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, itertools, 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 to solve a toy geometric reasoning task."
  31 )
  32
  33 parser.add_argument("--task", type=str, default="picoclvr")
  34
  35 parser.add_argument("--log_filename", type=str, default="train.log")
  36
  37 parser.add_argument("--result_dir", type=str, default="results_default")
  38
  39 parser.add_argument("--seed", type=int, default=0)
  40
  41 parser.add_argument("--nb_epochs", type=int, default=25)
  42
  43 parser.add_argument("--batch_size", type=int, default=25)
  44
  45 parser.add_argument("--nb_train_samples", type=int, default=250000)
  46
  47 parser.add_argument("--nb_test_samples", type=int, default=10000)
  48
  49 parser.add_argument("--optim", type=str, default="adam")
  50
  51 parser.add_argument("--learning_rate", type=float, default=1e-4)
  52
  53 parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
  54
  55 parser.add_argument("--dim_model", type=int, default=512)
  56
  57 parser.add_argument("--dim_keys", type=int, default=64)
  58
  59 parser.add_argument("--dim_hidden", type=int, default=2048)
  60
  61 parser.add_argument("--nb_heads", type=int, default=8)
  62
  63 parser.add_argument("--nb_blocks", type=int, default=12)
  64
  65 parser.add_argument("--dropout", type=float, default=0.1)
  66
  67 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  68
  69 parser.add_argument("--no_checkpoint", action="store_true", default=False)
  70
  71 parser.add_argument("--overwrite_results", action="store_true", default=False)
  72
  73 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
  74
  75 ##############################
  76 # picoclvr options
  77
  78 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
  79
  80 parser.add_argument("--picoclvr_height", type=int, default=12)
  81
  82 parser.add_argument("--picoclvr_width", type=int, default=16)
  83
  84 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
  85
  86 ##############################
  87 # Maze options
  88
  89 parser.add_argument("--maze_height", type=int, default=13)
  90
  91 parser.add_argument("--maze_width", type=int, default=21)
  92
  93 parser.add_argument("--maze_nb_walls", type=int, default=15)
  94
  95 ##############################
  96 # Snake options
  97
  98 parser.add_argument("--snake_height", type=int, default=6)
  99
 100 parser.add_argument("--snake_width", type=int, default=8)
 101
 102 parser.add_argument("--snake_nb_colors", type=int, default=3)
 103
 104 parser.add_argument("--snake_length", type=int, default=400)
 105
 106 ######################################################################
 107
 108 args = parser.parse_args()
 109
 110 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 111
 112 try:
 113     os.mkdir(args.result_dir)
 114 except FileExistsError:
 115     if not args.overwrite_results:
 116         print(f"result directory {args.result_dir} already exists")
 117         exit(1)
 118
 119 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 120
 121 if args.seed >= 0:
 122     # torch.backends.cudnn.deterministic = True
 123     # torch.backends.cudnn.benchmark = False
 124     # torch.use_deterministic_algorithms(True)
 125     torch.manual_seed(args.seed)
 126     if torch.cuda.is_available():
 127         torch.cuda.manual_seed_all(args.seed)
 128
 129 ######################################################################
 130
 131
 132 def log_string(s):
 133     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 134
 135     if log_file is not None:
 136         log_file.write(t + s + "\n")
 137         log_file.flush()
 138
 139     print(t + s)
 140     sys.stdout.flush()
 141
 142
 143 for n in vars(args):
 144     log_string(f"args.{n} {getattr(args, n)}")
 145
 146 ######################################################################
 147
 148
 149 def masked_inplace_autoregression(
 150     model, batch_size, input, ar_mask, forbidden_tokens=None, device=torch.device("cpu")
 151 ):
 152     for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
 153         i = (ar_mask.sum(0) > 0).nonzero()
 154         if i.min() > 0:
 155             model(
 156                 mygpt.BracketedSequence(input, 0, i.min())
 157             )  # Needed to initialize the model's cache
 158         for s in range(i.min(), i.max() + 1):
 159             output = model(mygpt.BracketedSequence(input, s, 1)).x
 160             logits = output[:, s]
 161             if forbidden_tokens is not None:
 162                 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
 163             if args.deterministic_synthesis:
 164                 t_next = logits.argmax(1)
 165             else:
 166                 dist = torch.distributions.categorical.Categorical(logits=logits)
 167                 t_next = dist.sample()
