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="results_default")
  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=250000)
  49
  50 parser.add_argument("--nb_test_samples", type=int, default=10000)
  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_values", type=int, default=10)
 117
 118 ######################################################################
 119
 120 args = parser.parse_args()
 121
 122 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 123
 124 try:
 125     os.mkdir(args.result_dir)
 126 except FileExistsError:
 127     if not args.overwrite_results:
 128         print(f"result directory {args.result_dir} already exists")
 129         exit(1)
 130
 131 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 132
 133 if args.seed >= 0:
 134     # torch.backends.cudnn.deterministic = True
 135     # torch.backends.cudnn.benchmark = False
 136     # torch.use_deterministic_algorithms(True)
 137     torch.manual_seed(args.seed)
 138     if torch.cuda.is_available():
 139         torch.cuda.manual_seed_all(args.seed)
 140
 141 ######################################################################
 142
 143 default_args = {
 144     "picoclvr": {
 145         "nb_epochs": 25,
 146         "batch_size": 25,
 147         "nb_train_samples": 250000,
 148         "nb_test_samples": 10000,
 149     },
 150     "mnist": {
 151         "nb_epochs": 25,
 152         "batch_size": 10,
 153         "nb_train_samples": 250000,
 154         "nb_test_samples": 10000,
 155     },
 156     "maze": {
 157         "nb_epochs": 25,
 158         "batch_size": 25,
 159         "nb_train_samples": 250000,
 160         "nb_test_samples": 10000,
 161     },
 162     "snake": {
 163         "nb_epochs": 5,
 164         "batch_size": 25,
 165         "nb_train_samples": 250000,
 166         "nb_test_samples": 10000,
 167     },
 168     "stack": {
 169         "nb_epochs": 5,
 170         "batch_size": 25,
 171         "nb_train_samples": 100000,
 172         "nb_test_samples": 1000,
 173     },
 174 }
 175
 176 if args.task in default_args:
 177     for k, v in default_args[args.task].items():
 178         if getattr(args, k) is None:
 179             setattr(args, k, v)
 180
 181 ######################################################################
 182
 183
 184 def log_string(s):
 185     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 186
 187     if log_file is not None:
 188         log_file.write(t + s + "\n")
 189         log_file.flush()
 190
 191     print(t + s)
 192     sys.stdout.flush()
 193
 194
 195 for n in vars(args):
 196     log_string(f"args.{n} {getattr(args, n)}")
 197
 198 ######################################################################
 199
 200
 201 # ra_mask is boolean, with 1s on the values to generate
 202
 203
 204 def masked_inplace_autoregression(
 205     model,
 206     batch_size,
 207     input,
 208     ar_mask,
 209     forbidden_tokens=None,
 210     progress_bar_desc="autoregression",
 211     device=torch.device("cpu"),
 212 ):
 213     # p = logits.softmax(1)
 214     # entropy[:,s]= p.xlogy(p).sum(1) / math.log(2)
 215     batches = zip(input.split(batch_size), ar_mask.split(batch_size))
 216     if progress_bar_desc is not None:
 217         tqdm.tqdm(
 218             batches,
 219             dynamic_ncols=True,
 220             desc=progress_bar_desc,
 221             total=input.size(0) // batch_size,
 222         )
 223     for input, ar_mask in batches:
 224         i = (ar_mask.sum(0) > 0).nonzero()
 225         if i.min() > 0:
 226             model(
 227                 mygpt.BracketedSequence(input, 0, i.min())
 228             )  # Needed to initialize the model's cache
 229         for s in range(i.min(), i.max() + 1):
 230             output = model(mygpt.BracketedSequence(input, s, 1)).x
 231             logits = output[:, s]
 232             if forbidden_tokens is not None:
 233                 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
 234             if args.deterministic_synthesis:
 235                 t_next = logits.argmax(1)
 236             else:
 237                 dist = torch.distributions.categorical.Categorical(logits=logits)
 238                 t_next = dist.sample()
 239             input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 240
 241
 242 ######################################################################
 243
 244
 245 class Task:
 246     def batches(self, split="train"):
 247         pass
 248
 249     def vocabulary_size(self):
 250         pass
 251
 252     def produce_results(self, n_epoch, model):
 253         pass
 254
 255
 256 ######################################################################
 257
 258 import picoclvr
 259
 260
 261 class TaskPicoCLVR(Task):
 262     # Make a tensor from a list of strings
 263     def tensorize(self, descr):
 264         token_descr = [s.strip().split(" ") for s in descr]
