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 lambda_file = open(os.path.join(args.result_dir, "lambda.dat"), "a")
 457
 458 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 459
 460 if args.seed >= 0:
 461     # torch.backends.cudnn.deterministic = True
 462     # torch.backends.cudnn.benchmark = False
 463     # torch.use_deterministic_algorithms(True)
 464     torch.manual_seed(args.seed)
 465     if torch.cuda.is_available():
 466         torch.cuda.manual_seed_all(args.seed)
 467
 468 ######################################################################
 469
 470
 471 def log_string(s):
 472     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 473
 474     if log_file is not None:
 475         log_file.write(t + s + "\n")
 476         log_file.flush()
 477
 478     print(t + s)
 479     sys.stdout.flush()
 480
 481
 482 with os.popen("sha256sum *.py") as f:
 483     for l in f:
 484         log_string(f"sha256sum {l.strip()}")
 485
 486 now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
 487 os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh")
 488
 489 log_string(f"argv {' '.join(sys.argv)}")
 490
 491 for n in vars(args):
 492     log_string(f"args.{n} {getattr(args, n)}")
 493
 494 for k, v in sup_args.items():
 495     log_string(f'sup_args["{k}"] "{v}"')
 496
 497
 498 ######################################################################
 499
 500
 501 def get_lr(n_epoch, it):
 502     if args.legacy_lr_schedule:
 503         # my crude scheduling to compare to previous baseline, added
 504         # warmup though
 505
 506         if it < args.nb_warmup_iter:
 507             return args.legacy_large_lr * it / args.nb_warmup_iter
 508         elif n_epoch < args.legacy_nb_epoch_large_lr:
 509             return args.legacy_large_lr
 510         else:
 511             return args.legacy_small_lr
 512
 513     # from nanoGPT
 514
 515     # 1) linear warmup for warmup_iter steps
 516     if it < args.nb_warmup_iter:
 517         return args.learning_rate * it / args.nb_warmup_iter
 518     # 2) if it > nb_decay_iter, return min learning rate
 519     if it > args.nb_decay_iter:
 520         return args.min_learning_rate
 521     # 3) in between, use cosine decay down to min learning rate
 522     decay_ratio = (it - args.nb_warmup_iter) / (
 523         args.nb_decay_iter - args.nb_warmup_iter
 524     )
 525     coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))  # coeff ranges 0..1
 526     return args.min_learning_rate + coeff * (
 527         args.learning_rate - args.min_learning_rate
 528     )
 529
 530
 531 ######################################################################
 532
 533
 534 def add_memex_v1(batches, memex_proba, marker_token):
 535     for input in batches:
 536         if torch.rand(1).item() < memex_proba:
 537             t = (
 538                 torch.arange(1 + 2 * input.size(1), device=input.device)[None, :]
 539                 .expand(input.size(0), -1)
 540                 .clone()
 541             )
 542
 543             u0 = torch.randint(input.size(1), (input.size(0), 1), device=input.device)
 544             caterpillar_length = args.nb_lines // args.caterpillar_height
 545             u1 = (
 546                 u0
 547                 + torch.randint(
 548                     caterpillar_length, (input.size(0), 1), device=input.device
 549                 )
 550                 + 1
 551             )
 552
 553             m0 = (t < u0).long()
 554             m1 = (t >= u1).long() * (t < u1 + input.size(1)).long()
 555
 556             t = t * m0 + ((-1) * (1 - m0) * (1 - m1)) + (t - u1) * m1
 557             m = (t < 0).long()
 558             n = torch.arange(input.size(0), device=input.device)[:, None].expand(
 559                 -1, t.size(1)
 560             )
 561
 562             new_input = input[n, t.clamp(min=0)]
 563             new_input = (1 - m) * new_input + m * (marker_token)
 564
 565             memex_mask = new_input.new_zeros(new_input.size())
 566             memex_mask[:, input.size(1) :] = 1.0
 567
 568             yield new_input, memex_mask
 569
 570         yield input
 571