 168             input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 169
 170
 171 ######################################################################
 172
 173
 174 class Task:
 175     def batches(self, split="train"):
 176         pass
 177
 178     def vocabulary_size(self):
 179         pass
 180
 181     def produce_results(self, n_epoch, model):
 182         pass
 183
 184
 185 ######################################################################
 186
 187 import picoclvr
 188
 189
 190 class TaskPicoCLVR(Task):
 191     # Make a tensor from a list of strings
 192     def tensorize(self, descr):
 193         token_descr = [s.strip().split(" ") for s in descr]
 194         l = max([len(s) for s in token_descr])
 195         token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
 196         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
 197         return torch.tensor(id_descr, device=self.device)
 198
 199     # Make a list of strings from a tensor
 200     def detensorize(self, x):
 201         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
 202
 203     # trim all the tensors in the tuple z to remove as much token from
 204     # left and right in the first tensor. If z is a tuple, all its
 205     # elements are trimed according to the triming for the first
 206     def trim(self, z, token="<nul>"):
 207         n = self.token2id[token]
 208         if type(z) == tuple:
 209             x = z[0]
 210             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 211             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 212             return tuple([t[:, a:b] for t in z])
 213         else:
 214             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 215             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 216             return z[:, a:b]
 217
 218     ######################
 219     # Not the cleanest part of the code
 220
 221     # Extract the last image of each sequence, from the last <img>
 222     # included, and set to <nul> all the tokens from the beginning of
 223     # that image to the end
 224     def excise_last_image(self, input):
 225         t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 226         nb_img_tokens = self.height * self.width + 1
 227
 228         input = input.clone()
 229         t = (input == t_img).long()
 230         tail_masks = (t.cumsum(dim=1) == t.sum(dim=1, keepdim=True)).long()
 231         i = (t * tail_masks).nonzero(as_tuple=True)
 232         j = (
 233             i[0][:, None],
 234             i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
 235         )
 236         images = self.trim(input[j])
 237         input[j] = t_nul
 238         loss_masks = 1 - tail_masks
 239         input, loss_masks = self.trim((input, loss_masks))
 240         return input, loss_masks, images
 241
 242     def add_true_image(self, input, images, loss_masks):
 243         t_nul = self.token2id["<nul>"]
 244         nb_img_tokens = self.height * self.width + 1
 245         input = F.pad(input, (0, nb_img_tokens), value=t_nul)
 246         loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
 247         t = (input == t_nul).long()
 248         i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
 249         j = (
 250             i[0][:, None],
 251             i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
 252         )
 253         input[j] = images
 254         loss_masks[j] = 1
 255         input, loss_masks = self.trim((input, loss_masks))
 256         return input, loss_masks
 257
 258     def add_generated_image(self, input, loss_masks, model):
 259         t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 260         nb_img_tokens = self.height * self.width + 1
 261
 262         input = F.pad(input, (0, nb_img_tokens), value=t_nul)
 263         loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
 264         t = (input == t_nul).long()
 265         i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
 266         input[i] = t_img
 267
 268         j = (
 269             i[0][:, None],
 270             i[1][:, None]
 271             + 1
 272             + torch.arange(nb_img_tokens - 1, device=input.device)[None, :],
 273         )
 274         ar_masks = input.new_zeros(input.size(), dtype=torch.int64)
 275         ar_masks[j] = 1
 276         forbidden_tokens = (
 277             torch.arange(self.vocabulary_size(), device=input.device) == t_nul
 278         )
 279         with torch.autograd.no_grad():
 280             t = model.training
 281             model.eval()
 282             masked_inplace_autoregression(
 283                 model,
 284                 self.batch_size,
 285                 input,
 286                 ar_masks,
 287                 forbidden_tokens,
 288                 device=self.device,
 289             )
 290             model.train(t)
 291
