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 # torch.autograd.set_detect_anomaly(True) #!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
  15
  16 import ffutils
  17 import mygpt, tasks, problems
  18
  19 ######################################################################
  20
  21
  22 def str2bool(x):
  23     x = x.lower()
  24     if x in {"1", "true", "yes"}:
  25         return True
  26     elif x in {"0", "false", "no"}:
  27         return False
  28     else:
  29         raise ValueError
  30
  31
  32 parser = argparse.ArgumentParser(
  33     description="An implementation of GPT with cache.",
  34     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
  35 )
  36
  37 parser.add_argument(
  38     "--task",
  39     type=str,
  40     default="twotargets",
  41     help="byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp",
  42 )
  43
  44 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
  45
  46 parser.add_argument("--result_dir", type=str, default=None)
  47
  48 parser.add_argument("--seed", type=int, default=0)
  49
  50 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
  51
  52 parser.add_argument("--force_cpu", type=str2bool, default=False)
  53
  54 ########################################
  55
  56 parser.add_argument("--nb_epochs", type=int, default=25)
  57
  58 parser.add_argument("--physical_batch_size", type=int, default=None)
  59
  60 parser.add_argument("--batch_size", type=int, default=25)
  61
  62 parser.add_argument("--nb_train_samples", type=int, default=None)
  63
  64 parser.add_argument("--nb_test_samples", type=int, default=None)
  65
  66 parser.add_argument("--optim", type=str, default="adam")
  67
  68 ########################################
  69
  70 parser.add_argument("--nb_warmup_iter", type=int, default=100)
  71
  72 parser.add_argument("--nb_decay_iter", type=int, default=5000)
  73
  74 parser.add_argument("--learning_rate", type=float, default=6e-4)
  75
  76 parser.add_argument("--min_learning_rate", type=float, default=6e-5)
  77
  78 # legacy
  79
  80 parser.add_argument("--legacy_lr_schedule", type=str2bool, default=True)
  81
  82 parser.add_argument("--legacy_large_lr", type=float, default=1e-4)
  83
  84 parser.add_argument("--legacy_small_lr", type=float, default=2e-5)
  85
  86 parser.add_argument("--legacy_nb_epoch_large_lr", type=float, default=10)
  87
  88 ########################################
  89
  90 parser.add_argument("--model", type=str, default=None)
  91
  92 parser.add_argument("--attention", type=str, default=None)
  93
  94 parser.add_argument("--memex_proba", type=float, default=0)
  95
  96 parser.add_argument("--memex_nb_epochs", type=float, default=None)
  97
  98 parser.add_argument("--dim_model", type=int, default=None)
  99
 100 parser.add_argument("--dim_keys", type=int, default=None)
 101
 102 parser.add_argument("--dim_hidden", type=int, default=None)
 103
 104 parser.add_argument("--nb_heads", type=int, default=None)
 105
 106 parser.add_argument("--nb_lines", type=int, default=None)
 107
 108 parser.add_argument("--caterpillar_height", type=int, default=None)
 109
 110 parser.add_argument("--gate_dropout_proba", type=float, default=0.0)
 111
 112 parser.add_argument("--gate_dropout_sync", type=str2bool, default=False)
 113
 114 parser.add_argument("--gate_dropout_replace", type=str2bool, default=False)
 115
 116 parser.add_argument("--rho_inner_loss", type=float, default=0.0)
 117
 118 parser.add_argument("--nb_blocks", type=int, default=None)
 119
 120 parser.add_argument("--dropout", type=float, default=0.1)
 121
 122 ########################################
 123
 124 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
 125
 126 parser.add_argument("--no_checkpoint", action="store_true", default=False)
 127
 128 parser.add_argument("--continue_training", action="store_true", default=False)
 129
 130 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
 131
 132 ##############################
 133 # rpl options
 134
 135 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
 136
 137 parser.add_argument("--rpl_max_input", type=int, default=9)
 138
 139 parser.add_argument("--rpl_prog_len", type=int, default=8)
 140
 141 parser.add_argument("--rpl_nb_runs", type=int, default=5)
 142
 143 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
 144
 145 ##############################
 146 # grid options
 147
 148 parser.add_argument("--grid_size", type=int, default=6)
 149
 150 parser.add_argument("--grid_nb_colors", type=int, default=6)
 151
 152 parser.add_argument("--grid_nb_shapes", type=int, default=6)
 153
 154 ##############################
