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 # torch.backends.cuda.matmul.allow_tf23
   9 # torch.autocast(torch.bfloat16)
  10
  11 import math, sys, argparse, time, tqdm, os
  12
  13 import torch, torchvision
  14 from torch import nn
  15 from torch.nn import functional as F
  16
  17 import mygpt, tensorstack
  18
  19 ######################################################################
  20
  21 if torch.cuda.is_available():
  22     device = torch.device("cuda")
  23     torch.backends.cuda.matmul.allow_tf32 = True
  24 else:
  25     device = torch.device("cpu")
  26
  27 ######################################################################
  28
  29 parser = argparse.ArgumentParser(
  30     description="An implementation of GPT with cache.",
  31     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
  32 )
  33
  34 parser.add_argument(
  35     "--task", type=str, default="picoclvr", help="picoclvr, mnist, maze, snake, stack"
  36 )
  37
  38 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
  39
  40 parser.add_argument("--result_dir", type=str, default=None)
  41
  42 parser.add_argument("--seed", type=int, default=0)
  43
  44 parser.add_argument("--nb_epochs", type=int, default=None)
  45
  46 parser.add_argument("--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("--optim", type=str, default="adam")
  53
  54 parser.add_argument("--learning_rate", type=float, default=1e-4)
  55
  56 parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
  57
  58 parser.add_argument("--dim_model", type=int, default=512)
  59
  60 parser.add_argument("--dim_keys", type=int, default=64)
  61
  62 parser.add_argument("--dim_hidden", type=int, default=2048)
  63
  64 parser.add_argument("--nb_heads", type=int, default=8)
  65
  66 parser.add_argument("--nb_blocks", type=int, default=12)
  67
  68 parser.add_argument("--dropout", type=float, default=0.1)
  69
  70 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  71
  72 parser.add_argument("--no_checkpoint", action="store_true", default=False)
  73
  74 parser.add_argument("--overwrite_results", action="store_true", default=False)
  75
  76 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
  77
  78 ##############################
  79 # picoclvr options
  80
  81 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
  82
  83 parser.add_argument("--picoclvr_height", type=int, default=12)
  84
  85 parser.add_argument("--picoclvr_width", type=int, default=16)
  86
  87 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
  88
  89 ##############################
  90 # Maze options
  91
  92 parser.add_argument("--maze_height", type=int, default=13)
  93
  94 parser.add_argument("--maze_width", type=int, default=21)
  95
  96 parser.add_argument("--maze_nb_walls", type=int, default=15)
  97
  98 ##############################
  99 # Snake options
 100
 101 parser.add_argument("--snake_height", type=int, default=6)
 102
 103 parser.add_argument("--snake_width", type=int, default=8)
 104
 105 parser.add_argument("--snake_nb_colors", type=int, default=5)
 106
 107 parser.add_argument("--snake_length", type=int, default=200)
 108
 109 ##############################
 110 # Snake options
 111
 112 parser.add_argument("--stack_nb_steps", type=int, default=100)
 113
 114 parser.add_argument("--stack_nb_stacks", type=int, default=1)
 115
 116 parser.add_argument("--stack_nb_digits", type=int, default=3)
 117
 118 parser.add_argument("--stack_fraction_values_for_train", type=float, default=None)
 119
 120 ######################################################################
 121
 122 args = parser.parse_args()
 123
 124 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 125
 126 try:
 127     os.mkdir(args.result_dir)
 128 except FileExistsError:
 129     if not args.overwrite_results:
 130         print(f"result directory {args.result_dir} already exists")
 131         exit(1)
 132
 133 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 134
 135 if args.seed >= 0:
 136     # torch.backends.cudnn.deterministic = True
 137     # torch.backends.cudnn.benchmark = False
 138     # torch.use_deterministic_algorithms(True)
 139     torch.manual_seed(args.seed)
 140     if torch.cuda.is_available():
 141         torch.cuda.manual_seed_all(args.seed)
 142
 143 ######################################################################
 144
 145 default_args = {
 146     "picoclvr": {
 147         "result_dir": "results_picoclvr",
 148         "nb_epochs": 25,
 149         "batch_size": 25,
 150         "nb_train_samples": 250000,
 151         "nb_test_samples": 10000,
 152     },
 153     "mnist": {
 154         "result_dir": "results_mnist",
 155         "nb_epochs": 25,
 156         "batch_size": 10,
 157         "nb_train_samples": 250000,
 158         "nb_test_samples": 10000,
 159     },
 160     "maze": {
 161         "result_dir": "results_maze",
 162         "nb_epochs": 25,
 163         "batch_size": 25,
 164         "nb_train_samples": 250000,
 165         "nb_test_samples": 10000,
 166     },
 167     "snake": {
 168         "result_dir": "results_snake",
 169         "nb_epochs": 5,
 170         "batch_size": 25,
 171         "nb_train_samples": 250000,
 172         "nb_test_samples": 10000,
 173     },
 174     "stack": {
 175         "result_dir": "results_stack",
 176         "nb_epochs": 5,
 177         "batch_size": 25,
 178         "nb_train_samples": 100000,
 179         "nb_test_samples": 1000,
 180     },
 181 }
 182
 183 if args.task in default_args:
 184     for k, v in default_args[args.task].items():
 185         if getattr(args, k) is None:
 186             setattr(args, k, v)
 187
