad():
 312         model.eval()
 313
 314         nb_test_samples, acc_test_loss = 0, 0.0
 315         nb_samples_accumulated = 0
 316
 317         for input in task.batches(split="test"):
 318             input = input.to(device)
 319
 320             bs = model(mygpt.BracketedSequence(input))
 321             output = bs.x
 322
 323             loss = F.cross_entropy(output.transpose(1, 2), input)
 324
 325             acc_test_loss += loss.item() * input.size(0)
 326
 327             nb_test_samples += input.size(0)
 328
 329         main_test_accuracy = task.produce_results(
 330             n_epoch=n_epoch,
 331             model=model,
 332             result_dir=args.result_dir,
 333             logger=log_string,
 334             deterministic_synthesis=deterministic_synthesis,
 335         )
 336
 337         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 338
 339         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 340
 341     model.main_test_accuracy = main_test_accuracy
 342
 343
 344 ######################################################################
 345
 346
 347 def create_c_quizzes(
 348     model,
 349     other_models,
 350     task,
 351     nb_for_train=1000,
 352     nb_for_test=100,
 353     min_ave_seq_logproba=None,
 354 ):
 355     kept = []
 356
 357     sum_logits, sum_nb_c_quizzes = 0, 0
 358
 359     while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
 360         nb_to_generate = 4 * (nb_for_train + nb_for_test)
 361
 362         new_c_quizzes, nb_correct, ave_seq_logproba = task.create_c_quizzes(
 363             n_epoch=n_epoch,
 364             result_dir=args.result_dir,
 365             logger=log_string,
 366             nb=nb_to_generate,
 367             model=model,
 368             other_models=other_models,
 369             min_ave_seq_logproba=min_ave_seq_logproba,
 370         )
 371
 372         sum_logits += new_c_quizzes.size(0) * ave_seq_logproba
 373         sum_nb_c_quizzes += new_c_quizzes.size(0)
 374
 375         to_keep = new_c_quizzes[nb_correct == len(other_models) - 1]
 376
 377         if args.dirty_debug:
 378             to_keep = new_c_quizzes
 379
 380         log_string(
 381             f"keep {to_keep.size(0)}/{new_c_quizzes.size(0)} c_quizzes ({to_keep.size(0)*100/new_c_quizzes.size(0):.02f}%)"
 382         )
 383
 384         kept.append(to_keep)
 385
 386     new_c_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
 387
 388     task.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
 389     task.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 390
 391     task.save_quizzes(
 392         new_c_quizzes[:72],
 393         args.result_dir,
 394         f"culture_c_quiz_{n_epoch:04d}_{model.id:02d}",
 395         log_string,
 396     )
 397
 398     return sum_logits / sum_nb_c_quizzes
 399
 400
 401 ######################################################################
 402
 403 models = []
 404
 405 for k in range(args.nb_gpts):
 406     model = mygpt.MyGPT(
 407         vocabulary_size=vocabulary_size,
 408         dim_model=args.dim_model,
 409         dim_keys=args.dim_keys,
 410         dim_hidden=args.dim_hidden,
 411         nb_heads=args.nb_heads,
 412         nb_blocks=args.nb_blocks,
 413         causal=True,
 414         dropout=args.dropout,
 415     ).to(device)
 416
 417     model.main_test_accuracy = 0.0
 418     model.id = k
 419
 420     models.append(model)
 421
 422
 423 nb_parameters = sum(p.numel() for p in models[0].parameters())
 424 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 425
 426 ######################################################################
 427
 428 min_ave_seq_logproba = None
 429
 430 for n_epoch in range(args.nb_epochs):
 431     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 432
 433     a = [(model.id, float(model.main_test_accuracy)) for model in models]
 434     a.sort(key=lambda p: p[0])
 435     log_string(f"current accuracies {a}")
 436
 437     # select the model with lowest accuracy
 438     models.sort(key=lambda model: model.main_test_accuracy)
 439     model = models[0]
 440
 441     log_string(
 442         f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
 443     )
 444
 445     # improve it
 446     one_epoch(model, task)
 447
 448     task.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
 449
 450     log_string(
 451         f"train_set_composition w_quizzes {task.nb_batch_w_quizzes} c_quizzes {task.nb_batch_c_quizzes}"
 452     )
 453
 454     # test it
 455     run_tests(model, task, deterministic_synthesis=False)
 456
 457     log_string(
 458         f"test_set_composition w_quizzes {task.nb_batch_w_quizzes} c_quizzes {task.nb_batch_c_quizzes}"
 459     )
 460
 461     if min([m.main_test_accuracy for m in models]) >= accuracy_to_make_c_quizzes:
 462         other_models = models.copy()
 463         other_models.remove(model)
 464
 465         ave_seq_logproba = create_c_quizzes(
 466             model,
 467             other_models,
 468             task,
 469             nb_for_train=nb_new_c_quizzes_for_train,
 470             nb_for_test=nb_new_c_quizzes_for_test,
 471             min_ave_seq_logproba=min_ave_seq_logproba,
 472         )
 473
 474         # We keep the first average logits as a reference
 475         if min_ave_seq_logproba is None:
 476             min_ave_seq_logproba = ave_seq_logproba
 477         else:
 478             log_string(
 479                 f"min_ave_seq_logproba {min_ave_seq_logproba} ave_seq_logproba {ave_seq_logproba}"
 480             )
 481
 482         # We update everyone
 483         for model in models:
 484             run_tests(model, task, deterministic_synthesis=False)
 485
 486
 487 ######################################################################