 155 # picoclvr options
 156
 157 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 158
 159 parser.add_argument("--picoclvr_height", type=int, default=12)
 160
 161 parser.add_argument("--picoclvr_width", type=int, default=16)
 162
 163 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
 164
 165 ##############################
 166 # Maze options
 167
 168 parser.add_argument("--maze_height", type=int, default=13)
 169
 170 parser.add_argument("--maze_width", type=int, default=21)
 171
 172 parser.add_argument("--maze_nb_walls", type=int, default=15)
 173
 174 ##############################
 175 # Snake options
 176
 177 parser.add_argument("--snake_height", type=int, default=9)
 178
 179 parser.add_argument("--snake_width", type=int, default=12)
 180
 181 parser.add_argument("--snake_nb_colors", type=int, default=5)
 182
 183 parser.add_argument("--snake_length", type=int, default=200)
 184
 185 ##############################
 186 # Stack options
 187
 188 parser.add_argument("--stack_nb_steps", type=int, default=100)
 189
 190 parser.add_argument("--stack_nb_stacks", type=int, default=3)
 191
 192 parser.add_argument("--stack_nb_digits", type=int, default=3)
 193
 194 parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
 195
 196 ##############################
 197 # Expr options
 198
 199 parser.add_argument("--expr_nb_variables", type=int, default=5)
 200
 201 parser.add_argument("--expr_sequence_length", type=int, default=40)
 202
 203 parser.add_argument("--expr_operand_max", type=int, default=9)
 204
 205 parser.add_argument("--expr_result_max", type=int, default=99)
 206
 207 parser.add_argument("--expr_input_file", type=str, default=None)
 208
 209 ##############################
 210 # Memory
 211
 212 parser.add_argument("--memory_len_total", type=int, default=32)
 213
 214 ##############################
 215 # Mixing
 216
 217 parser.add_argument("--mixing_hard", action="store_true", default=False)
 218
 219 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
 220
 221 ######################################################################
 222
 223 # args = parser.parse_args()
 224
 225 args, sup_args = parser.parse_known_args()
 226
 227 sup_args = dict([x.removeprefix("--").split("=") for x in sup_args])
 228
 229 if args.result_dir is None:
 230     args.result_dir = f"results_{args.task}_{args.model}"
 231
 232 ######################################################################
 233
 234 if not args.force_cpu and torch.cuda.is_available():
 235     device = torch.device("cuda")
 236     torch.backends.cuda.matmul.allow_tf32 = True
 237 else:
 238     device = torch.device("cpu")
 239
 240 ######################################################################
 241
 242 default_task_args = {
 243     "addition": {
 244         "model": "352M",
 245         "physical_batch_size": 25,
 246         "nb_train_samples": 250000,
 247         "nb_test_samples": 10000,
 248     },
 249     "byheart": {
 250         "model": "37M",
 251         "physical_batch_size": 25,
 252         "nb_train_samples": 50000,
 253         "nb_test_samples": 10000,
 254     },
 255     "expr": {
 256         "model": "352M",
 257         "physical_batch_size": 25,
 258         "nb_train_samples": 2500000,
 259         "nb_test_samples": 10000,
 260     },
 261     "grid": {
 262         "model": "37M",
 263         "physical_batch_size": 25,
 264         "nb_train_samples": 250000,
 265         "nb_test_samples": 10000,
 266     },
 267     "qmlp": {
 268         "model": "37M",
 269         "physical_batch_size": 10,
 270         "nb_train_samples": 100000,
 271         "nb_test_samples": 1000,
 272     },
 273     "guessop": {
 274         "model": "352M",
 275         "physical_batch_size": 25,
 276         "nb_train_samples": 1000000,
 277         "nb_test_samples": 10000,
 278     },
 279     "learnop": {
 280         "model": "37M",
 281         "physical_batch_size": 25,
 282         "nb_train_samples": 50000,
 283         "nb_test_samples": 10000,
 284     },
 285     "maze": {
 286         "model": "37M",
 287         "physical_batch_size": 5,
 288         "nb_train_samples": 100000,
 289         "nb_test_samples": 10000,
 290     },
 291     "picoclvr": {
 292         "model": "37M",
 293         "physical_batch_size": 25,
 294         "nb_train_samples": 250000,
 295         "nb_test_samples": 10000,
 296     },
 297     "rpl": {
 298         "model": "352M",
 299         "physical_batch_size": 5,
 300         "nb_train_samples": 2500000,
 301         "nb_test_samples": 10000,
 302     },
 303     "snake": {
 304         "model": "37M",