 188 ######################################################################
 189
 190
 191 def log_string(s):
 192     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 193
 194     if log_file is not None:
 195         log_file.write(t + s + "\n")
 196         log_file.flush()
 197
 198     print(t + s)
 199     sys.stdout.flush()
 200
 201
 202 for n in vars(args):
 203     log_string(f"args.{n} {getattr(args, n)}")
 204
 205 ######################################################################
 206
 207
 208 # ra_mask is boolean, with 1s on the values to generate
 209
 210
 211 def masked_inplace_autoregression(
 212     model,
 213     batch_size,
 214     input,
 215     ar_mask,
 216     forbidden_tokens=None,
 217     progress_bar_desc="autoregression",
 218     device=torch.device("cpu"),
 219 ):
 220     # p = logits.softmax(1)
 221     # entropy[:,s]= p.xlogy(p).sum(1) / math.log(2)
 222     batches = zip(input.split(batch_size), ar_mask.split(batch_size))
 223     if progress_bar_desc is not None:
 224         batches = tqdm.tqdm(
 225             batches,
 226             dynamic_ncols=True,
 227             desc=progress_bar_desc,
 228             total=input.size(0) // batch_size,
 229         )
 230     for input, ar_mask in batches:
 231         i = (ar_mask.sum(0) > 0).nonzero()
 232         if i.min() > 0:
 233             model(
 234                 mygpt.BracketedSequence(input, 0, i.min())
 235             )  # Needed to initialize the model's cache
 236         for s in range(i.min(), i.max() + 1):
 237             output = model(mygpt.BracketedSequence(input, s, 1)).x
 238             logits = output[:, s]
 239             if forbidden_tokens is not None:
 240                 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
 241             if args.deterministic_synthesis:
 242                 t_next = logits.argmax(1)
 243             else:
 244                 dist = torch.distributions.categorical.Categorical(logits=logits)
 245                 t_next = dist.sample()
 246             input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 247
 248
 249 ######################################################################
 250
 251
 252 class Task:
 253     def batches(self, split="train"):
 254         pass
 255
 256     def vocabulary_size(self):
 257         pass
 258
 259     def produce_results(self, n_epoch, model):
 260         pass
 261
 262
 263 ######################################################################
 264
 265 import picoclvr
 266
 267
 268 class TaskPicoCLVR(Task):
 269     # Make a tensor from a list of strings
 270     def tensorize(self, descr):
 271         token_descr = [s.strip().split(" ") for s in descr]
 272         l = max([len(s) for s in token_descr])
 273         token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
 274         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
 275         return torch.tensor(id_descr, device=self.device)
 276
 277     # Make a list of strings from a tensor
 278     def detensorize(self, x):
 279         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
 280
 281     # trim all the tensors in the tuple z to remove as much token from
 282     # left and right in the first tensor. If z is a tuple, all its
 283     # elements are trimed according to the triming for the first
 284     def trim(self, z, token="<nul>"):
 285         n = self.token2id[token]
 286         if type(z) == tuple:
 287             x = z[0]
 288             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 289             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 290             return tuple([t[:, a:b] for t in z])
 291         else:
 292             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 293             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 294             return z[:, a:b]
 295
 296     ######################
 297     # Not the cleanest part of the code
 298
 299     # Extract the last image of each sequence, from the last <img>
 300     # included, and set to <nul> all the tokens from the beginning of
 301     # that image to the end
 302     def excise_last_image(self, input):
 303         t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 304         nb_img_tokens = self.height * self.width + 1
 305
 306         input = input.clone()
 307         t = (input == t_img).long()
 308         tail_masks = (t.cumsum(dim=1) == t.sum(dim=1, keepdim=True)).long()
 309         i = (t * tail_masks).nonzero(as_tuple=True)
 310         j = (
 311             i[0][:, None],
 312             i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
 313         )
 314         images = self.trim(input[j])
 315         input[j] = t_nul
 316         loss_masks = 1 - tail_masks
 317         input, loss_masks = self.trim((input, loss_masks))
 318         return input, loss_masks, images
 319
 320     def add_true_image(self, input, images, loss_masks):
 321         t_nul = self.token2id["<nul>"]
 322         nb_img_tokens = self.height * self.width + 1
 323         input = F.pad(input, (0, nb_img_tokens), value=t_nul)
 324         loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
 325         t = (input == t_nul).long()
 326         i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
 327         j = (
 328             i[0][:, None],
 329             i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
 330         )
 331         input[j] = images
 332         loss_masks[j] = 1
 333         input, loss_masks = self.trim((input, loss_masks))
 334         return input, loss_masks
 335
 336     def add_generated_image(self, input, loss_masks, model):