 292         input, loss_masks = self.trim((input, loss_masks))
 293
 294         return input, loss_masks
 295
 296     ######################
 297
 298     def __init__(
 299         self,
 300         nb_train_samples,
 301         nb_test_samples,
 302         batch_size,
 303         height,
 304         width,
 305         nb_colors=5,
 306         device=torch.device("cpu"),
 307         pruner_train=None,
 308         pruner_eval=None,
 309     ):
 310         def generate_descr(nb, cache_suffix, pruner):
 311             return picoclvr.generate(
 312                 nb,
 313                 height=self.height,
 314                 width=self.width,
 315                 nb_colors=nb_colors,
 316                 pruner=pruner,
 317             )
 318
 319         self.height = height
 320         self.width = width
 321         self.batch_size = batch_size
 322         self.device = device
 323         self.pruner_train = pruner_train
 324         self.pruner_eval = pruner_eval
 325
 326         param = {
 327             "nb_train_samples": nb_train_samples,
 328             "nb_test_samples": nb_test_samples,
 329             "height": height,
 330             "width": width,
 331             "nb_colors": nb_colors,
 332             "batch_size": batch_size,
 333             "rng_state": list(torch.get_rng_state()),
 334         }
 335
 336         log_string(
 337             f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
 338         )
 339         self.train_descr = generate_descr(
 340             nb_train_samples, "train", pruner=self.pruner_train
 341         )
 342         self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
 343
 344         # Build the tokenizer
 345         tokens = {"<nul>", "<img>"}
 346         for d in [self.train_descr, self.test_descr]:
 347             for s in d:
 348                 for t in s.strip().split(" "):
 349                     tokens.add(t)
 350         # make this set a sorted list to get the same tensors given
 351         # the same descr
 352         tokens = list(tokens)
 353         tokens.sort()
 354         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
 355         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
 356
 357         # Tokenize the train and test sets
 358         self.train_input = self.tensorize(self.train_descr)
 359         self.test_input = self.tensorize(self.test_descr)
 360
 361     def batches(self, split="train"):
 362         assert split in {"train", "test"}
 363         input = self.train_input if split == "train" else self.test_input
 364         for batch in tqdm.tqdm(
 365             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 366         ):
 367             yield self.trim(batch)
 368
 369     def vocabulary_size(self):
 370         return len(self.token2id)
 371
 372     def compute_missing_properties(self, n_epoch, model, pruner=None):
 373         acc_nb_requested_properties = []
 374         acc_nb_missing_properties = []
 375         acc_nb_results = 0
 376
 377         for input in tqdm.tqdm(
 378             self.test_input.split(self.batch_size),
 379             dynamic_ncols=True,
 380             desc=f"test-properties",
 381         ):
 382             tape, loss_masks, _ = self.excise_last_image(input)
 383             tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
 384             result_descr = self.detensorize(tape)
 385             np = picoclvr.nb_properties(
 386                 result_descr,
 387                 height=self.height,
 388                 width=self.width,
 389                 pruner=pruner,
 390             )
 391             nb_requested_properties, _, nb_missing_properties = zip(*np)
 392             acc_nb_requested_properties += nb_requested_properties
 393             acc_nb_missing_properties += nb_missing_properties
 394             acc_nb_results += len(result_descr)
 395
 396         nb_requested_properties = sum(acc_nb_requested_properties)
 397         nb_missing_properties = sum(acc_nb_missing_properties)
 398
 399         prefix = "" if pruner is None else "pruned_"
 400         log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 401         log_string(
 402             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 403         )
 404         log_string(
 405             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 406         )
 407
 408     ######################################################################
 409
 410     def produce_results(self, n_epoch, model):
 411         self.compute_missing_properties(n_epoch, model)
 412
 413         if self.pruner_eval is not None:
 414             self.compute_missing_properties(n_epoch, model, self.pruner_eval)
 415
 416         nb_tokens_to_generate = self.height * self.width + 3
 417         result_descr = []
 418         nb_per_primer = 8
 419         primer = []