 305         "physical_batch_size": 25,
 306         "nb_train_samples": 250000,
 307         "nb_test_samples": 10000,
 308     },
 309     "stack": {
 310         "model": "37M",
 311         "physical_batch_size": 25,
 312         "nb_train_samples": 100000,
 313         "nb_test_samples": 1000,
 314     },
 315     "twotargets": {
 316         "model": "37M",
 317         "physical_batch_size": 25,
 318         "nb_train_samples": 50000,
 319         "nb_test_samples": 10000,
 320     },
 321     "memory": {
 322         "model": "37M",
 323         "physical_batch_size": 25,
 324         "nb_train_samples": 25000,
 325         "nb_test_samples": 10000,
 326     },
 327     "mixing": {
 328         "model": "37M",
 329         "physical_batch_size": 25,
 330         "nb_train_samples": 250000,
 331         "nb_test_samples": 10000,
 332     },
 333     "mnist": {
 334         "model": "37M",
 335         "physical_batch_size": 5,
 336         "nb_train_samples": 60000,
 337         "nb_test_samples": 10000,
 338     },
 339 }
 340
 341 if args.task in default_task_args:
 342     for k, v in default_task_args[args.task].items():
 343         if getattr(args, k) is None:
 344             setattr(args, k, v)
 345
 346 ######################################################################
 347
 348 default_model_args = {
 349     "17K": {
 350         "attention": "mha",
 351         "dim_model": 32,
 352         "dim_keys": 32,
 353         "dim_hidden": 32,
 354         "nb_heads": 2,
 355         "nb_blocks": 2,
 356     },
 357     "17K-C": {
 358         "attention": "caterpillar",
 359         "dim_model": 32,
 360         "dim_keys": 32,
 361         "dim_hidden": 32,
 362         "nb_heads": 2,
 363         "nb_lines": 16,
 364         "caterpillar_height": 4,
 365         "nb_blocks": 2,
 366     },
 367     "4M": {
 368         "attention": "mha",
 369         "dim_model": 256,
 370         "dim_keys": 32,
 371         "dim_hidden": 1024,
 372         "nb_heads": 4,
 373         "nb_blocks": 6,
 374     },
 375     "4M-C": {
 376         "attention": "caterpillar",
 377         "dim_model": 256,
 378         "dim_keys": 32,
 379         "dim_hidden": 1024,
 380         "nb_heads": 4,
 381         "nb_lines": 32,
 382         "caterpillar_height": 4,
 383         "nb_blocks": 6,
 384     },
 385     "37M": {
 386         "attention": "mha",
 387         "dim_model": 512,
 388         "dim_keys": 64,
 389         "dim_hidden": 2048,
 390         "nb_heads": 8,
 391         "nb_blocks": 12,
 392     },
 393     "37M-C": {
 394         "attention": "caterpillar",
 395         "dim_model": 512,
 396         "dim_keys": 64,
 397         "dim_hidden": 2048,
 398         "nb_heads": 8,
 399         "nb_lines": 256,
 400         "caterpillar_height": 32,
 401         "nb_blocks": 12,
 402     },
 403     "122M": {
 404         "attention": "mha",
 405         "dim_model": 768,
 406         "dim_keys": 64,
 407         "dim_hidden": 2048,
 408         "nb_heads": 8,
 409         "nb_blocks": 24,
 410     },
 411     "122M-C": {
 412         "attention": "caterpillar",
 413         "dim_model": 768,
 414         "dim_keys": 64,
 415         "dim_hidden": 2048,
 416         "nb_heads": 8,
 417         "nb_lines": 128,
 418         "nb_blocks": 24,
 419     },
 420     "352M": {
 421         "attention": "mha",
 422         "dim_model": 1024,
 423         "dim_keys": 64,
 424         "dim_hidden": 2048,
 425         "nb_heads": 8,
 426         "nb_blocks": 48,
 427     },
 428     "352M-C": {
 429         "attention": "caterpillar",
 430         "dim_model": 1024,
 431         "dim_keys": 64,
 432         "dim_hidden": 2048,
 433         "nb_heads": 8,
 434         "nb_lines": 128,
 435         "nb_blocks": 48,
 436     },
 437 }
 438
 439 if args.model in default_model_args:
 440     for k, v in default_model_args[args.model].items():
 441         if getattr(args, k) is None:
 442             setattr(args, k, v)
 443 else:
 444     raise ValueError(f"Unknown model {args.model}")
 445
 446 ######################################################################
 447
 448 try:
 449     os.mkdir(args.result_dir)
 450 except FileExistsError:
 451     if not args.continue_training:
 452         print(f"result directory {args.result_dir} already exists")
 453         exit(1)
 454
 455 loss_file = open(os.path.join(args.result_dir, "loss.dat"), "a")
 456
 457 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 458
 459 if args.seed >= 0:
 460     # torch.backends.cudnn.deterministic = True
 461     # torch.backends.cudnn.benchmark = False
 462     # torch.use_deterministic_algorithms(True)