 265         l = max([len(s) for s in token_descr])
 266         token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
 267         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
 268         return torch.tensor(id_descr, device=self.device)
 269
 270     # Make a list of strings from a tensor
 271     def detensorize(self, x):
 272         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
 273
 274     # trim all the tensors in the tuple z to remove as much token from
 275     # left and right in the first tensor. If z is a tuple, all its
 276     # elements are trimed according to the triming for the first
 277     def trim(self, z, token="<nul>"):
 278         n = self.token2id[token]
 279         if type(z) == tuple:
 280             x = z[0]
 281             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 282             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 283             return tuple([t[:, a:b] for t in z])
 284         else:
 285             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 286             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 287             return z[:, a:b]
 288
 289     ######################
 290     # Not the cleanest part of the code
 291
 292     # Extract the last image of each sequence, from the last <img>
 293     # included, and set to <nul> all the tokens from the beginning of
 294     # that image to the end
 295     def excise_last_image(self, input):
 296         t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 297         nb_img_tokens = self.height * self.width + 1
 298
 299         input = input.clone()
 300         t = (input == t_img).long()
 301         tail_masks = (t.cumsum(dim=1) == t.sum(dim=1, keepdim=True)).long()
 302         i = (t * tail_masks).nonzero(as_tuple=True)
 303         j = (
 304             i[0][:, None],
 305             i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
 306         )
 307         images = self.trim(input[j])
 308         input[j] = t_nul
 309         loss_masks = 1 - tail_masks
 310         input, loss_masks = self.trim((input, loss_masks))
 311         return input, loss_masks, images
 312
 313     def add_true_image(self, input, images, loss_masks):
 314         t_nul = self.token2id["<nul>"]
 315         nb_img_tokens = self.height * self.width + 1
 316         input = F.pad(input, (0, nb_img_tokens), value=t_nul)
 317         loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
 318         t = (input == t_nul).long()
 319         i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
 320         j = (
 321             i[0][:, None],
 322             i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
 323         )
 324         input[j] = images
 325         loss_masks[j] = 1
 326         input, loss_masks = self.trim((input, loss_masks))
 327         return input, loss_masks
 328
 329     def add_generated_image(self, input, loss_masks, model):
 330         t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 331         nb_img_tokens = self.height * self.width + 1
 332
 333         input = F.pad(input, (0, nb_img_tokens), value=t_nul)
 334         loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
 335         t = (input == t_nul).long()
 336         i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
 337         input[i] = t_img
 338
 339         j = (
 340             i[0][:, None],
 341             i[1][:, None]
 342             + 1
 343             + torch.arange(nb_img_tokens - 1, device=input.device)[None, :],
 344         )
 345         ar_masks = input.new_zeros(input.size(), dtype=torch.int64)
 346         ar_masks[j] = 1
 347         forbidden_tokens = (
 348             torch.arange(self.vocabulary_size(), device=input.device) == t_nul
 349         )
 350         with torch.autograd.no_grad():
 351             t = model.training
 352             model.eval()
 353             masked_inplace_autoregression(
 354                 model,
 355                 self.batch_size,
 356                 input,
 357                 ar_masks,
 358                 forbidden_tokens,
 359                 progress_bar_desc=None,
 360                 device=self.device,
 361             )
 362             model.train(t)
 363
 364         input, loss_masks = self.trim((input, loss_masks))
 365
 366         return input, loss_masks
 367
 368     ######################
 369
 370     def __init__(
 371         self,
 372         nb_train_samples,
 373         nb_test_samples,
 374         batch_size,
 375         height,
 376         width,
 377         nb_colors=5,
 378         device=torch.device("cpu"),
 379         pruner_train=None,
 380         pruner_eval=None,
 381     ):
 382         def generate_descr(nb, cache_suffix, pruner):
 383             return picoclvr.generate(
 384                 nb,
 385                 height=self.height,
 386                 width=self.width,