 572
 573 # The marker token is not used for this one
 574 def add_memex_v2(batches, memex_proba, marker_token):
 575     for input in batches:
 576         if torch.rand(1).item() < memex_proba:
 577             t = torch.arange(input.size(1) // 4, device=input.device)[None, :].expand(
 578                 input.size(0), -1
 579             )
 580             t = t + torch.randint(
 581                 input.size(1) - t.size(1), (t.size(0), 1), device=t.device
 582             )
 583             n = torch.arange(input.size(0), device=input.device)[:, None].expand(
 584                 -1, t.size(1)
 585             )
 586
 587             flash = input[n, t]
 588             new_input = torch.cat([input, flash], dim=1)
 589
 590             memex_mask = new_input.new_zeros(new_input.size())
 591             memex_mask[:, input.size(1) :] = 1.0
 592
 593             yield new_input, memex_mask
 594
 595         else:
 596             yield input
 597
 598
 599 def add_memex_v3(batches, memex_proba, marker_token):
 600     for input in batches:
 601         if torch.rand(1).item() < memex_proba:
 602             memex_len = input.size(1) // 4
 603
 604             t = torch.arange(input.size(1) + memex_len, device=input.device)[
 605                 None, :
 606             ].expand(input.size(0), -1)
 607             n = torch.arange(input.size(0), device=input.device)[:, None].expand(
 608                 -1, t.size(1)
 609             )
 610
 611             # Call me the tensor-spaghetti master
 612
 613             trigger = torch.rand(t.size(), device=t.device)
 614             trigger[:, -memex_len:] = 2.0
 615             trigger[:, 0] = 2.0
 616             trigger = (trigger == trigger.min(dim=1, keepdim=True).values).long()
 617             memex_mask = trigger.clone()
 618             memex_mask[:, memex_len:] -= trigger[:, :-memex_len]
 619             memex_mask = memex_mask.cumsum(dim=1)
 620
 621             u = 1 - memex_mask
 622             u[:, 0] = 0
 623             u = u.cumsum(dim=1)
 624             assert u.min() == 0
 625             assert u.max() == input.size(1) - 1
 626
 627             v = (
 628                 (trigger.cumsum(dim=1) - trigger).cumsum(dim=1)
 629                 + torch.randint(
 630                     input.size(1) - memex_len, (input.size(0), 1), device=t.device
 631                 )
 632             ) * memex_mask
 633             assert v.min() >= 0
 634             assert v.max() < input.size(1)
 635             u = u * (1 - memex_mask) + v * memex_mask
 636
 637             new_input = input[n, u]
 638             assert input.max() < vocabulary_size
 639             assert new_input.max() < vocabulary_size
 640             limits = trigger.clone()
 641             limits[:, memex_len - 1 :] += limits[:, : -(memex_len - 1)]
 642             assert limits.min() == 0
 643             assert limits.max() == 1
 644             new_input = new_input * (1 - limits) + marker_token * limits
 645             assert marker_token < vocabulary_size
 646             assert new_input.max() < vocabulary_size
 647
 648             yield new_input, memex_mask
 649
 650         else:
 651             yield input
 652
 653
 654 ######################################################################
 655
 656 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 657
 658 assert args.batch_size % args.physical_batch_size == 0
 659
 660
 661 def picoclvr_pruner_horizontal_green(p):
 662     return not ("green" in p and ("left" in p or "right" in p))
 663
 664
 665 picoclvr_pruner_train = (
 666     picoclvr_pruner_horizontal_green
 667     if args.picocvlr_prune_properties in {"train+eval"}
 668     else None
 669 )
 670
 671 picoclvr_pruner_eval = (
 672     (lambda p: not picoclvr_pruner_horizontal_green(p))
 673     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 674     else None
 675 )
 676
 677 ######################################################################
 678
 679 device_data = device
 680
 681 if args.task == "byheart":
 682     task = tasks.SandBox(
 683         problem=problems.ProblemByHeart(),
 684         nb_train_samples=args.nb_train_samples,