 463     torch.manual_seed(args.seed)
 464     if torch.cuda.is_available():
 465         torch.cuda.manual_seed_all(args.seed)
 466
 467 ######################################################################
 468
 469
 470 def log_string(s):
 471     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 472
 473     if log_file is not None:
 474         log_file.write(t + s + "\n")
 475         log_file.flush()
 476
 477     print(t + s)
 478     sys.stdout.flush()
 479
 480
 481 with os.popen("sha256sum *.py") as f:
 482     for l in f:
 483         log_string(f"sha256sum {l.strip()}")
 484
 485 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
 486 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
 487
 488 log_string(f"argv {' '.join(sys.argv)}")
 489
 490 for n in vars(args):
 491     log_string(f"args.{n} {getattr(args, n)}")
 492
 493 for k, v in sup_args.items():
 494     log_string(f'sup_args["{k}"] "{v}"')
 495
 496
 497 ######################################################################
 498
 499
 500 def get_lr(n_epoch, it):
 501     if args.legacy_lr_schedule:
 502         # my crude scheduling to compare to previous baseline, added
 503         # warmup though
 504
 505         if it < args.nb_warmup_iter:
 506             return args.legacy_large_lr * it / args.nb_warmup_iter
 507         elif n_epoch < args.legacy_nb_epoch_large_lr:
 508             return args.legacy_large_lr
 509         else:
 510             return args.legacy_small_lr
 511
 512     # from nanoGPT
 513
 514     # 1) linear warmup for warmup_iter steps
 515     if it < args.nb_warmup_iter:
 516         return args.learning_rate * it / args.nb_warmup_iter
 517     # 2) if it > nb_decay_iter, return min learning rate
 518     if it > args.nb_decay_iter:
 519         return args.min_learning_rate
 520     # 3) in between, use cosine decay down to min learning rate
 521     decay_ratio = (it - args.nb_warmup_iter) / (
 522         args.nb_decay_iter - args.nb_warmup_iter
 523     )
 524     coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))  # coeff ranges 0..1
 525     return args.min_learning_rate + coeff * (
 526         args.learning_rate - args.min_learning_rate
 527     )
 528
 529
 530 ######################################################################
 531
 532
 533 def add_memex_v2(batches, memex_proba, marker_token):
 534     for input in batches:
 535         if torch.rand(1).item() < memex_proba:
 536             t = (
 537                 torch.arange(1 + 2 * input.size(1), device=input.device)[None, :]
 538                 .expand(input.size(0), -1)
 539                 .clone()
 540             )
 541
 542             u0 = torch.randint(input.size(1), (input.size(0), 1), device=input.device)
 543             caterpillar_length = args.nb_lines // args.caterpillar_height
 544             u1 = (
 545                 u0
 546                 + torch.randint(
 547                     caterpillar_length, (input.size(0), 1), device=input.device
 548                 )
 549                 + 1
 550             )
 551
 552             m0 = (t < u0).long()
 553             m1 = (t >= u1).long() * (t < u1 + input.size(1)).long()
 554
 555             t = t * m0 + ((-1) * (1 - m0) * (1 - m1)) + (t - u1) * m1
 556             m = (t < 0).long()
 557             n = torch.arange(input.size(0), device=input.device)[:, None].expand(
 558                 -1, t.size(1)
 559             )
 560
 561             new_input = input[n, t.clamp(min=0)]
 562             new_input = (1 - m) * new_input + m * (marker_token)
 563
 564             yield new_input
 565
 566         yield input
 567
 568
 569 def add_memex_v3(batches, memex_proba, marker_token):
 570     for input in batches:
 571         if torch.rand(1).item() < memex_proba:
 572             t = (
 573                 torch.arange(2 * input.size(1), device=input.device)[None, :]
 574                 .expand(input.size(0), -1)
 575                 .clone()
 576             )
 577
 578             u = torch.rand(t.size(), device=t.device)
 579             u[:, : input.size(1)] = 1.0
 580             memex_v3_proba_fragment = 1 / 20
 581             u = (u < memex_v3_proba_fragment).long()
 582             v = u * torch.randint(input.size(1), u.size())
 583             u[:, input.size(1) + 1 :] = v[:, input.size(1) + 1 :] - u[
 584                 :, : input.size(1) - 1
 585             ] * input.size(1)
 586             u = u.cumsum().clamp(min=0)
 587
 588             u0 = torch.randint(input.size(1), (input.size(0), 1), device=input.device)
 589             caterpillar_length = args.nb_lines // args.caterpillar_height