 337         t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 338         nb_img_tokens = self.height * self.width + 1
 339
 340         input = F.pad(input, (0, nb_img_tokens), value=t_nul)
 341         loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
 342         t = (input == t_nul).long()
 343         i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
 344         input[i] = t_img
 345
 346         j = (
 347             i[0][:, None],
 348             i[1][:, None]
 349             + 1
 350             + torch.arange(nb_img_tokens - 1, device=input.device)[None, :],
 351         )
 352         ar_masks = input.new_zeros(input.size(), dtype=torch.int64)
 353         ar_masks[j] = 1
 354         forbidden_tokens = (
 355             torch.arange(self.vocabulary_size(), device=input.device) == t_nul
 356         )
 357         with torch.autograd.no_grad():
 358             t = model.training
 359             model.eval()
 360             masked_inplace_autoregression(
 361                 model,
 362                 self.batch_size,
 363                 input,
 364                 ar_masks,
 365                 forbidden_tokens,
 366                 progress_bar_desc=None,
 367                 device=self.device,
 368             )
 369             model.train(t)
 370
 371         input, loss_masks = self.trim((input, loss_masks))
 372
 373         return input, loss_masks
 374
 375     ######################
 376
 377     def __init__(
 378         self,
 379         nb_train_samples,
 380         nb_test_samples,
 381         batch_size,
 382         height,
 383         width,
 384         nb_colors=5,
 385         device=torch.device("cpu"),
 386         pruner_train=None,
 387         pruner_eval=None,
 388     ):
 389         def generate_descr(nb, cache_suffix, pruner):
 390             return picoclvr.generate(
 391                 nb,
 392                 height=self.height,
 393                 width=self.width,
 394                 nb_colors=nb_colors,
 395                 pruner=pruner,
 396             )
 397
 398         self.height = height
 399         self.width = width
 400         self.batch_size = batch_size
 401         self.device = device
 402         self.pruner_train = pruner_train
 403         self.pruner_eval = pruner_eval
 404
 405         param = {
 406             "nb_train_samples": nb_train_samples,
 407             "nb_test_samples": nb_test_samples,
 408             "height": height,
 409             "width": width,
 410             "nb_colors": nb_colors,
 411             "batch_size": batch_size,
 412             "rng_state": list(torch.get_rng_state()),
 413         }
 414
 415         log_string(
 416             f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
 417         )
 418         self.train_descr = generate_descr(
 419             nb_train_samples, "train", pruner=self.pruner_train
 420         )
 421         self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
 422
 423         # Build the tokenizer
 424         tokens = {"<nul>", "<img>"}
 425         for d in [self.train_descr, self.test_descr]:
 426             for s in d:
 427                 for t in s.strip().split(" "):
 428                     tokens.add(t)
 429         # make this set a sorted list to get the same tensors given
 430         # the same descr
 431         tokens = list(tokens)
 432         tokens.sort()
 433         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
 434         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
 435
 436         # Tokenize the train and test sets
 437         self.train_input = self.tensorize(self.train_descr)
 438         self.test_input = self.tensorize(self.test_descr)
 439
 440     def batches(self, split="train"):
 441         assert split in {"train", "test"}
 442         input = self.train_input if split == "train" else self.test_input
 443         for batch in tqdm.tqdm(
 444             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 445         ):
 446             yield self.trim(batch)
 447
 448     def vocabulary_size(self):
 449         return len(self.token2id)
 450
 451     def compute_missing_properties(self, n_epoch, model, pruner=None):
 452         acc_nb_requested_properties = []
 453         acc_nb_missing_properties = []
 454         acc_nb_results = 0
 455
 456         for input in tqdm.tqdm(
 457             self.test_input.split(self.batch_size),
 458             dynamic_ncols=True,
 459             desc=f"test-properties",
 460         ):
 461             tape, loss_masks, _ = self.excise_last_image(input)
 462             tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
 463             result_descr = self.detensorize(tape)
 464             np = picoclvr.nb_properties(
 465                 result_descr,
 466                 height=self.height,
 467                 width=self.width,
 468                 pruner=pruner,
 469             )
 470             nb_requested_properties, _, nb_missing_properties = zip(*np)
 471             acc_nb_requested_properties += nb_requested_properties
 472             acc_nb_missing_properties += nb_missing_properties
 473             acc_nb_results += len(result_descr)
 474
 475         nb_requested_properties = sum(acc_nb_requested_properties)
 476         nb_missing_properties = sum(acc_nb_missing_properties)
 477
 478         prefix = "" if pruner is None else "pruned_"
 479         log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 480         log_string(
 481             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 482         )
 483         log_string(
 484             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 485         )