 420
 421         for primer_descr in [
 422             "red above green <sep> green top <sep> blue right of red",
 423             "there is red <sep> there is yellow <sep> there is blue",
 424             "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
 425             "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
 426         ]:
 427             primer += [primer_descr] * nb_per_primer
 428
 429         tape = self.tensorize(primer)
 430         loss_masks = 1 - (tape == self.token2id["<nul>"]).long()
 431         tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
 432         result_descr = self.detensorize(tape)
 433
 434         np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
 435
 436         acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
 437         acc_nb_results = len(result_descr)
 438
 439         nb_requested_properties = sum(acc_nb_requested_properties)
 440         nb_missing_properties = sum(acc_nb_missing_properties)
 441
 442         prefix = "demo_"
 443         log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 444         log_string(
 445             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 446         )
 447         log_string(
 448             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 449         )
 450
 451         img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
 452
 453         if img.dim() == 5:
 454             if img.size(1) == 1:
 455                 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
 456             else:
 457                 img = torch.cat(
 458                     [
 459                         torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
 460                         for x in img
 461                     ],
 462                     0,
 463                 )
 464
 465         image_name = os.path.join(args.result_dir, f"picoclvr_result_{n_epoch:04d}.png")
 466         torchvision.utils.save_image(
 467             img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=1.0
 468         )
 469         log_string(f"wrote {image_name}")
 470
 471
 472 ######################################################################
 473
 474
 475 class TaskMNIST(Task):
 476     def __init__(self, batch_size, device=torch.device("cpu")):
 477         self.device = device
 478         self.batch_size = batch_size
 479
 480     def batches(self, split="train"):
 481         assert split in {"train", "test"}
 482         data_set = torchvision.datasets.MNIST(
 483             root="./data", train=(split == "train"), download=True
 484         )
 485         data_input = data_set.data.view(-1, 28 * 28).long()
 486         if args.nb_train_samples is not None:
 487             data_input = data_input[: args.nb_train_samples]
 488         for batch in tqdm.tqdm(
 489             data_input.split(self.batch_size), desc=f"epoch-{split}"
 490         ):
 491             yield batch
 492
 493     def vocabulary_size(self):
 494         return 256
 495
 496     def produce_results(self, n_epoch, model):
 497         results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
 498         ar_mask = torch.full_like(results, 1)
 499         masked_inplace_autoregression(
 500             model, self.batch_size, results, ar_mask, device=self.device
 501         )
 502         image_name = os.path.join(args.result_dir, f"mnist_result_{n_epoch:04d}.png")
 503         torchvision.utils.save_image(
 504             1 - results.reshape(-1, 1, 28, 28) / 255.0,
 505             image_name,
 506             nrow=16,
 507             pad_value=0.8,
 508         )
 509         log_string(f"wrote {image_name}")
 510
 511
 512 ######################################################################
 513
 514 import maze
 515
 516
 517 class TaskMaze(Task):
 518     def map2seq(self, *m):
 519         return torch.cat([x.flatten(1) for x in m], 1)
 520
 521     def seq2map(self, s):
 522         s = s.reshape(s.size(0), -1, self.height, self.width)
 523         return (s[:, k] for k in range(s.size(1)))
 524
 525     def __init__(
 526         self,
 527         nb_train_samples,
 528         nb_test_samples,
 529         batch_size,
 530         height,
 531         width,
 532         nb_walls,
 533         device=torch.device("cpu"),
 534     ):
 535         self.batch_size = batch_size
 536         self.height = height
 537         self.width = width
 538         self.device = device
 539
 540         train_mazes, train_paths, _ = maze.create_maze_data(
 541             nb_train_samples,
 542             height=height,
 543             width=width,
 544             nb_walls=nb_walls,
 545             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 546         )
 547         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 548
 549         test_mazes, test_paths, _ = maze.create_maze_data(