 685         nb_test_samples=args.nb_test_samples,
 686         batch_size=args.physical_batch_size,
 687         logger=log_string,
 688         device=device_data,
 689     )
 690     args.max_percents_of_test_in_train = -1
 691
 692 elif args.task == "learnop":
 693     task = tasks.SandBox(
 694         problem=problems.ProblemLearnOperator(),
 695         nb_train_samples=args.nb_train_samples,
 696         nb_test_samples=args.nb_test_samples,
 697         batch_size=args.physical_batch_size,
 698         logger=log_string,
 699         device=device_data,
 700     )
 701
 702
 703 elif args.task == "guessop":
 704     task = tasks.SandBox(
 705         problem=problems.ProblemGuessOperator(),
 706         nb_train_samples=args.nb_train_samples,
 707         nb_test_samples=args.nb_test_samples,
 708         batch_size=args.physical_batch_size,
 709         logger=log_string,
 710         device=device_data,
 711     )
 712
 713
 714 elif args.task == "twotargets":
 715     task = tasks.SandBox(
 716         problem=problems.ProblemTwoTargets(),
 717         nb_train_samples=args.nb_train_samples,
 718         nb_test_samples=args.nb_test_samples,
 719         batch_size=args.physical_batch_size,
 720         logger=log_string,
 721         device=device_data,
 722     )
 723
 724 elif args.task == "memory":
 725     task = tasks.SandBox(
 726         problem=problems.ProblemMemory(len_total=args.memory_len_total),
 727         nb_train_samples=args.nb_train_samples,
 728         nb_test_samples=args.nb_test_samples,
 729         batch_size=args.physical_batch_size,
 730         logger=log_string,
 731         device=device_data,
 732     )
 733
 734 elif args.task == "mixing":
 735     task = tasks.SandBox(
 736         problem=problems.ProblemMixing(
 737             hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
 738         ),
 739         nb_train_samples=args.nb_train_samples,
 740         nb_test_samples=args.nb_test_samples,
 741         batch_size=args.physical_batch_size,
 742         logger=log_string,
 743         device=device_data,
 744     )
 745
 746 elif args.task == "addition":
 747     task = tasks.SandBox(
 748         problem=problems.ProblemAddition(),
 749         nb_train_samples=args.nb_train_samples,
 750         nb_test_samples=args.nb_test_samples,
 751         batch_size=args.physical_batch_size,
 752         logger=log_string,
 753         device=device_data,
 754     )
 755
 756 elif args.task == "picoclvr":
 757     task = tasks.PicoCLVR(
 758         nb_train_samples=args.nb_train_samples,
 759         nb_test_samples=args.nb_test_samples,
 760         batch_size=args.physical_batch_size,
 761         height=args.picoclvr_height,
 762         width=args.picoclvr_width,
 763         nb_colors=args.picoclvr_nb_colors,
 764         logger=log_string,
 765         device=device_data,
 766         pruner_train=picoclvr_pruner_train,
 767         pruner_eval=picoclvr_pruner_eval,
 768     )
 769
 770 elif args.task == "mnist":
 771     task = tasks.MNIST(
 772         nb_train_samples=args.nb_train_samples,
 773         nb_test_samples=args.nb_test_samples,
 774         batch_size=args.physical_batch_size,
 775         device=device_data,
 776     )
 777
 778 elif args.task == "maze":
 779     task = tasks.Maze(
 780         nb_train_samples=args.nb_train_samples,
 781         nb_test_samples=args.nb_test_samples,
 782         batch_size=args.physical_batch_size,
 783         height=args.maze_height,
 784         width=args.maze_width,
 785         nb_walls=args.maze_nb_walls,
 786         device=device_data,
 787     )
 788
 789 elif args.task == "snake":
 790     task = tasks.Snake(
 791         nb_train_samples=args.nb_train_samples,
 792         nb_test_samples=args.nb_test_samples,
 793         batch_size=args.physical_batch_size,
 794         height=args.snake_height,
 795         width=args.snake_width,
 796         nb_colors=args.snake_nb_colors,
 797         length=args.snake_length,
 798         prompt_length=args.snake_length // 2,
 799         device=device_data,
 800     )
 801
 802 elif args.task == "stack":