 486
 487     ######################################################################
 488
 489     def produce_results(self, n_epoch, model):
 490         self.compute_missing_properties(n_epoch, model)
 491
 492         if self.pruner_eval is not None:
 493             self.compute_missing_properties(n_epoch, model, self.pruner_eval)
 494
 495         nb_tokens_to_generate = self.height * self.width + 3
 496         result_descr = []
 497         nb_per_primer = 8
 498         primer = []
 499
 500         for primer_descr in [
 501             "red above green <sep> green top <sep> blue right of red",
 502             "there is red <sep> there is yellow <sep> there is blue",
 503             "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
 504             "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
 505         ]:
 506             primer += [primer_descr] * nb_per_primer
 507
 508         tape = self.tensorize(primer)
 509         loss_masks = 1 - (tape == self.token2id["<nul>"]).long()
 510         tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
 511         result_descr = self.detensorize(tape)
 512
 513         np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
 514
 515         acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
 516         acc_nb_results = len(result_descr)
 517
 518         nb_requested_properties = sum(acc_nb_requested_properties)
 519         nb_missing_properties = sum(acc_nb_missing_properties)
 520
 521         prefix = "demo_"
 522         log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 523         log_string(
 524             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 525         )
 526         log_string(
 527             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 528         )
 529
 530         img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
 531
 532         if img.dim() == 5:
 533             if img.size(1) == 1:
 534                 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
 535             else:
 536                 img = torch.cat(
 537                     [
 538                         torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
 539                         for x in img
 540                     ],
 541                     0,
 542                 )
 543
 544         image_name = os.path.join(args.result_dir, f"picoclvr_result_{n_epoch:04d}.png")
 545         torchvision.utils.save_image(
 546             img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0
 547         )
 548         log_string(f"wrote {image_name}")
 549
 550
 551 ######################################################################
 552
 553
 554 class TaskMNIST(Task):
 555     def __init__(self, batch_size, device=torch.device("cpu")):
 556         self.device = device
 557         self.batch_size = batch_size
 558
 559     def batches(self, split="train"):
 560         assert split in {"train", "test"}
 561         data_set = torchvision.datasets.MNIST(
 562             root="./data", train=(split == "train"), download=True
 563         )
 564         data_input = data_set.data.view(-1, 28 * 28).long()
 565         if args.nb_train_samples is not None:
 566             data_input = data_input[: args.nb_train_samples]
 567         for batch in tqdm.tqdm(
 568             data_input.split(self.batch_size), desc=f"epoch-{split}"
 569         ):
 570             yield batch
 571
 572     def vocabulary_size(self):
 573         return 256
 574
 575     def produce_results(self, n_epoch, model):
 576         results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
 577         ar_mask = torch.full_like(results, 1)
 578         masked_inplace_autoregression(
 579             model, self.batch_size, results, ar_mask, device=self.device
 580         )
 581         image_name = os.path.join(args.result_dir, f"mnist_result_{n_epoch:04d}.png")
 582         torchvision.utils.save_image(
 583             1 - results.reshape(-1, 1, 28, 28) / 255.0,
 584             image_name,
 585             nrow=16,
 586             pad_value=0.8,
 587         )
 588         log_string(f"wrote {image_name}")
 589
 590
 591 ######################################################################
 592
 593 import maze
 594
 595
 596 class TaskMaze(Task):
 597     def map2seq(self, *m):
 598         return torch.cat([x.flatten(1) for x in m], 1)
 599
 600     def seq2map(self, s):
 601         s = s.reshape(s.size(0), -1, self.height, self.width)
 602         return (s[:, k] for k in range(s.size(1)))
 603
 604     def __init__(
 605         self,
 606         nb_train_samples,
 607         nb_test_samples,
 608         batch_size,
 609         height,
 610         width,
 611         nb_walls,
 612         device=torch.device("cpu"),
 613     ):
 614         self.batch_size = batch_size
 615         self.height = height
 616         self.width = width
 617         self.device = device
 618
 619         train_mazes, train_paths, _ = maze.create_maze_data(
 620             nb_train_samples,
 621             height=height,
 622             width=width,
 623             nb_walls=nb_walls,
 624             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 625         )
 626         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 627
 628         test_mazes, test_paths, _ = maze.create_maze_data(
 629             nb_test_samples,
 630             height=height,
 631             width=width,
 632             nb_walls=nb_walls,
 633             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 634         )
 635         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 636
