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 ######################################################################
  18
  19 if torch.cuda.is_available():
  20     device = torch.device("cuda")
  21     torch.backends.cuda.matmul.allow_tf32 = True
  22 else:
  23     device = torch.device("cpu")
  24
  25 ######################################################################
  26
  27 parser = argparse.ArgumentParser(
  28     description="An implementation of GPT with cache.",
  29     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
  30 )
  31
  32 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
  33
  34 parser.add_argument("--result_dir", type=str, default=None)
  35
  36 parser.add_argument("--seed", type=int, default=0)
  37
  38 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
  39
  40 ########################################
  41
  42 parser.add_argument("--nb_epochs", type=int, default=10000)
  43
  44 parser.add_argument("--batch_size", type=int, default=None)
  45
  46 parser.add_argument("--physical_batch_size", type=int, default=None)
  47
  48 parser.add_argument("--nb_train_samples", type=int, default=None)
  49
  50 parser.add_argument("--nb_test_samples", type=int, default=None)
  51
  52 parser.add_argument("--learning_rate", type=float, default=1e-4)
  53
  54 ########################################
  55
  56 parser.add_argument("--model", type=str, default=None)
  57
  58 parser.add_argument("--dim_model", type=int, default=None)
  59
  60 parser.add_argument("--dim_keys", type=int, default=None)
  61
  62 parser.add_argument("--dim_hidden", type=int, default=None)
  63
  64 parser.add_argument("--nb_heads", type=int, default=None)
  65
  66 parser.add_argument("--nb_blocks", type=int, default=None)
  67
  68 parser.add_argument("--dropout", type=float, default=0.1)
  69
  70 ########################################
  71
  72 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  73
  74 parser.add_argument("--nb_gpts", type=int, default=5)
  75
  76 parser.add_argument("--check", action="store_true", default=False)
  77
  78 ######################################################################
  79
  80 args = parser.parse_args()
  81
  82 if args.result_dir is None:
  83     args.result_dir = f"results_culture"
  84
  85 ######################################################################
  86
  87 default_args = {
  88     "model": "37M",
  89     "batch_size": 100,
  90     "nb_train_samples": 250000,
  91     "nb_test_samples": 10000,
  92 }
  93
  94 for k, v in default_args.items():
  95     if getattr(args, k) is None:
  96         setattr(args, k, v)
  97
  98 ######################################################################
  99
 100 default_model_args = {
 101     "17K": {
 102         "dim_model": 32,
 103         "dim_keys": 32,
 104         "dim_hidden": 32,
 105         "nb_heads": 2,
 106         "nb_blocks": 2,
 107     },
 108     "4M": {
 109         "dim_model": 256,
 110         "dim_keys": 32,
 111         "dim_hidden": 1024,
 112         "nb_heads": 4,
 113         "nb_blocks": 6,
 114     },
 115     "37M": {
 116         "dim_model": 512,
 117         "dim_keys": 64,
 118         "dim_hidden": 2048,
 119         "nb_heads": 8,
 120         "nb_blocks": 12,
 121     },
 122     "122M": {
 123         "dim_model": 768,
 124         "dim_keys": 64,
 125         "dim_hidden": 2048,
 126         "nb_heads": 8,
 127         "nb_blocks": 24,
 128     },
 129     "352M": {
 130         "dim_model": 1024,
 131         "dim_keys": 64,
 132         "dim_hidden": 2048,
 133         "nb_heads": 8,
 134         "nb_blocks": 48,
 135     },
 136 }
 137
 138 if args.model in default_model_args:
 139     for k, v in default_model_args[args.model].items():
 140         if getattr(args, k) is None:
 141             setattr(args, k, v)
 142 else:
 143     raise ValueError(f"Unknown model {args.model}")
 144
 145 ######################################################################
 146
 147 try:
 148     os.mkdir(args.result_dir)
 149 except FileExistsError:
 150     print(f"result directory {args.result_dir} already exists")
 151     exit(1)
 152
 153 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 154
 155 if args.seed >= 0:
 156     # torch.backends.cudnn.deterministic = True
 157     # torch.backends.cudnn.benchmark = False
 158     # torch.use_deterministic_algorithms(True)
 159     torch.manual_seed(args.seed)
 160     if torch.cuda.is_available():
 161         torch.cuda.manual_seed_all(args.seed)
 162
 163 ######################################################################
 164
 165
 166 def log_string(s):
 167     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 168
 169     if log_file is not None:
 170         log_file.write(t + s + "\n")
 171         log_file.flush()
 172
 173     print(t + s)
 174     sys.stdout.flush()
 175
 176
 177 log_string(f"argv {' '.join(sys.argv)}")