 387                 nb_colors=nb_colors,
 388                 pruner=pruner,
 389             )
 390
 391         self.height = height
 392         self.width = width
 393         self.batch_size = batch_size
 394         self.device = device
 395         self.pruner_train = pruner_train
 396         self.pruner_eval = pruner_eval
 397
 398         param = {
 399             "nb_train_samples": nb_train_samples,
 400             "nb_test_samples": nb_test_samples,
 401             "height": height,
 402             "width": width,
 403             "nb_colors": nb_colors,
 404             "batch_size": batch_size,
 405             "rng_state": list(torch.get_rng_state()),
 406         }
 407
 408         log_string(
 409             f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
 410         )
 411         self.train_descr = generate_descr(
 412             nb_train_samples, "train", pruner=self.pruner_train
 413         )
 414         self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
 415
 416         # Build the tokenizer
 417         tokens = {"<nul>", "<img>"}
 418         for d in [self.train_descr, self.test_descr]:
 419             for s in d:
 420                 for t in s.strip().split(" "):
 421                     tokens.add(t)
 422         # make this set a sorted list to get the same tensors given
 423         # the same descr
 424         tokens = list(tokens)
 425         tokens.sort()
 426         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
 427         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
 428
 429         # Tokenize the train and test sets
 430         self.train_input = self.tensorize(self.train_descr)
 431         self.test_input = self.tensorize(self.test_descr)
 432
 433     def batches(self, split="train"):
 434         assert split in {"train", "test"}
 435         input = self.train_input if split == "train" else self.test_input
 436         for batch in tqdm.tqdm(
 437             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 438         ):
 439             yield self.trim(batch)
 440
 441     def vocabulary_size(self):
 442         return len(self.token2id)
 443
 444     def compute_missing_properties(self, n_epoch, model, pruner=None):
 445         acc_nb_requested_properties = []
 446         acc_nb_missing_properties = []
 447         acc_nb_results = 0
 448
 449         for input in tqdm.tqdm(
 450             self.test_input.split(self.batch_size),
 451             dynamic_ncols=True,
 452             desc=f"test-properties",
 453         ):
 454             tape, loss_masks, _ = self.excise_last_image(input)
 455             tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
 456             result_descr = self.detensorize(tape)
 457             np = picoclvr.nb_properties(
 458                 result_descr,
 459                 height=self.height,
 460                 width=self.width,
 461                 pruner=pruner,
 462             )
 463             nb_requested_properties, _, nb_missing_properties = zip(*np)
 464             acc_nb_requested_properties += nb_requested_properties
 465             acc_nb_missing_properties += nb_missing_properties
 466             acc_nb_results += len(result_descr)
 467
 468         nb_requested_properties = sum(acc_nb_requested_properties)
 469         nb_missing_properties = sum(acc_nb_missing_properties)
 470
 471         prefix = "" if pruner is None else "pruned_"
 472         log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 473         log_string(
 474             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 475         )
 476         log_string(
 477             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 478         )
 479
 480     ######################################################################
 481
 482     def produce_results(self, n_epoch, model):
 483         self.compute_missing_properties(n_epoch, model)
 484
 485         if self.pruner_eval is not None:
 486             self.compute_missing_properties(n_epoch, model, self.pruner_eval)
 487
 488         nb_tokens_to_generate = self.height * self.width + 3
 489         result_descr = []
 490         nb_per_primer = 8
 491         primer = []
 492
 493         for primer_descr in [
 494             "red above green <sep> green top <sep> blue right of red",
 495             "there is red <sep> there is yellow <sep> there is blue",
 496             "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
 497             "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
 498         ]:
 499             primer += [primer_descr] * nb_per_primer
 500
 501         tape = self.tensorize(primer)
 502         loss_masks = 1 - (tape == self.token2id["<nul>"]).long()
 503         tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
 504         result_descr = self.detensorize(tape)
 505