 590             u1 = (
 591                 u0
 592                 + torch.randint(
 593                     caterpillar_length, (input.size(0), 1), device=input.device
 594                 )
 595                 + 1
 596             )
 597
 598             m0 = (t < u0).long()
 599             m1 = (t >= u1).long() * (t < u1 + input.size(1)).long()
 600
 601             t = t * m0 + ((-1) * (1 - m0) * (1 - m1)) + (t - u1) * m1
 602             m = (t < 0).long()
 603             n = torch.arange(input.size(0), device=input.device)[:, None].expand(
 604                 -1, t.size(1)
 605             )
 606
 607             new_input = input[n, t.clamp(min=0)]
 608             new_input = (1 - m) * new_input + m * (marker_token)
 609
 610             yield new_input
 611
 612         yield input
 613
 614
 615 ######################################################################
 616
 617 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 618
 619
 620 def picoclvr_pruner_horizontal_green(p):
 621     return not ("green" in p and ("left" in p or "right" in p))
 622
 623
 624 picoclvr_pruner_train = (
 625     picoclvr_pruner_horizontal_green
 626     if args.picocvlr_prune_properties in {"train+eval"}
 627     else None
 628 )
 629
 630 picoclvr_pruner_eval = (
 631     (lambda p: not picoclvr_pruner_horizontal_green(p))
 632     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 633     else None
 634 )
 635
 636 ######################################################################
 637
 638 device_data = device
 639
 640 if args.task == "byheart":
 641     task = tasks.SandBox(
 642         problem=problems.ProblemByHeart(),
 643         nb_train_samples=args.nb_train_samples,
 644         nb_test_samples=args.nb_test_samples,
 645         batch_size=args.physical_batch_size,
 646         logger=log_string,
 647         device=device_data,
 648     )
 649     args.max_percents_of_test_in_train = -1
 650
 651 elif args.task == "learnop":
 652     task = tasks.SandBox(
 653         problem=problems.ProblemLearnOperator(),
 654         nb_train_samples=args.nb_train_samples,
 655         nb_test_samples=args.nb_test_samples,
 656         batch_size=args.physical_batch_size,
 657         logger=log_string,
 658         device=device_data,
 659     )
 660
 661
 662 elif args.task == "guessop":
 663     task = tasks.SandBox(
 664         problem=problems.ProblemGuessOperator(),
 665         nb_train_samples=args.nb_train_samples,
 666         nb_test_samples=args.nb_test_samples,
 667         batch_size=args.physical_batch_size,
 668         logger=log_string,
 669         device=device_data,
 670     )
 671
 672
 673 elif args.task == "twotargets":
 674     task = tasks.SandBox(
 675         problem=problems.ProblemTwoTargets(),
 676         nb_train_samples=args.nb_train_samples,
 677         nb_test_samples=args.nb_test_samples,
 678         batch_size=args.physical_batch_size,
 679         logger=log_string,
 680         device=device_data,
 681     )
 682
 683 elif args.task == "memory":
 684     task = tasks.SandBox(
 685         problem=problems.ProblemMemory(len_total=args.memory_len_total),
 686         nb_train_samples=args.nb_train_samples,
 687         nb_test_samples=args.nb_test_samples,
 688         batch_size=args.physical_batch_size,
 689         logger=log_string,
 690         device=device_data,
 691     )
 692
 693 elif args.task == "mixing":
 694     task = tasks.SandBox(
 695         problem=problems.ProblemMixing(
 696             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 697         ),
 698         nb_train_samples=args.nb_train_samples,
 699         nb_test_samples=args.nb_test_samples,
 700         batch_size=args.physical_batch_size,
 701         logger=log_string,
 702         device=device_data,
 703     )
 704
 705 elif args.task == "addition":
 706     task = tasks.SandBox(
 707         problem=problems.ProblemAddition(),
 708         nb_train_samples=args.nb_train_samples,
 709         nb_test_samples=args.nb_test_samples,
 710         batch_size=args.physical_batch_size,
 711         logger=log_string,
 712         device=device_data,
 713     )
 714
 715 elif args.task == "picoclvr":
 716     task = tasks.PicoCLVR(
 717         nb_train_samples=args.nb_train_samples,
 718         nb_test_samples=args.nb_test_samples,
 719         batch_size=args.physical_batch_size,
 720         height=args.picoclvr_height,
 721         width=args.picoclvr_width,
 722         nb_colors=args.picoclvr_nb_colors,
 723         logger=log_string,
 724         device=device_data,
 725         pruner_train=picoclvr_pruner_train,