 550             nb_test_samples,
 551             height=height,
 552             width=width,
 553             nb_walls=nb_walls,
 554             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 555         )
 556         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 557
 558         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 559
 560     def batches(self, split="train", nb_to_use=-1, desc=None):
 561         assert split in {"train", "test"}
 562         input = self.train_input if split == "train" else self.test_input
 563         if nb_to_use > 0:
 564             input = input[:nb_to_use]
 565         if desc is None:
 566             desc = f"epoch-{split}"
 567         for batch in tqdm.tqdm(
 568             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 569         ):
 570             yield batch
 571
 572     def vocabulary_size(self):
 573         return self.nb_codes
 574
 575     def compute_error(self, model, split="train", nb_to_use=-1):
 576         nb_total, nb_correct = 0, 0
 577         for input in task.batches(split, nb_to_use):
 578             result = input.clone()
 579             ar_mask = result.new_zeros(result.size())
 580             ar_mask[:, self.height * self.width :] = 1
 581             result *= 1 - ar_mask
 582             masked_inplace_autoregression(
 583                 model, self.batch_size, result, ar_mask, device=self.device
 584             )
 585             mazes, paths = self.seq2map(result)
 586             nb_correct += maze.path_correctness(mazes, paths).long().sum()
 587             nb_total += mazes.size(0)
 588
 589         return nb_total, nb_correct
 590
 591     def produce_results(self, n_epoch, model):
 592         with torch.autograd.no_grad():
 593             t = model.training
 594             model.eval()
 595
 596             train_nb_total, train_nb_correct = self.compute_error(
 597                 model, "train", nb_to_use=1000
 598             )
 599             log_string(
 600                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 601             )
 602
 603             test_nb_total, test_nb_correct = self.compute_error(
 604                 model, "test", nb_to_use=1000
 605             )
 606             log_string(
 607                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 608             )
 609
 610             input = self.test_input[:48]
 611             result = input.clone()
 612             ar_mask = result.new_zeros(result.size())
 613             ar_mask[:, self.height * self.width :] = 1
 614             result *= 1 - ar_mask
 615             masked_inplace_autoregression(
 616                 model, self.batch_size, result, ar_mask, device=self.device
 617             )
 618
 619             mazes, paths = self.seq2map(input)
 620             _, predicted_paths = self.seq2map(result)
 621
 622             filename = os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.png")
 623             maze.save_image(
 624                 filename,
 625                 mazes=mazes,
 626                 target_paths=paths,
 627                 predicted_paths=predicted_paths,
 628                 path_correct=maze.path_correctness(mazes, predicted_paths),
 629             )
 630             log_string(f"wrote {filename}")
 631
 632             model.train(t)
 633
 634
 635 ######################################################################
 636
 637
 638 def generate_snake_sequences(
 639     nb, height, width, nb_colors, length, device=torch.device("cpu")
 640 ):
 641     worlds = torch.randint(nb_colors, (nb, height, width), device=device)
 642     # nb x 2
 643     snake_position = torch.cat(
 644         (
 645             torch.randint(height, (nb, 1), device=device),
 646             torch.randint(width, (nb, 1), device=device),
 647         ),
 648         1,
 649     )
 650     snake_direction = torch.randint(4, (nb,), device=device)
 651     sequences = torch.empty(nb, 2 * length, device=device, dtype=torch.int64)
 652     i = torch.arange(nb, device=device)  # [:,None]
 653
 654     for l in range(length):
 655         # nb x 3
 656         snake_next_direction = torch.cat(
 657             (
 658                 (snake_direction[:, None] - 1) % 4,
 659                 snake_direction[:, None],
 660                 (snake_direction[:, None] + 1) % 4,
 661             ),
 662             1,
 663         )
 664
 665         # nb x 3
 666         vh = (snake_next_direction + 1) % 2 * (snake_next_direction - 1)
 667         vw = snake_next_direction % 2 * (snake_next_direction - 2)
 668
 669         # nb x 3 x 2
 670         snake_next_speed = torch.cat((vh[:, :, None], vw[:, :, None]), 2)
 671         snake_next_position = snake_position[:, None, :] + snake_next_speed
 672
 673         # nb x 3
 674         val = torch.logical_and(
 675             torch.logical_and(