 506         np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
 507
 508         acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
 509         acc_nb_results = len(result_descr)
 510
 511         nb_requested_properties = sum(acc_nb_requested_properties)
 512         nb_missing_properties = sum(acc_nb_missing_properties)
 513
 514         prefix = "demo_"
 515         log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 516         log_string(
 517             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 518         )
 519         log_string(
 520             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 521         )
 522
 523         img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
 524
 525         if img.dim() == 5:
 526             if img.size(1) == 1:
 527                 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
 528             else:
 529                 img = torch.cat(
 530                     [
 531                         torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
 532                         for x in img
 533                     ],
 534                     0,
 535                 )
 536
 537         image_name = os.path.join(args.result_dir, f"picoclvr_result_{n_epoch:04d}.png")
 538         torchvision.utils.save_image(
 539             img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0
 540         )
 541         log_string(f"wrote {image_name}")
 542
 543
 544 ######################################################################
 545
 546
 547 class TaskMNIST(Task):
 548     def __init__(self, batch_size, device=torch.device("cpu")):
 549         self.device = device
 550         self.batch_size = batch_size
 551
 552     def batches(self, split="train"):
 553         assert split in {"train", "test"}
 554         data_set = torchvision.datasets.MNIST(
 555             root="./data", train=(split == "train"), download=True
 556         )
 557         data_input = data_set.data.view(-1, 28 * 28).long()
 558         if args.nb_train_samples is not None:
 559             data_input = data_input[: args.nb_train_samples]
 560         for batch in tqdm.tqdm(
 561             data_input.split(self.batch_size), desc=f"epoch-{split}"
 562         ):
 563             yield batch
 564
 565     def vocabulary_size(self):
 566         return 256
 567
 568     def produce_results(self, n_epoch, model):
 569         results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
 570         ar_mask = torch.full_like(results, 1)
 571         masked_inplace_autoregression(
 572             model, self.batch_size, results, ar_mask, device=self.device
 573         )
 574         image_name = os.path.join(args.result_dir, f"mnist_result_{n_epoch:04d}.png")
 575         torchvision.utils.save_image(
 576             1 - results.reshape(-1, 1, 28, 28) / 255.0,
 577             image_name,
 578             nrow=16,
 579             pad_value=0.8,
 580         )
 581         log_string(f"wrote {image_name}")
 582
 583
 584 ######################################################################
 585
 586 import maze
 587
 588
 589 class TaskMaze(Task):
 590     def map2seq(self, *m):
 591         return torch.cat([x.flatten(1) for x in m], 1)
 592
 593     def seq2map(self, s):
 594         s = s.reshape(s.size(0), -1, self.height, self.width)
 595         return (s[:, k] for k in range(s.size(1)))
 596
 597     def __init__(
 598         self,
 599         nb_train_samples,
 600         nb_test_samples,
 601         batch_size,
 602         height,
 603         width,
 604         nb_walls,
 605         device=torch.device("cpu"),
 606     ):
 607         self.batch_size = batch_size
 608         self.height = height
 609         self.width = width
 610         self.device = device
 611
 612         train_mazes, train_paths, _ = maze.create_maze_data(
 613             nb_train_samples,
 614             height=height,
 615             width=width,
 616             nb_walls=nb_walls,
 617             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 618         )
 619         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 620
 621         test_mazes, test_paths, _ = maze.create_maze_data(
 622             nb_test_samples,
 623             height=height,
 624             width=width,
 625             nb_walls=nb_walls,
 626             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 627         )
 628         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 629
 630         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 631
 632     def batches(self, split="train", nb_to_use=-1, desc=None):
 633         assert split in {"train", "test"}
 634         input = self.train_input if split == "train" else self.test_input
 635         if nb_to_use > 0:
 636             input = input[:nb_to_use]
 637         if desc is None:
 638             desc = f"epoch-{split}"
 639         for batch in tqdm.tqdm(