 726         pruner_eval=picoclvr_pruner_eval,
 727     )
 728
 729 elif args.task == "mnist":
 730     task = tasks.MNIST(
 731         nb_train_samples=args.nb_train_samples,
 732         nb_test_samples=args.nb_test_samples,
 733         batch_size=args.physical_batch_size,
 734         device=device_data,
 735     )
 736
 737 elif args.task == "maze":
 738     task = tasks.Maze(
 739         nb_train_samples=args.nb_train_samples,
 740         nb_test_samples=args.nb_test_samples,
 741         batch_size=args.physical_batch_size,
 742         height=args.maze_height,
 743         width=args.maze_width,
 744         nb_walls=args.maze_nb_walls,
 745         device=device_data,
 746     )
 747
 748 elif args.task == "snake":
 749     task = tasks.Snake(
 750         nb_train_samples=args.nb_train_samples,
 751         nb_test_samples=args.nb_test_samples,
 752         batch_size=args.physical_batch_size,
 753         height=args.snake_height,
 754         width=args.snake_width,
 755         nb_colors=args.snake_nb_colors,
 756         length=args.snake_length,
 757         prompt_length=args.snake_length // 2,
 758         device=device_data,
 759     )
 760
 761 elif args.task == "stack":
 762     task = tasks.Stack(
 763         nb_train_samples=args.nb_train_samples,
 764         nb_test_samples=args.nb_test_samples,
 765         batch_size=args.physical_batch_size,
 766         logger=log_string,
 767         nb_steps=args.stack_nb_steps,
 768         nb_stacks=args.stack_nb_stacks,
 769         nb_digits=args.stack_nb_digits,
 770         fraction_values_for_train=args.stack_fraction_values_for_train,
 771         device=device_data,
 772     )
 773
 774 elif args.task == "expr":
 775     task = tasks.Expr(
 776         nb_train_samples=args.nb_train_samples,
 777         nb_test_samples=args.nb_test_samples,
 778         nb_variables=args.expr_nb_variables,
 779         sequence_length=args.expr_sequence_length,
 780         operand_max=args.expr_operand_max,
 781         result_max=args.expr_result_max,
 782         batch_size=args.physical_batch_size,
 783         device=device_data,
 784     )
 785
 786 elif args.task == "rpl":
 787     task = tasks.RPL(
 788         nb_train_samples=args.nb_train_samples,
 789         nb_test_samples=args.nb_test_samples,
 790         batch_size=args.physical_batch_size,
 791         nb_starting_values=args.rpl_nb_starting_values,
 792         max_input=args.rpl_max_input,
 793         prog_len=args.rpl_prog_len,
 794         nb_runs=args.rpl_nb_runs,
 795         no_prog=args.rpl_no_prog,
 796         logger=log_string,
 797         device=device_data,
 798     )
 799
 800 elif args.task == "grid":
 801     task = tasks.Grid(
 802         nb_train_samples=args.nb_train_samples,
 803         nb_test_samples=args.nb_test_samples,
 804         batch_size=args.physical_batch_size,
 805         size=args.grid_size,
 806         nb_shapes=args.grid_nb_shapes,
 807         nb_colors=args.grid_nb_colors,
 808         logger=log_string,
 809         device=device_data,
 810     )
 811
 812 elif args.task == "qmlp":
 813     task = tasks.QMLP(
 814         nb_train_samples=args.nb_train_samples,
 815         nb_test_samples=args.nb_test_samples,
 816         batch_size=args.physical_batch_size,
 817         result_dir=args.result_dir,
 818         logger=log_string,
 819         device=device_data,
 820     )
 821
 822 else:
 823     raise ValueError(f"Unknown task {args.task}")
 824
 825 ######################################################################
 826
 827 log_string(f"device {device}")
 828
 829 vocabulary_size = task.vocabulary_size()
 830
 831 if args.memex_proba > 0:
 832     vocabulary_size += 1
 833
 834 log_string(f"vocabulary_size {vocabulary_size}")
 835
 836 ##############################
 837
 838 model = mygpt.MyGPT(
 839     vocabulary_size=vocabulary_size,
 840     dim_model=args.dim_model,
 841     dim_keys=args.dim_keys,
 842     dim_hidden=args.dim_hidden,
 843     nb_heads=args.nb_heads,
 844     nb_lines=args.nb_lines,
 845     caterpillar_height=args.caterpillar_height,
 846     nb_blocks=args.nb_blocks,
 847     causal=True,
 848     dropout=args.dropout,
 849     attention_layer=args.attention,
 850     logger=log_string,
 851     args=args,
 852 )
 853
 854 model.to(device)
 855
 856 nb_parameters = sum(p.numel() for p in model.parameters())
 857 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 858
 859 ######################################################################
 860
 861 nb_epochs_finished = 0
 862
 863 if args.no_checkpoint:
 864     log_string(f"not trying to load checkpoint.")