 676                 snake_next_position[:, :, 0] >= 0, snake_next_position[:, :, 0] < height
 677             ),
 678             torch.logical_and(
 679                 snake_next_position[:, :, 1] >= 0, snake_next_position[:, :, 1] < width
 680             ),
 681         ).float()
 682         val = (
 683             torch.rand_like(val) * val * torch.tensor([[1.0, 4.0, 1.0]], device=device)
 684         )
 685
 686         # nb
 687         j = val.argmax(1)
 688         snake_direction = snake_next_direction[i, j]
 689
 690         sequences[:, 2 * l] = worlds[i, snake_position[:, 0], snake_position[:, 1]] + 4
 691         sequences[:, 2 * l + 1] = snake_direction
 692
 693         # nb x 2
 694         snake_position = snake_next_position[i, j]
 695
 696     return sequences, worlds
 697
 698
 699 # generate_snake_sequences(nb=1, height=4, width=6, nb_colors=3, length=20)
 700 # exit(0)
 701
 702
 703 class TaskSnake(Task):
 704     def __init__(
 705         self,
 706         nb_train_samples,
 707         nb_test_samples,
 708         batch_size,
 709         height,
 710         width,
 711         nb_colors,
 712         length,
 713         device=torch.device("cpu"),
 714     ):
 715         self.batch_size = batch_size
 716         self.height = height
 717         self.width = width
 718         self.device = device
 719
 720         self.train_input, self.train_worlds = generate_snake_sequences(
 721             nb_train_samples, height, width, nb_colors, length, self.device
 722         )
 723         self.test_input, self.test_worlds = generate_snake_sequences(
 724             nb_test_samples, height, width, nb_colors, length, self.device
 725         )
 726
 727         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 728
 729     def batches(self, split="train", nb_to_use=-1, desc=None):
 730         assert split in {"train", "test"}
 731         input = self.train_input if split == "train" else self.test_input
 732         if nb_to_use > 0:
 733             input = input[:nb_to_use]
 734         if desc is None:
 735             desc = f"epoch-{split}"
 736         for batch in tqdm.tqdm(
 737             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 738         ):
 739             yield batch
 740
 741     def vocabulary_size(self):
 742         return self.nb_codes
 743
 744     def produce_results(self, n_epoch, model):
 745         with torch.autograd.no_grad():
 746             t = model.training
 747             model.eval()
 748
 749             def compute_nb_correct(input):
 750                 result = input.clone()
 751                 i = torch.arange(result.size(1), device=result.device)
 752                 ar_mask = torch.logical_and(i >= i.size(0) // 2, i % 2 == 0)[
 753                     None, :
 754                 ].long()
 755                 result *= 1 - ar_mask
 756                 masked_inplace_autoregression(
 757                     model, self.batch_size, result, ar_mask, device=self.device
 758                 )
 759
 760                 nb_total = ar_mask.sum() * input.size(0)
 761                 nb_correct = ((result == input).long() * ar_mask).sum()
 762
 763                 # nb_total = result.size(0)
 764                 # nb_correct = ((result - input).abs().sum(1) == 0).sum()
 765
 766                 return nb_total, nb_correct
 767
 768             train_nb_total, train_nb_correct = compute_nb_correct(self.train_input)
 769
 770             log_string(
 771                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 772             )
 773
 774             test_nb_total, test_nb_correct = compute_nb_correct(self.test_input)
 775
 776             log_string(
 777                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 778             )
 779
 780             model.train(t)
 781
 782
 783 ######################################################################
 784
 785
 786 def picoclvr_pruner_horizontal_green(p):
 787     return not ("green" in p and ("left" in p or "right" in p))
 788
 789
 790 picoclvr_pruner_train = (
 791     picoclvr_pruner_horizontal_green
 792     if args.picocvlr_prune_properties in {"train+eval"}
 793     else None
 794 )
 795
 796 picoclvr_pruner_eval = (
 797     (lambda p: not picoclvr_pruner_horizontal_green(p))
 798     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 799     else None
 800 )
 801
 802 ######################################################################
 803
 804 if args.task == "picoclvr":
 805     task = TaskPicoCLVR(
 806         nb_train_samples=args.nb_train_samples,
 807         nb_test_samples=args.nb_test_samples,
 808         batch_size=args.batch_size,
 809         height=args.picoclvr_height,
 810         width=args.picoclvr_width,
 811         nb_colors=args.picoclvr_nb_colors,
 812         device=device,
 813         pruner_train=picoclvr_pruner_train,