 640             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 641         ):
 642             yield batch
 643
 644     def vocabulary_size(self):
 645         return self.nb_codes
 646
 647     def compute_error(self, model, split="train", nb_to_use=-1):
 648         nb_total, nb_correct = 0, 0
 649         count = torch.zeros(
 650             self.width * self.height,
 651             self.width * self.height,
 652             device=self.device,
 653             dtype=torch.int64,
 654         )
 655         for input in tqdm.tqdm(
 656             task.batches(split, nb_to_use),
 657             dynamic_ncols=True,
 658             desc=f"test-mazes",
 659         ):
 660             result = input.clone()
 661             ar_mask = result.new_zeros(result.size())
 662             ar_mask[:, self.height * self.width :] = 1
 663             result *= 1 - ar_mask
 664             masked_inplace_autoregression(
 665                 model,
 666                 self.batch_size,
 667                 result,
 668                 ar_mask,
 669                 progress_bar_desc=None,
 670                 device=self.device,
 671             )
 672             mazes, paths = self.seq2map(result)
 673             path_correctness = maze.path_correctness(mazes, paths)
 674             nb_correct += path_correctness.long().sum()
 675             nb_total += mazes.size(0)
 676
 677             optimal_path_lengths = (
 678                 (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
 679             )
 680             predicted_path_lengths = (
 681                 (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
 682             )
 683             optimal_path_lengths = optimal_path_lengths[path_correctness]
 684             predicted_path_lengths = predicted_path_lengths[path_correctness]
 685             count[optimal_path_lengths, predicted_path_lengths] += 1
 686
 687         if count.max() == 0:
 688             count = None
 689         else:
 690             count = count[
 691                 : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
 692             ]
 693
 694         return nb_total, nb_correct, count
 695
 696     def produce_results(self, n_epoch, model):
 697         with torch.autograd.no_grad():
 698             t = model.training
 699             model.eval()
 700
 701             train_nb_total, train_nb_correct, count = self.compute_error(
 702                 model, "train", nb_to_use=1000
 703             )
 704             log_string(
 705                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 706             )
 707
 708             test_nb_total, test_nb_correct, count = self.compute_error(
 709                 model, "test", nb_to_use=1000
 710             )
 711             log_string(
 712                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 713             )
 714
 715             if count is not None:
 716                 proportion_optimal = count.diagonal().sum().float() / count.sum()
 717                 log_string(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
 718                 with open(
 719                     os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
 720                 ) as f:
 721                     for i in range(count.size(0)):
 722                         for j in range(count.size(1)):
 723                             eol = " " if j < count.size(1) - 1 else "\n"
 724                             f.write(f"{count[i,j]}{eol}")
 725
 726             input = self.test_input[:48]
 727             result = input.clone()
 728             ar_mask = result.new_zeros(result.size())
 729             ar_mask[:, self.height * self.width :] = 1
 730             result *= 1 - ar_mask
 731             masked_inplace_autoregression(
 732                 model, self.batch_size, result, ar_mask, device=self.device
 733             )
 734
 735             mazes, paths = self.seq2map(input)
 736             _, predicted_paths = self.seq2map(result)
 737
 738             filename = os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.png")
 739             maze.save_image(
 740                 filename,
 741                 mazes=mazes,
 742                 target_paths=paths,
 743                 predicted_paths=predicted_paths,
 744                 path_correct=maze.path_correctness(mazes, predicted_paths),
 745                 path_optimal=maze.path_optimality(paths, predicted_paths),
 746             )
 747             log_string(f"wrote {filename}")
 748
 749             model.train(t)
 750
 751
 752 ######################################################################
 753
 754
 755 import snake
 756
 757
 758 class TaskSnake(Task):
 759     def __init__(
 760         self,
 761         nb_train_samples,
 762         nb_test_samples,
 763         batch_size,
 764         height,
 765         width,
 766         nb_colors,