 865
 866 else:
 867     try:
 868         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 869         checkpoint = torch.load(checkpoint_name)
 870         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 871         model.load_state_dict(checkpoint["model_state"])
 872         torch.set_rng_state(checkpoint["rng_state"])
 873         if torch.cuda.is_available():
 874             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 875
 876         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 877
 878     except FileNotFoundError:
 879         log_string("starting from scratch.")
 880
 881     except:
 882         log_string("error when loading the checkpoint.")
 883         exit(1)
 884
 885 ######################################################################
 886
 887 if args.task == "expr" and args.expr_input_file is not None:
 888     task.produce_results(
 889         n_epoch=nb_epochs_finished,
 890         model=model,
 891         result_dir=args.result_dir,
 892         logger=log_string,
 893         deterministic_synthesis=args.deterministic_synthesis,
 894         input_file=args.expr_input_file,
 895     )
 896
 897     exit(0)
 898
 899 ######################################################################
 900
 901 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 902
 903 # Compute the entropy of the training tokens
 904
 905 token_count = 0
 906 for input in task.batches(split="train"):
 907     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 908 token_probas = token_count / token_count.sum()
 909 entropy = -torch.xlogy(token_probas, token_probas).sum()
 910 train_set_perplexity = math.exp(entropy)
 911
 912 ######################################################################
 913 # A bit of paranoia never hurts
 914
 915 if args.max_percents_of_test_in_train >= 0:
 916
 917     def subsets_as_tuples(batches, cs):
 918         s = set()
 919         for batch in batches:
 920             for x in batch:
 921                 s.add(tuple([v.item() for v in x]))
 922                 if len(s) == cs:
 923                     yield s
 924                     s = set()
 925         yield s
 926
 927     nb_test, nb_in_train = 0, 0
 928     for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 929         in_train = set()
 930         for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 931             in_train.update(test_subset.intersection(train_subset))
 932         nb_in_train += len(in_train)
 933         nb_test += len(test_subset)
 934
 935     log_string(
 936         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 937     )
 938
 939     assert (
 940         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 941     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 942
 943 ##############################
 944
 945 if "calibrate" in sup_args:
 946     for input in task.batches(split="train", desc="calibrate"):
 947         input = input.to(device)
 948         output = model(mygpt.BracketedSequence(input)).x
 949
 950     for n, m in model.named_modules():
 951         for a in dir(m):
 952             x = getattr(m, a)
 953             if isinstance(x, mygpt.Calibrator):
 954                 print(f"####### ${n} | ${a} ########################")
 955                 mean, std = x.moments()
 956                 print("mean\n", mean, "\n")
 957                 print("std\n", std, "\n")
 958                 print(f"############################################\n\n")
 959
 960     exit(0)
 961
 962 ##############################
 963
 964 nb_samples_seen = 0
 965
 966 if nb_epochs_finished >= nb_epochs:
 967     task.produce_results(
 968         n_epoch=nb_epochs_finished,
 969         model=model,
 970         result_dir=args.result_dir,
 971         logger=log_string,
 972         deterministic_synthesis=args.deterministic_synthesis,
 973     )
 974
 975 time_pred_result = datetime.datetime.now()
 976
 977 it = 0
 978
 979 n_batch = 0
 980
 981 for n_epoch in range(nb_epochs_finished, nb_epochs):
 982     if args.optim == "sgd":
 983         optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
 984     elif args.optim == "adam":
 985         optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 986     elif args.optim == "adamw":
 987         optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate)
 988     else:
 989         raise ValueError(f"Unknown optimizer {args.optim}.")