 803     task = tasks.Stack(
 804         nb_train_samples=args.nb_train_samples,
 805         nb_test_samples=args.nb_test_samples,
 806         batch_size=args.physical_batch_size,
 807         logger=log_string,
 808         nb_steps=args.stack_nb_steps,
 809         nb_stacks=args.stack_nb_stacks,
 810         nb_digits=args.stack_nb_digits,
 811         fraction_values_for_train=args.stack_fraction_values_for_train,
 812         device=device_data,
 813     )
 814
 815 elif args.task == "expr":
 816     task = tasks.Expr(
 817         nb_train_samples=args.nb_train_samples,
 818         nb_test_samples=args.nb_test_samples,
 819         nb_variables=args.expr_nb_variables,
 820         sequence_length=args.expr_sequence_length,
 821         operand_max=args.expr_operand_max,
 822         result_max=args.expr_result_max,
 823         batch_size=args.physical_batch_size,
 824         device=device_data,
 825     )
 826
 827 elif args.task == "rpl":
 828     task = tasks.RPL(
 829         nb_train_samples=args.nb_train_samples,
 830         nb_test_samples=args.nb_test_samples,
 831         batch_size=args.physical_batch_size,
 832         nb_starting_values=args.rpl_nb_starting_values,
 833         max_input=args.rpl_max_input,
 834         prog_len=args.rpl_prog_len,
 835         nb_runs=args.rpl_nb_runs,
 836         no_prog=args.rpl_no_prog,
 837         logger=log_string,
 838         device=device_data,
 839     )
 840
 841 elif args.task == "grid":
 842     task = tasks.Grid(
 843         nb_train_samples=args.nb_train_samples,
 844         nb_test_samples=args.nb_test_samples,
 845         batch_size=args.physical_batch_size,
 846         size=args.grid_size,
 847         nb_shapes=args.grid_nb_shapes,
 848         nb_colors=args.grid_nb_colors,
 849         logger=log_string,
 850         device=device_data,
 851     )
 852
 853 elif args.task == "qmlp":
 854     task = tasks.QMLP(
 855         nb_train_samples=args.nb_train_samples,
 856         nb_test_samples=args.nb_test_samples,
 857         batch_size=args.physical_batch_size,
 858         result_dir=args.result_dir,
 859         logger=log_string,
 860         device=device_data,
 861     )
 862
 863 else:
 864     raise ValueError(f"Unknown task {args.task}")
 865
 866 ######################################################################
 867
 868 log_string(f"device {device}")
 869
 870 vocabulary_size = task.vocabulary_size()
 871
 872 if args.memex_proba > 0:
 873     memex_marker = vocabulary_size
 874     vocabulary_size += 1
 875
 876 log_string(f"vocabulary_size {vocabulary_size}")
 877
 878 ##############################
 879
 880 model = mygpt.MyGPT(
 881     vocabulary_size=vocabulary_size,
 882     dim_model=args.dim_model,
 883     dim_keys=args.dim_keys,
 884     dim_hidden=args.dim_hidden,
 885     nb_heads=args.nb_heads,
 886     nb_lines=args.nb_lines,
 887     caterpillar_height=args.caterpillar_height,
 888     nb_blocks=args.nb_blocks,
 889     causal=True,
 890     dropout=args.dropout,
 891     attention_layer=args.attention,
 892     logger=log_string,
 893     args=args,
 894 )
 895
 896 model.to(device)
 897
 898 nb_parameters = sum(p.numel() for p in model.parameters())
 899 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 900
 901 ######################################################################
 902
 903 nb_epochs_finished = 0
 904
 905 if args.no_checkpoint:
 906     log_string(f"not trying to load checkpoint.")
 907
 908 else:
 909     try:
 910         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 911         checkpoint = torch.load(checkpoint_name)
 912         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 913         model.load_state_dict(checkpoint["model_state"])
 914         torch.set_rng_state(checkpoint["rng_state"])
 915         if torch.cuda.is_available():
 916             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 917
 918         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 919
 920     except FileNotFoundError:
 921         log_string("starting from scratch.")
 922
 923     except:
 924         log_string("error when loading the checkpoint.")