 767         length,
 768         prompt_length,
 769         device=torch.device("cpu"),
 770     ):
 771         self.batch_size = batch_size
 772         self.height = height
 773         self.width = width
 774         self.device = device
 775         self.prompt_length = prompt_length
 776
 777         self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
 778             nb_train_samples,
 779             height,
 780             width,
 781             nb_colors,
 782             length,
 783             prompt_length,
 784             self.device,
 785         )
 786         self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
 787             nb_test_samples,
 788             height,
 789             width,
 790             nb_colors,
 791             length,
 792             prompt_length,
 793             self.device,
 794         )
 795
 796         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 797
 798     def batches(self, split="train", nb_to_use=-1, desc=None):
 799         assert split in {"train", "test"}
 800         input = self.train_input if split == "train" else self.test_input
 801         if nb_to_use > 0:
 802             input = input[:nb_to_use]
 803         if desc is None:
 804             desc = f"epoch-{split}"
 805         for batch in tqdm.tqdm(
 806             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 807         ):
 808             yield batch
 809
 810     def vocabulary_size(self):
 811         return self.nb_codes
 812
 813     def produce_results(self, n_epoch, model):
 814         with torch.autograd.no_grad():
 815             t = model.training
 816             model.eval()
 817
 818             def compute_nb_correct(input, prior_visits):
 819                 result = input.clone()
 820                 i = torch.arange(result.size(1), device=result.device)[None, :]
 821                 ar_mask = (
 822                     torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
 823                     .long()
 824                     .expand_as(result)
 825                 )
 826                 result *= 1 - ar_mask
 827
 828                 # snake.solver(result,ar_mask)
 829
 830                 masked_inplace_autoregression(
 831                     model, self.batch_size, result, ar_mask, device=self.device
 832                 )
 833
 834                 nb_total = ((prior_visits > 0) * ar_mask).sum()
 835
 836                 nb_correct = (
 837                     (result == input).long() * (prior_visits > 0) * ar_mask
 838                 ).sum()
 839
 840                 # nb_total = result.size(0)
 841                 # nb_correct = ((result - input).abs().sum(1) == 0).sum()
 842
 843                 return nb_total, nb_correct
 844
 845             # train_nb_total, train_nb_correct = compute_nb_correct(
 846             # self.train_input, self.train_prior_visits
 847             # )
 848
 849             # log_string(
 850             # f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 851             # )
 852
 853             test_nb_total, test_nb_correct = compute_nb_correct(
 854                 self.test_input[:1000], self.test_prior_visits[:1000]
 855             )
 856
 857             log_string(
 858                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 859             )
 860
 861             model.train(t)
 862
 863
 864 ######################################################################
 865
 866
 867 import stack
 868
 869
 870 class TaskStack(Task):
 871     def __init__(
 872         self,
 873         nb_train_samples,
 874         nb_test_samples,
 875         batch_size,
 876         nb_steps,
 877         nb_stacks,
 878         nb_values,
 879         device=torch.device("cpu"),
 880     ):
 881         self.batch_size = batch_size
 882         self.nb_steps = nb_steps
 883         self.nb_stacks = nb_stacks
 884         self.nb_values = nb_values
 885         self.device = device
 886
 887         self.train_input, self.train_stack_counts = stack.generate_sequences(
 888             nb_train_samples, nb_steps, nb_stacks, nb_values, self.device
 889         )
 890
 891         self.test_input, self.test_stack_counts = stack.generate_sequences(
 892             nb_test_samples, nb_steps, nb_stacks, nb_values, self.device
 893         )
 894
 895         mask = self.test_input.clone()
 896         stack.remove_poped_values(mask,self.nb_stacks)
 897         mask=(mask!=self.test_input)
 898         counts = self.test_stack_counts.flatten()[mask.flatten()]
 899         counts=F.one_hot(counts).sum(0)
 900         log_string(f"stack_count {counts}")
 901
 902         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 903
 904     def batches(self, split="train", nb_to_use=-1, desc=None):
 905         assert split in {"train", "test"}
 906         input = self.train_input if split == "train" else self.test_input