 637         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 638
 639     def batches(self, split="train", nb_to_use=-1, desc=None):
 640         assert split in {"train", "test"}
 641         input = self.train_input if split == "train" else self.test_input
 642         if nb_to_use > 0:
 643             input = input[:nb_to_use]
 644         if desc is None:
 645             desc = f"epoch-{split}"
 646         for batch in tqdm.tqdm(
 647             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 648         ):
 649             yield batch
 650
 651     def vocabulary_size(self):
 652         return self.nb_codes
 653
 654     def compute_error(self, model, split="train", nb_to_use=-1):
 655         nb_total, nb_correct = 0, 0
 656         count = torch.zeros(
 657             self.width * self.height,
 658             self.width * self.height,
 659             device=self.device,
 660             dtype=torch.int64,
 661         )
 662         for input in tqdm.tqdm(
 663             task.batches(split, nb_to_use),
 664             dynamic_ncols=True,
 665             desc=f"test-mazes",
 666         ):
 667             result = input.clone()
 668             ar_mask = result.new_zeros(result.size())
 669             ar_mask[:, self.height * self.width :] = 1
 670             result *= 1 - ar_mask
 671             masked_inplace_autoregression(
 672                 model,
 673                 self.batch_size,
 674                 result,
 675                 ar_mask,
 676                 progress_bar_desc=None,
 677                 device=self.device,
 678             )
 679             mazes, paths = self.seq2map(result)
 680             path_correctness = maze.path_correctness(mazes, paths)
 681             nb_correct += path_correctness.long().sum()
 682             nb_total += mazes.size(0)
 683
 684             optimal_path_lengths = (
 685                 (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
 686             )
 687             predicted_path_lengths = (
 688                 (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
 689             )
 690             optimal_path_lengths = optimal_path_lengths[path_correctness]
 691             predicted_path_lengths = predicted_path_lengths[path_correctness]
 692             count[optimal_path_lengths, predicted_path_lengths] += 1
 693
 694         if count.max() == 0:
 695             count = None
 696         else:
 697             count = count[
 698                 : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
 699             ]
 700
 701         return nb_total, nb_correct, count
 702
 703     def produce_results(self, n_epoch, model):
 704         with torch.autograd.no_grad():
 705             t = model.training
 706             model.eval()
 707
 708             train_nb_total, train_nb_correct, count = self.compute_error(
 709                 model, "train", nb_to_use=1000
 710             )
 711             log_string(
 712                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 713             )
 714
 715             test_nb_total, test_nb_correct, count = self.compute_error(
 716                 model, "test", nb_to_use=1000
 717             )
 718             log_string(
 719                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 720             )
 721
 722             if count is not None:
 723                 proportion_optimal = count.diagonal().sum().float() / count.sum()
 724                 log_string(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
 725                 with open(
 726                     os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
 727                 ) as f:
 728                     for i in range(count.size(0)):
 729                         for j in range(count.size(1)):
 730                             eol = " " if j < count.size(1) - 1 else "\n"
 731                             f.write(f"{count[i,j]}{eol}")
 732
 733             input = self.test_input[:48]
 734             result = input.clone()
 735             ar_mask = result.new_zeros(result.size())
 736             ar_mask[:, self.height * self.width :] = 1
 737             result *= 1 - ar_mask
 738             masked_inplace_autoregression(
 739                 model, self.batch_size, result, ar_mask, device=self.device
 740             )
 741
 742             mazes, paths = self.seq2map(input)
 743             _, predicted_paths = self.seq2map(result)
 744
 745             filename = os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.png")
 746             maze.save_image(
 747                 filename,
 748                 mazes=mazes,
 749                 target_paths=paths,
 750                 predicted_paths=predicted_paths,
 751                 path_correct=maze.path_correctness(mazes, predicted_paths),
 752                 path_optimal=maze.path_optimality(paths, predicted_paths),
 753             )
 754             log_string(f"wrote {filename}")
 755
 756             model.train(t)
 757
 758
 759 ######################################################################
 760
 761
 762 import snake
 763
 764
 765 class TaskSnake(Task):
 766     def __init__(
 767         self,
 768         nb_train_samples,
 769         nb_test_samples,
 770         batch_size,
 771         height,
 772         width,
 773         nb_colors,
 774         length,
 775         prompt_length,
 776         device=torch.device("cpu"),
 777     ):
 778         self.batch_size = batch_size
 779         self.height = height
 780         self.width = width
 781         self.device = device
 782         self.prompt_length = prompt_length
 783
 784         self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