 178
 179 for n in vars(args):
 180     log_string(f"args.{n} {getattr(args, n)}")
 181
 182
 183 ######################################################################
 184
 185 if args.check:
 186     args.nb_train_samples = 2500
 187     args.nb_test_samples = 100
 188
 189 if args.physical_batch_size is None:
 190     args.physical_batch_size = args.batch_size
 191 else:
 192     assert args.batch_size % args.physical_batch_size == 0
 193
 194 assert args.nb_train_samples % args.batch_size == 0
 195 assert args.nb_test_samples % args.batch_size == 0
 196
 197 task = tasks.World(
 198     nb_train_samples=args.nb_train_samples,
 199     nb_test_samples=args.nb_test_samples,
 200     batch_size=args.physical_batch_size,
 201     result_dir=args.result_dir,
 202     logger=log_string,
 203     device=device,
 204 )
 205
 206 ######################################################################
 207
 208 log_string(f"device {device}")
 209
 210 vocabulary_size = task.vocabulary_size()
 211
 212 log_string(f"vocabulary_size {vocabulary_size}")
 213
 214 ######################################################################
 215
 216 # Compute the entropy of the training tokens
 217
 218 token_count = 0
 219 for input in task.batches(split="train", desc="train-entropy"):
 220     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 221 token_probas = token_count / token_count.sum()
 222 entropy = -torch.xlogy(token_probas, token_probas).sum()
 223 train_set_perplexity = math.exp(entropy)
 224
 225 ######################################################################
 226 # A bit of paranoia never hurts
 227
 228 if args.max_percents_of_test_in_train >= 0:
 229
 230     def subsets_as_tuples(batches, cs):
 231         s = set()
 232         for batch in batches:
 233             for x in batch:
 234                 s.add(tuple([v.item() for v in x]))
 235                 if len(s) == cs:
 236                     yield s
 237                     s = set()
 238         yield s
 239
 240     nb_test, nb_in_train = 0, 0
 241     for test_subset in subsets_as_tuples(
 242         task.batches(split="test", desc="test-check"), 25000
 243     ):
 244         in_train = set()
 245         for train_subset in subsets_as_tuples(
 246             task.batches(split="train", desc="train-check"), 25000
 247         ):
 248             in_train.update(test_subset.intersection(train_subset))
 249         nb_in_train += len(in_train)
 250         nb_test += len(test_subset)
 251
 252     log_string(
 253         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 254     )
 255
 256     assert (
 257         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 258     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 259
 260 ##############################
 261
 262
 263 def one_epoch(model, task):
 264     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 265
 266     model.train()
 267
 268     nb_train_samples, acc_train_loss = 0, 0.0
 269
 270     for input in task.batches(split="train"):
 271         input = input.to(device)
 272
 273         if nb_train_samples % args.batch_size == 0:
 274             optimizer.zero_grad()
 275
 276         output = model(mygpt.BracketedSequence(input)).x
 277         loss = F.cross_entropy(output.transpose(1, 2), input)
 278         acc_train_loss += loss.item() * input.size(0)
 279
 280         nb_train_samples += input.size(0)
 281
 282         loss.backward()
 283
 284         if nb_train_samples % args.batch_size == 0:
 285             optimizer.step()
 286
 287     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 288
 289     log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 290
 291
 292 ######################################################################
 293
 294
 295 def run_tests(model, task, deterministic_synthesis):
 296     with torch.autograd.no_grad():
 297         model.eval()
 298
 299         nb_test_samples, acc_test_loss = 0, 0.0
 300         nb_samples_accumulated = 0
 301
 302         for input in task.batches(split="test"):
 303             input = input.to(device)
 304
 305             bs = model(mygpt.BracketedSequence(input))
 306             output = bs.x
 307
 308             loss = F.cross_entropy(output.transpose(1, 2), input)
 309
 310             acc_test_loss += loss.item() * input.size(0)
 311
 312             nb_test_samples += input.size(0)
 313
 314         main_test_accuracy = task.produce_results(
 315             n_epoch=n_epoch,
 316             model=model,
 317             result_dir=args.result_dir,
 318             logger=log_string,
 319             deterministic_synthesis=deterministic_synthesis,
 320         )
 321
 322         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 323
 324         log_string(f"test_perplexity {n_epoch} {test_perplexity}")
 325
 326     model.main_test_accuracy = main_test_accuracy
 327
 328