 907         if nb_to_use > 0:
 908             input = input[:nb_to_use]
 909         if desc is None:
 910             desc = f"epoch-{split}"
 911         for batch in tqdm.tqdm(
 912             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 913         ):
 914             yield batch
 915
 916     def vocabulary_size(self):
 917         return self.nb_codes
 918
 919     def produce_results(self, n_epoch, model):
 920         with torch.autograd.no_grad():
 921             t = model.training
 922             model.eval()
 923
 924             def compute_nb_correct(input):
 925                 result = input.clone()
 926                 stack.remove_poped_values(result,self.nb_stacks)
 927                 ar_mask = (result != input).long()
 928                 result *= 1 - ar_mask
 929
 930                 masked_inplace_autoregression(
 931                     model, self.batch_size, result, ar_mask, device=self.device
 932                 )
 933
 934                 nb_total = ar_mask.sum()
 935
 936                 nb_correct = (
 937                     (result == input).long() * ar_mask
 938                 ).sum()
 939
 940                 return nb_total, nb_correct
 941
 942             test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
 943
 944             log_string(
 945                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 946             )
 947
 948             model.train(t)
 949
 950
 951 ######################################################################
 952
 953
 954 def picoclvr_pruner_horizontal_green(p):
 955     return not ("green" in p and ("left" in p or "right" in p))
 956
 957
 958 picoclvr_pruner_train = (
 959     picoclvr_pruner_horizontal_green
 960     if args.picocvlr_prune_properties in {"train+eval"}
 961     else None
 962 )
 963
 964 picoclvr_pruner_eval = (
 965     (lambda p: not picoclvr_pruner_horizontal_green(p))
 966     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 967     else None
 968 )
 969
 970 ######################################################################
 971
 972 if args.task == "picoclvr":
 973     task = TaskPicoCLVR(
 974         nb_train_samples=args.nb_train_samples,
 975         nb_test_samples=args.nb_test_samples,
 976         batch_size=args.batch_size,
 977         height=args.picoclvr_height,
 978         width=args.picoclvr_width,
 979         nb_colors=args.picoclvr_nb_colors,
 980         device=device,
 981         pruner_train=picoclvr_pruner_train,
 982         pruner_eval=picoclvr_pruner_eval,
 983     )
 984
 985 elif args.task == "mnist":
 986     task = TaskMNIST(
 987         batch_size=args.batch_size,
 988         device=device,
 989     )
 990
 991 elif args.task == "maze":
 992     task = TaskMaze(
 993         nb_train_samples=args.nb_train_samples,
 994         nb_test_samples=args.nb_test_samples,
 995         batch_size=args.batch_size,
 996         height=args.maze_height,
 997         width=args.maze_width,
 998         nb_walls=args.maze_nb_walls,
 999         device=device,
1000     )
1001
1002 elif args.task == "snake":
1003     task = TaskSnake(
1004         nb_train_samples=args.nb_train_samples,
1005         nb_test_samples=args.nb_test_samples,
1006         batch_size=args.batch_size,
1007         height=args.snake_height,
1008         width=args.snake_width,
1009         nb_colors=args.snake_nb_colors,
1010         length=args.snake_length,
1011         prompt_length=args.snake_length // 2,
1012         device=device,
1013     )
1014
1015 elif args.task == "stack":
1016     task = TaskStack(
1017         nb_train_samples=args.nb_train_samples,
1018         nb_test_samples=args.nb_test_samples,
1019         batch_size=args.batch_size,
1020         nb_steps = args.stack_nb_steps,
1021         nb_stacks = args.stack_nb_stacks,
1022         nb_values = args.stack_nb_values,
1023         device=device,
1024     )
1025
1026 else:
1027     raise ValueError(f"Unknown task {args.task}")
1028
1029 ######################################################################
1030
1031 log_string(f"device {device}")
1032
1033 vocabulary_size = task.vocabulary_size()
1034
1035 log_string(f"vocabulary_size {vocabulary_size}")
1036
1037 ##############################
1038
1039 model = mygpt.MyGPT(
1040     vocabulary_size=vocabulary_size,
1041     dim_model=args.dim_model,
1042     dim_keys=args.dim_keys,
1043     dim_hidden=args.dim_hidden,
1044     nb_heads=args.nb_heads,
1045     nb_blocks=args.nb_blocks,
1046     causal=True,
1047     dropout=args.dropout,
1048 )
1049
1050 model.to(device)
1051
1052 nb_parameters = sum(p.numel() for p in model.parameters())
1053 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
1054
1055 ######################################################################
1056
1057 nb_epochs_finished = 0
1058
1059 if args.no_checkpoint:
1060     log_string(f"not trying to load checkpoint.")