 990
 991     model.train()
 992
 993     nb_train_samples, acc_train_loss, acc_train_inner_loss = 0, 0.0, 0.0
 994
 995     memex_proba = (
 996         args.memex_proba
 997         if args.memex_nb_epochs is None or n_epoch < args.memex_nb_epochs
 998         else 0.0
 999     )
1000
1001     log_string(f"memex_proba {memex_proba}")
1002
1003     train_batches = add_memex_v2(
1004         batches=task.batches(split="train"),
1005         memex_proba=memex_proba,
1006         marker_token=vocabulary_size - 1,
1007     )
1008
1009     def add_none(it):
1010         for x in it:
1011             yield x
1012         yield None
1013
1014     nb_acc_samples = 0
1015
1016     for input in add_none(train_batches):
1017         if input is not None:
1018             model.reset_inner_loss()
1019             input = input.to(device)
1020
1021             output = model(mygpt.BracketedSequence(input)).x
1022             loss = F.cross_entropy(output.transpose(1, 2), input)
1023             inner_loss = model.get_inner_loss()
1024
1025             acc_train_loss += loss.item() * input.size(0)
1026             acc_train_inner_loss += inner_loss.item() * input.size(0)
1027
1028             nb_train_samples += input.size(0)
1029             nb_samples_seen += input.size(0)
1030
1031             total_loss = loss + (
1032                 args.rho_inner_loss * inner_loss if args.rho_inner_loss > 0 else 0.0
1033             )
1034
1035             it += 1
1036             lr = get_lr(n_epoch, it)
1037             for param_group in optimizer.param_groups:
1038                 param_group["lr"] = lr
1039
1040                 # log_string(f"learning_rate {lr}")
1041
1042             total_loss.backward()
1043             nb_acc_samples += input.size(0)
1044
1045         if (input is None and nb_acc_samples > 0) or nb_acc_samples == args.batch_size:
1046             assert nb_acc_samples <= args.batch_size
1047             optimizer.step()
1048             grad_norm = sum([p.grad.pow(2).sum() for p in model.parameters()]).sqrt()
1049             loss_file.write(f"{n_epoch} {n_batch} {loss.item()} {grad_norm.item()}\n")
1050             optimizer.zero_grad()
1051             nb_acc_samples = 0
1052
1053         n_batch += 1
1054
1055     with torch.autograd.no_grad():
1056         model.eval()
1057
1058         nb_test_samples, acc_test_loss = 0, 0.0
1059
1060         for input in task.batches(split="test"):
1061             input = input.to(device)
1062
1063             output = model(mygpt.BracketedSequence(input)).x
1064             loss = F.cross_entropy(output.transpose(1, 2), input)
1065             acc_test_loss += loss.item() * input.size(0)
1066             nb_test_samples += input.size(0)
1067
1068         log_string(
1069             f"loss {n_epoch} train_loss {acc_train_loss/nb_train_samples} train_inner_loss {acc_train_inner_loss/nb_train_samples} test_prediction {acc_test_loss/nb_test_samples}"
1070         )
1071
1072         task.produce_results(
1073             n_epoch=n_epoch,
1074             model=model,
1075             result_dir=args.result_dir,
1076             logger=log_string,
1077             deterministic_synthesis=args.deterministic_synthesis,
1078         )
1079
1080         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
1081         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
1082
1083         log_string(
1084             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
1085         )
1086
1087         time_current_result = datetime.datetime.now()
1088         log_string(
1089             f"next_result {time_current_result + (time_current_result - time_pred_result)}"
1090         )
1091         time_pred_result = time_current_result
1092
1093     checkpoint = {
1094         "nb_epochs_finished": n_epoch + 1,
1095         "model_state": model.state_dict(),
1096         "rng_state": torch.get_rng_state(),
1097     }
1098
1099     if torch.cuda.is_available():
1100         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
1101
1102     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
1103     torch.save(checkpoint, checkpoint_name)
1104     log_string(f"saved checkpoint {checkpoint_name}")
1105
1106 ######################################################################