 925         exit(1)
 926
 927 ######################################################################
 928
 929 if args.task == "expr" and args.expr_input_file is not None:
 930     task.produce_results(
 931         n_epoch=nb_epochs_finished,
 932         model=model,
 933         result_dir=args.result_dir,
 934         logger=log_string,
 935         deterministic_synthesis=args.deterministic_synthesis,
 936         input_file=args.expr_input_file,
 937     )
 938
 939     exit(0)
 940
 941 ######################################################################
 942
 943 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 944
 945 # Compute the entropy of the training tokens
 946
 947 token_count = 0
 948 for input in task.batches(split="train"):
 949     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 950 token_probas = token_count / token_count.sum()
 951 entropy = -torch.xlogy(token_probas, token_probas).sum()
 952 train_set_perplexity = math.exp(entropy)
 953
 954 ######################################################################
 955 # A bit of paranoia never hurts
 956
 957 if args.max_percents_of_test_in_train >= 0:
 958
 959     def subsets_as_tuples(batches, cs):
 960         s = set()
 961         for batch in batches:
 962             for x in batch:
 963                 s.add(tuple([v.item() for v in x]))
 964                 if len(s) == cs:
 965                     yield s
 966                     s = set()
 967         yield s
 968
 969     nb_test, nb_in_train = 0, 0
 970     for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
 971         in_train = set()
 972         for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
 973             in_train.update(test_subset.intersection(train_subset))
 974         nb_in_train += len(in_train)
 975         nb_test += len(test_subset)
 976
 977     log_string(
 978         f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
 979     )
 980
 981     assert (
 982         nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
 983     ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
 984
 985 ##############################
 986
 987 if "calibrate" in sup_args:
 988     for input in task.batches(split="train", desc="calibrate"):
 989         input = input.to(device)
 990         output = model(mygpt.BracketedSequence(input)).x
 991
 992     for n, m in model.named_modules():
 993         for a in dir(m):
 994             x = getattr(m, a)
 995             if isinstance(x, mygpt.Calibrator):
 996                 print(f"####### ${n} | ${a} ########################")
 997                 mean, std = x.moments()
 998                 print("mean\n", mean, "\n")
 999                 print("std\n", std, "\n")
1000                 print(f"############################################\n\n")
1001
1002     exit(0)
1003
1004 ##############################
1005
1006 nb_samples_seen = 0
1007
1008 if nb_epochs_finished >= nb_epochs:
1009     task.produce_results(
1010         n_epoch=nb_epochs_finished,
1011         model=model,
1012         result_dir=args.result_dir,
1013         logger=log_string,
1014         deterministic_synthesis=args.deterministic_synthesis,
1015     )
1016
1017 time_pred_result = datetime.datetime.now()
1018
1019 it = 0
1020
1021 n_batch = 0
1022
1023
1024 def the_dot_products(value1, value2, params):
1025     g1g1, g1g2, g2g2 = 0, 0, 0
1026     for p in params:
1027         g1 = torch.autograd.grad(value1, p, retain_graph=True)[0]
1028         g2 = torch.autograd.grad(value2, p, retain_graph=True)[0]
1029         g1g1 += g1.pow(2).sum()[None]
1030         g2g2 += g2.pow(2).sum()[None]
1031         g1g2 += (g1 * g2).sum()[None]
1032     return torch.cat([g1g1, g1g2, g2g2])
1033
1034
1035 def update_ave_grad(value, params, name, eps=1e-3):
1036     for p in params:
1037         g = torch.autograd.grad(value, p, retain_graph=True)[0]
1038         ag = getattr(p, name) if hasattr(p, name) else 0
1039         setattr(p, name, (1 - eps) * ag + eps * g)
1040
1041
1042 def norm(params, name):
1043     s = 0
1044     for p in params:
1045         s += getattr(p, name).pow(2).sum()
1046     return s
1047
1048
1049 for n_epoch in range(nb_epochs_finished, nb_epochs):
1050     if args.optim == "sgd":
1051         optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
1052     elif args.optim == "adam":
1053         optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
1054     elif args.optim == "adamw":
1055         optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate)
1056     else:
1057         raise ValueError(f"Unknown optimizer {args.optim}.")