1061
1062 else:
1063     try:
1064         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
1065         checkpoint = torch.load(checkpoint_name)
1066         nb_epochs_finished = checkpoint["nb_epochs_finished"]
1067         model.load_state_dict(checkpoint["model_state"])
1068         torch.set_rng_state(checkpoint["rng_state"])
1069         if torch.cuda.is_available():
1070             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
1071
1072         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
1073
1074     except FileNotFoundError:
1075         log_string("starting from scratch.")
1076
1077     except:
1078         log_string("error when loading the checkpoint.")
1079         exit(1)
1080
1081 ######################################################################
1082
1083 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
1084
1085 token_count = 0
1086 for input in task.batches(split="train"):
1087     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
1088 token_probas = token_count / token_count.sum()
1089 entropy = -torch.xlogy(token_probas, token_probas).sum()
1090 train_set_perplexity = math.exp(entropy)
1091
1092 ##############################
1093
1094 if args.learning_rate_schedule == "cos":
1095     learning_rate_schedule = {}
1096     for n_epoch in range(args.nb_epochs):
1097         u = n_epoch / args.nb_epochs * math.pi
1098         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
1099 else:
1100     u = {
1101         int(k): float(v)
1102         for k, v in [
1103             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
1104         ]
1105     }
1106
1107     learning_rate_schedule = {}
1108     learning_rate = args.learning_rate
1109     for n_epoch in range(args.nb_epochs):
1110         if n_epoch in u:
1111             learning_rate = u[n_epoch]
1112         learning_rate_schedule[n_epoch] = learning_rate
1113
1114 log_string(f"learning_rate_schedule {learning_rate_schedule}")
1115
1116 ##############################
1117
1118 nb_samples_seen = 0
1119
1120 if nb_epochs_finished >= nb_epochs:
1121     task.produce_results(nb_epochs_finished, model)
1122
1123 for n_epoch in range(nb_epochs_finished, nb_epochs):
1124     learning_rate = learning_rate_schedule[n_epoch]
1125
1126     log_string(f"learning_rate {learning_rate}")
1127
1128     if args.optim == "sgd":
1129         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
1130     elif args.optim == "adam":
1131         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
1132     elif args.optim == "adamw":
1133         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
1134     else:
1135         raise ValueError(f"Unknown optimizer {args.optim}.")
1136
1137     model.train()
1138
1139     nb_train_samples, acc_train_loss = 0, 0.0
1140
1141     for input in task.batches(split="train"):
1142         input = input.to(device)
1143         output = model(mygpt.BracketedSequence(input)).x
1144         loss = F.cross_entropy(output.transpose(1, 2), input)
1145         acc_train_loss += loss.item() * input.size(0)
1146         nb_train_samples += input.size(0)
1147         nb_samples_seen += input.size(0)
1148
1149         optimizer.zero_grad()
1150         loss.backward()
1151         optimizer.step()
1152
1153     with torch.autograd.no_grad():
1154         model.eval()
1155
1156         nb_test_samples, acc_test_loss = 0, 0.0
1157
1158         for input in task.batches(split="test"):
1159             input = input.to(device)
1160
1161             output = model(mygpt.BracketedSequence(input)).x
1162             loss = F.cross_entropy(output.transpose(1, 2), input)
1163             acc_test_loss += loss.item() * input.size(0)
1164             nb_test_samples += input.size(0)
1165
1166         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
1167         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
1168
1169         log_string(
1170             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
1171         )
1172
1173         task.produce_results(n_epoch, model)
1174
1175     checkpoint = {
1176         "nb_epochs_finished": n_epoch + 1,
1177         "model_state": model.state_dict(),
1178         "rng_state": torch.get_rng_state(),
1179     }
1180
1181     if torch.cuda.is_available():
1182         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
1183
1184     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
1185     torch.save(checkpoint, checkpoint_name)
1186     log_string(f"saved checkpoint {checkpoint_name}")
1187
1188 ######################################################################