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