1058
1059     model.train()
1060
1061     nb_train_samples, acc_train_loss, acc_train_inner_loss = 0, 0.0, 0.0
1062
1063     memex_proba = (
1064         args.memex_proba
1065         if args.memex_nb_epochs is None or n_epoch < args.memex_nb_epochs
1066         else 0.0
1067     )
1068
1069     log_string(f"memex_proba {memex_proba}")
1070
1071     warnings.warn("memex v3", RuntimeWarning)
1072     train_batches = add_memex_v3(
1073         batches=task.batches(split="train"),
1074         memex_proba=memex_proba,
1075         marker_token=memex_marker,
1076     )
1077
1078     def add_none(it):
1079         for x in it:
1080             yield x
1081         yield None
1082
1083     nb_acc_samples = 0
1084
1085     for input in add_none(train_batches):
1086         if input is not None:
1087             if type(input) is tuple:
1088                 input, memex_mask = input
1089                 memex_mask = memex_mask.to(device)
1090             else:
1091                 memex_mask = None
1092
1093             model.reset_inner_loss()
1094             input = input.to(device)
1095
1096             output = model(mygpt.BracketedSequence(input)).x
1097
1098             if memex_mask is None:
1099                 loss = F.cross_entropy(output.transpose(1, 2), input)
1100             else:
1101                 loss = F.cross_entropy(output.transpose(1, 2), input, reduction="none")
1102                 loss_regular = (loss * (1 - memex_mask)).mean()
1103                 loss_memex = (loss * memex_mask).mean()
1104
1105                 if it < 100 or torch.rand(1) < 0.01:
1106                     update_ave_grad(loss_regular, model.parameters(), "grad_regular")
1107                     update_ave_grad(loss_memex, model.parameters(), "grad_memex")
1108                     norm_regular = norm(model.parameters(), "grad_regular")
1109                     norm_memex = norm(model.parameters(), "grad_memex")
1110                     l_memex = (
1111                         max(norm_regular, norm_memex) - norm_regular
1112                     ) / norm_memex
1113
1114                 loss = loss_regular + l_memex * loss_memex
1115
1116             inner_loss = model.get_inner_loss()
1117
1118             acc_train_loss += loss.item() * input.size(0)
1119             acc_train_inner_loss += inner_loss.item() * input.size(0)
1120
1121             nb_train_samples += input.size(0)
1122             nb_samples_seen += input.size(0)
1123
1124             total_loss = loss + (
1125                 args.rho_inner_loss * inner_loss if args.rho_inner_loss > 0 else 0.0
1126             )
1127
1128             it += 1
1129             lr = get_lr(n_epoch, it)
1130             for param_group in optimizer.param_groups:
1131                 param_group["lr"] = lr
1132
1133                 # log_string(f"learning_rate {lr}")
1134
1135             total_loss.backward()
1136             nb_acc_samples += input.size(0)
1137
1138         if (input is None and nb_acc_samples > 0) or nb_acc_samples == args.batch_size:
1139             assert nb_acc_samples <= args.batch_size
1140             optimizer.step()
1141             grad_norm = sum([p.grad.pow(2).sum() for p in model.parameters()]).sqrt()
1142             loss_file.write(f"{n_epoch} {n_batch} {loss.item()} {grad_norm.item()}\n")
1143             lambda_file.write(
1144                 f"{n_epoch} {n_batch} {l_memex} {norm_regular} {norm_memex}\n"
1145             )
1146             optimizer.zero_grad()
1147             nb_acc_samples = 0
1148             n_batch += 1
1149
1150     with torch.autograd.no_grad():
1151         model.eval()
1152
1153         nb_test_samples, acc_test_loss = 0, 0.0
1154
1155         for input in task.batches(split="test"):
1156             input = input.to(device)
1157
1158             output = model(mygpt.BracketedSequence(input)).x
1159             loss = F.cross_entropy(output.transpose(1, 2), input)
1160             acc_test_loss += loss.item() * input.size(0)
1161             nb_test_samples += input.size(0)
1162
1163         log_string(
1164             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}"
1165         )
1166
1167         task.produce_results(
1168             n_epoch=n_epoch,
1169             model=model,
1170             result_dir=args.result_dir,
1171             logger=log_string,
1172             deterministic_synthesis=args.deterministic_synthesis,
1173         )
1174
1175         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
1176         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
1177
1178         log_string(
1179             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
1180         )
1181
1182         time_current_result = datetime.datetime.now()
1183         log_string(
1184             f"next_result {time_current_result + (time_current_result - time_pred_result)}"
1185         )
1186         time_pred_result = time_current_result
1187
1188     checkpoint = {
1189         "nb_epochs_finished": n_epoch + 1,
1190         "model_state": model.state_dict(),
1191         "rng_state": torch.get_rng_state(),
1192     }
1193
1194     if torch.cuda.is_available():
1195         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
1196
1197     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
1198     torch.save(checkpoint, checkpoint_name)
1199     log_string(f"saved checkpoint {checkpoint_name}")
1200
1201 ######################################################################