 785             nb_train_samples,
 786             height,
 787             width,
 788             nb_colors,
 789             length,
 790             prompt_length,
 791             self.device,
 792         )
 793         self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
 794             nb_test_samples,
 795             height,
 796             width,
 797             nb_colors,
 798             length,
 799             prompt_length,
 800             self.device,
 801         )
 802
 803         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 804
 805     def batches(self, split="train", nb_to_use=-1, desc=None):
 806         assert split in {"train", "test"}
 807         input = self.train_input if split == "train" else self.test_input
 808         if nb_to_use > 0:
 809             input = input[:nb_to_use]
 810         if desc is None:
 811             desc = f"epoch-{split}"
 812         for batch in tqdm.tqdm(
 813             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 814         ):
 815             yield batch
 816
 817     def vocabulary_size(self):
 818         return self.nb_codes
 819
 820     def produce_results(self, n_epoch, model):
 821         with torch.autograd.no_grad():
 822             t = model.training
 823             model.eval()
 824
 825             def compute_nb_correct(input, prior_visits):
 826                 result = input.clone()
 827                 i = torch.arange(result.size(1), device=result.device)[None, :]
 828                 ar_mask = (
 829                     torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
 830                     .long()
 831                     .expand_as(result)
 832                 )
 833                 result *= 1 - ar_mask
 834
 835                 # snake.solver(result,ar_mask)
 836
 837                 masked_inplace_autoregression(
 838                     model, self.batch_size, result, ar_mask, device=self.device
 839                 )
 840
 841                 nb_total = ((prior_visits > 0) * ar_mask).sum()
 842
 843                 nb_correct = (
 844                     (result == input).long() * (prior_visits > 0) * ar_mask
 845                 ).sum()
 846
 847                 # nb_total = result.size(0)
 848                 # nb_correct = ((result - input).abs().sum(1) == 0).sum()
 849
 850                 return nb_total, nb_correct
 851
 852             # train_nb_total, train_nb_correct = compute_nb_correct(
 853             # self.train_input, self.train_prior_visits
 854             # )
 855
 856             # log_string(
 857             # f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 858             # )
 859
 860             test_nb_total, test_nb_correct = compute_nb_correct(
 861                 self.test_input[:1000], self.test_prior_visits[:1000]
 862             )
 863
 864             log_string(
 865                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 866             )
 867
 868             model.train(t)
 869
 870
 871 ######################################################################
 872
 873
 874 import stack
 875
 876
 877 class TaskStack(Task):
 878     def __init__(
 879         self,
 880         nb_train_samples,
 881         nb_test_samples,
 882         batch_size,
 883         nb_steps,
 884         nb_stacks,
 885         nb_digits,
 886         fraction_values_for_train=None,
 887         device=torch.device("cpu"),
 888     ):
 889         self.batch_size = batch_size
 890         self.nb_steps = nb_steps
 891         self.nb_stacks = nb_stacks
 892         self.nb_digits = nb_digits
 893         self.device = device
 894
 895         if fraction_values_for_train is None:
 896             values_for_train = None
 897             values_for_test = None
 898         else:
 899             all = torch.randperm(10**nb_digits)
 900             nb_for_train = int(all.size(0) * fraction_values_for_train)
 901             values_for_train = all[:nb_for_train]
 902             values_for_test = all[nb_for_train:]
 903
 904         self.train_input, self.train_stack_counts = stack.generate_sequences(
 905             nb_train_samples,
 906             nb_steps,
 907             nb_stacks,
 908             nb_digits,
 909             values_for_train,
 910             self.device,
 911         )
 912
 913         self.test_input, self.test_stack_counts = stack.generate_sequences(
 914             nb_test_samples,
 915             nb_steps,
 916             nb_stacks,
 917             nb_digits,
 918             values_for_test,
 919             self.device,
 920         )
 921
 922         mask = self.test_input.clone()
 923         stack.remove_popped_values(mask, self.nb_stacks, self.nb_digits)
 924         mask = mask != self.test_input
 925         counts = self.test_stack_counts.flatten()[mask.flatten()]
 926         counts = F.one_hot(counts).sum(0)
 927         log_string(f"stack_count {counts}")
 928
 929         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 930
 931     def batches(self, split="train", nb_to_use=-1, desc=None):
 932         assert split in {"train", "test"}
 933         input = self.train_input if split == "train" else self.test_input
 934         if nb_to_use > 0:
 935             input = input[:nb_to_use]
 936         if desc is None:
 937             desc = f"epoch-{split}"
 938         for batch in tqdm.tqdm(
 939             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 940         ):
 941             yield batch
 942
 943     def vocabulary_size(self):
 944         return self.nb_codes
 945
 946     def produce_results(self, n_epoch, model):