 329 ######################################################################
 330
 331
 332 def create_quizzes(
 333     model,
 334     other_models,
 335     task,
 336     nb_for_train=1000,
 337     nb_for_test=100,
 338     desired_average_logits=None,
 339 ):
 340     kept = []
 341     nb_generated_tokens, sum_logits = 0, 0
 342
 343     while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
 344         nb_to_generate = 4 * (nb_for_train + nb_for_test)
 345         new_quizzes, nb_correct, average_logits = task.create_new_quizzes(
 346             n_epoch=n_epoch,
 347             result_dir=args.result_dir,
 348             logger=log_string,
 349             nb=nb_to_generate,
 350             model=model,
 351             other_models=other_models,
 352             desired_average_logits=desired_average_logits,
 353         )
 354
 355         nb_generated_tokens += new_quizzes.numel()
 356         sum_logits += average_logits * new_quizzes.numel()
 357
 358         to_keep = new_quizzes[nb_correct == len(other_models) - 1]
 359         log_string(
 360             f"keep {to_keep.size(0)}/{new_quizzes.size(0)} quizzes ({to_keep.size(0)*100/new_quizzes.size(0):.02f}%)"
 361         )
 362         kept.append(to_keep)
 363
 364     new_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
 365
 366     task.store_new_quizzes(new_quizzes[:nb_for_train], for_train=True)
 367     task.store_new_quizzes(new_quizzes[nb_for_train:], for_train=False)
 368
 369     task.save_image(
 370         new_quizzes[:72],
 371         args.result_dir,
 372         f"world_quiz_{n_epoch:04d}_{model.id:02d}.png",
 373         log_string,
 374     )
 375
 376     return sum_logits / nb_generated_tokens
 377
 378
 379 ######################################################################
 380
 381 models = []
 382
 383 for k in range(args.nb_gpts):
 384     model = mygpt.MyGPT(
 385         vocabulary_size=vocabulary_size,
 386         dim_model=args.dim_model,
 387         dim_keys=args.dim_keys,
 388         dim_hidden=args.dim_hidden,
 389         nb_heads=args.nb_heads,
 390         nb_blocks=args.nb_blocks,
 391         causal=True,
 392         dropout=args.dropout,
 393     ).to(device)
 394
 395     model.main_test_accuracy = 0.0
 396     model.id = k
 397
 398     models.append(model)
 399
 400
 401 nb_parameters = sum(p.numel() for p in models[0].parameters())
 402 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 403
 404 ######################################################################
 405
 406 accuracy_to_make_quizzes = 0.975
 407 nb_new_quizzes_for_train = 1000
 408 nb_new_quizzes_for_test = 100
 409
 410 if args.check:
 411     accuracy_to_make_quizzes = 0.0
 412     nb_new_quizzes_for_train = 10
 413     nb_new_quizzes_for_test = 10
 414
 415 desired_average_logits = None
 416
 417 for n_epoch in range(args.nb_epochs):
 418     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 419
 420     a = [(model.id, float(model.main_test_accuracy)) for model in models]
 421     a.sort(key=lambda p: p[0])
 422     log_string(f"current accuracies {a}")
 423
 424     # select the model with lowest accuracy
 425     models.sort(key=lambda model: model.main_test_accuracy)
 426     model = models[0]
 427
 428     log_string(
 429         f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
 430     )
 431
 432     # improve it
 433     one_epoch(model, task)
 434
 435     task.renew_samples(args.nb_train_samples // args.nb_gpts)
 436
 437     log_string(
 438         f"train_set_composition world {task.nb_batch_samples_world} quizzes {task.nb_batch_samples_quizzes}"
 439     )
 440
 441     # test it
 442     run_tests(model, task, deterministic_synthesis=False)
 443
 444     log_string(
 445         f"test_set_composition world {task.nb_batch_samples_world} quizzes {task.nb_batch_samples_quizzes}"
 446     )
 447
 448     if min([m.main_test_accuracy for m in models]) >= accuracy_to_make_quizzes:
 449         other_models = models.copy()
 450         other_models.remove(model)
 451
 452         average_logits = create_quizzes(
 453             model,
 454             other_models,
 455             task,
 456             nb_for_train=nb_new_quizzes_for_train,
 457             nb_for_test=nb_new_quizzes_for_test,
 458             desired_average_logits=desired_average_logits,
 459         )
 460
 461         # We keep the first average logits as a reference
 462         if desired_average_logits is None:
 463             desired_average_logits = average_logits
 464         else:
 465             log_string(
 466                 f"desired_average_logits {desired_average_logits} average_logits {average_logits}"
 467             )
 468
 469         # We update everyone
 470         for model in models:
 471             run_tests(model, task, deterministic_synthesis=False)
 472
 473
 474 ######################################################################