 947         with torch.autograd.no_grad():
 948             t = model.training
 949             model.eval()
 950
 951             def compute_nb_correct(input):
 952                 result = input.clone()
 953                 stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 954                 ar_mask = (result != input).long()
 955                 masked_inplace_autoregression(
 956                     model, self.batch_size, result, ar_mask, device=self.device
 957                 )
 958
 959                 errors = ((result != input).long() * ar_mask).reshape(
 960                     -1, 1 + self.nb_digits
 961                 )
 962                 ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits)
 963
 964                 nb_total = ar_mask.max(1).values.sum()
 965                 nb_correct = nb_total - errors.max(1).values.sum()
 966
 967                 return nb_total, nb_correct
 968
 969             test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
 970
 971             log_string(
 972                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 973             )
 974
 975             #!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
 976             l = 50
 977             l = l - l % (1 + self.nb_digits)
 978             input = self.test_input[:10, :l]
 979             result = input.clone()
 980             stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
 981             ar_mask = (result != input).long()
 982             for n in range(result.size(0)):
 983                 log_string(
 984                     f"test_before {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 985                 )
 986             masked_inplace_autoregression(
 987                 model, self.batch_size, result, ar_mask, device=self.device
 988             )
 989             for n in range(result.size(0)):
 990                 log_string(
 991                     f"test_after  {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
 992                 )
 993             #!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
 994
 995             model.train(t)
 996
 997
 998 ######################################################################
 999
1000
1001 def picoclvr_pruner_horizontal_green(p):
1002     return not ("green" in p and ("left" in p or "right" in p))
1003
1004
1005 picoclvr_pruner_train = (
1006     picoclvr_pruner_horizontal_green
1007     if args.picocvlr_prune_properties in {"train+eval"}
1008     else None
1009 )
1010
1011 picoclvr_pruner_eval = (
1012     (lambda p: not picoclvr_pruner_horizontal_green(p))
1013     if args.picocvlr_prune_properties in {"train+eval", "eval"}
1014     else None
1015 )
1016
1017 ######################################################################
1018
1019 if args.task == "picoclvr":
1020     task = TaskPicoCLVR(
1021         nb_train_samples=args.nb_train_samples,
1022         nb_test_samples=args.nb_test_samples,
1023         batch_size=args.batch_size,
1024         height=args.picoclvr_height,
1025         width=args.picoclvr_width,
1026         nb_colors=args.picoclvr_nb_colors,
1027         device=device,
1028         pruner_train=picoclvr_pruner_train,
1029         pruner_eval=picoclvr_pruner_eval,
1030     )
1031
1032 elif args.task == "mnist":
1033     task = TaskMNIST(
1034         batch_size=args.batch_size,
1035         device=device,
1036     )
1037
1038 elif args.task == "maze":
1039     task = TaskMaze(
1040         nb_train_samples=args.nb_train_samples,
1041         nb_test_samples=args.nb_test_samples,
1042         batch_size=args.batch_size,
1043         height=args.maze_height,
1044         width=args.maze_width,
1045         nb_walls=args.maze_nb_walls,
1046         device=device,
1047     )
1048
1049 elif args.task == "snake":
1050     task = TaskSnake(
1051         nb_train_samples=args.nb_train_samples,
1052         nb_test_samples=args.nb_test_samples,
1053         batch_size=args.batch_size,
1054         height=args.snake_height,
1055         width=args.snake_width,
1056         nb_colors=args.snake_nb_colors,
1057         length=args.snake_length,
1058         prompt_length=args.snake_length // 2,
1059         device=device,
1060     )
1061
1062 elif args.task == "stack":
1063     task = TaskStack(
1064         nb_train_samples=args.nb_train_samples,
1065         nb_test_samples=args.nb_test_samples,
1066         batch_size=args.batch_size,
1067         nb_steps=args.stack_nb_steps,
1068         nb_stacks=args.stack_nb_stacks,
1069         nb_digits=args.stack_nb_digits,
1070         fraction_values_for_train=args.stack_fraction_values_for_train,
1071         device=device,
1072     )
1073
1074 else:
1075     raise ValueError(f"Unknown task {args.task}")
1076
1077 ######################################################################
1078
1079 log_string(f"device {device}")
1080
1081 vocabulary_size = task.vocabulary_size()
1082
1083 log_string(f"vocabulary_size {vocabulary_size}")
1084
1085 ##############################
1086
1087 model = mygpt.MyGPT(
1088     vocabulary_size=vocabulary_size,
1089     dim_model=args.dim_model,
1090     dim_keys=args.dim_keys,
1091     dim_hidden=args.dim_hidden,
1092     nb_heads=args.nb_heads,
1093     nb_blocks=args.nb_blocks,
1094     causal=True,
1095     dropout=args.dropout,
1096 )
1097
1098 model.to(device)
1099
1100 nb_parameters = sum(p.numel() for p in model.parameters())
1101 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
1102
1103 ######################################################################
1104
1105 nb_epochs_finished = 0
1106
1107 if args.no_checkpoint:
1108     log_string(f"not trying to load checkpoint.")
1109
1110 else:
1111     try:
1112         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
1113         checkpoint = torch.load(checkpoint_name)
1114         nb_epochs_finished = checkpoint["nb_epochs_finished"]
1115         model.load_state_dict(checkpoint["model_state"])
1116         torch.set_rng_state(checkpoint["rng_state"])
1117         if torch.cuda.is_available():
1118             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
1119
1120         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
1121
1122     except FileNotFoundError:
1123         log_string("starting from scratch.")
1124
1125     except:
1126         log_string("error when loading the checkpoint.")
1127         exit(1)
1128
1129 ######################################################################
1130
1131 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
1132
1133 token_count = 0
1134 for input in task.batches(split="train"):
1135     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
1136 token_probas = token_count / token_count.sum()
1137 entropy = -torch.xlogy(token_probas, token_probas).sum()
1138 train_set_perplexity = math.exp(entropy)
1139
1140 ##############################
1141
1142 if args.learning_rate_schedule == "cos":
1143     learning_rate_schedule = {}
1144     for n_epoch in range(args.nb_epochs):
1145         u = n_epoch / args.nb_epochs * math.pi
1146         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
1147 else:
1148     u = {
1149         int(k): float(v)
1150         for k, v in [
1151             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
1152         ]
1153     }
1154
1155     learning_rate_schedule = {}
1156     learning_rate = args.learning_rate
1157     for n_epoch in range(args.nb_epochs):
1158         if n_epoch in u:
1159             learning_rate = u[n_epoch]
1160         learning_rate_schedule[n_epoch] = learning_rate
1161
1162 log_string(f"learning_rate_schedule {learning_rate_schedule}")
1163
1164 ##############################
1165
1166 nb_samples_seen = 0
1167
1168 if nb_epochs_finished >= nb_epochs:
1169     task.produce_results(nb_epochs_finished, model)
1170
1171 for n_epoch in range(nb_epochs_finished, nb_epochs):
1172     learning_rate = learning_rate_schedule[n_epoch]
1173
1174     log_string(f"learning_rate {learning_rate}")
1175
1176     if args.optim == "sgd":
1177         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
1178     elif args.optim == "adam":
1179         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
1180     elif args.optim == "adamw":
1181         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
1182     else:
1183         raise ValueError(f"Unknown optimizer {args.optim}.")
1184
1185     model.train()
1186
1187     nb_train_samples, acc_train_loss = 0, 0.0
1188
1189     for input in task.batches(split="train"):
1190         input = input.to(device)
1191         output = model(mygpt.BracketedSequence(input)).x
1192         loss = F.cross_entropy(output.transpose(1, 2), input)
1193         acc_train_loss += loss.item() * input.size(0)
1194         nb_train_samples += input.size(0)
1195         nb_samples_seen += input.size(0)
1196
1197         optimizer.zero_grad()
1198         loss.backward()
1199         optimizer.step()
1200
1201     with torch.autograd.no_grad():
1202         model.eval()
1203
1204         nb_test_samples, acc_test_loss = 0, 0.0
1205
1206         for input in task.batches(split="test"):
1207             input = input.to(device)
1208
1209             output = model(mygpt.BracketedSequence(input)).x
1210             loss = F.cross_entropy(output.transpose(1, 2), input)
1211             acc_test_loss += loss.item() * input.size(0)
1212             nb_test_samples += input.size(0)
1213
1214         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
1215         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
1216
1217         log_string(
1218             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
1219         )
1220
1221         task.produce_results(n_epoch, model)
1222
1223     checkpoint = {
1224         "nb_epochs_finished": n_epoch + 1,
1225         "model_state": model.state_dict(),
1226         "rng_state": torch.get_rng_state(),
1227     }
1228
1229     if torch.cuda.is_available():
1230         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
1231
1232     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
1233     torch.save(checkpoint, checkpoint_name)
1234     log_string(f"saved checkpoint {checkpoint_name}")
1235
1236 ######################################################################