 814         pruner_eval=picoclvr_pruner_eval,
 815     )
 816
 817 elif args.task == "mnist":
 818     task = TaskMNIST(
 819         batch_size=args.batch_size,
 820         device=device,
 821     )
 822
 823 elif args.task == "maze":
 824     task = TaskMaze(
 825         nb_train_samples=args.nb_train_samples,
 826         nb_test_samples=args.nb_test_samples,
 827         batch_size=args.batch_size,
 828         height=args.maze_height,
 829         width=args.maze_width,
 830         nb_walls=args.maze_nb_walls,
 831         device=device,
 832     )
 833
 834 elif args.task == "snake":
 835     task = TaskSnake(
 836         nb_train_samples=args.nb_train_samples,
 837         nb_test_samples=args.nb_test_samples,
 838         batch_size=args.batch_size,
 839         height=args.snake_height,
 840         width=args.snake_width,
 841         nb_colors=args.snake_nb_colors,
 842         length=args.snake_length,
 843         device=device,
 844     )
 845
 846 else:
 847     raise ValueError(f"Unknown task {args.task}")
 848
 849 ######################################################################
 850
 851 log_string(f"device {device}")
 852
 853 vocabulary_size = task.vocabulary_size()
 854
 855 log_string(f"vocabulary_size {vocabulary_size}")
 856
 857 ##############################
 858
 859 model = mygpt.MyGPT(
 860     vocabulary_size=vocabulary_size,
 861     dim_model=args.dim_model,
 862     dim_keys=args.dim_keys,
 863     dim_hidden=args.dim_hidden,
 864     nb_heads=args.nb_heads,
 865     nb_blocks=args.nb_blocks,
 866     causal=True,
 867     dropout=args.dropout,
 868 )
 869
 870 model.to(device)
 871
 872 nb_parameters = sum(p.numel() for p in model.parameters())
 873 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 874
 875 ######################################################################
 876
 877 nb_epochs_finished = 0
 878
 879 if args.no_checkpoint:
 880     log_string(f"not trying to load checkpoint.")
 881
 882 else:
 883     try:
 884         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 885         checkpoint = torch.load(checkpoint_name)
 886         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 887         model.load_state_dict(checkpoint["model_state"])
 888         torch.set_rng_state(checkpoint["rng_state"])
 889         if torch.cuda.is_available():
 890             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 891
 892         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 893
 894     except FileNotFoundError:
 895         log_string("starting from scratch.")
 896
 897     except:
 898         log_string("error when loading the checkpoint.")
 899         exit(1)
 900
 901 ######################################################################
 902
 903 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 904
 905 token_count = 0
 906 for input in task.batches(split="train"):
 907     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 908 token_probas = token_count / token_count.sum()
 909 entropy = -torch.xlogy(token_probas, token_probas).sum()
 910 train_set_perplexity = math.exp(entropy)
 911
 912 ##############################
 913
 914 if args.learning_rate_schedule == "cos":
 915     learning_rate_schedule = {}
 916     for n_epoch in range(args.nb_epochs):
 917         u = n_epoch / args.nb_epochs * math.pi
 918         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 919 else:
 920     u = {
 921         int(k): float(v)
 922         for k, v in [
 923             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 924         ]
 925     }
 926
 927     learning_rate_schedule = {}
 928     learning_rate = args.learning_rate
 929     for n_epoch in range(args.nb_epochs):
 930         if n_epoch in u:
 931             learning_rate = u[n_epoch]
 932         learning_rate_schedule[n_epoch] = learning_rate
 933
 934 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 935
 936 ##############################
 937
 938 nb_samples_seen = 0
 939
 940 if nb_epochs_finished >= nb_epochs:
 941     task.produce_results(nb_epochs_finished, model)
 942
 943 for n_epoch in range(nb_epochs_finished, nb_epochs):
 944     learning_rate = learning_rate_schedule[n_epoch]
 945
 946     log_string(f"learning_rate {learning_rate}")
 947
 948     if args.optim == "sgd":
 949         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 950     elif args.optim == "adam":
 951         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 952     elif args.optim == "adamw":
 953         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 954     else:
 955         raise ValueError(f"Unknown optimizer {args.optim}.")
 956
 957     model.train()
 958
 959     nb_train_samples, acc_train_loss = 0, 0.0
 960
 961     for input in task.batches(split="train"):
 962         input = input.to(device)
 963         output = model(mygpt.BracketedSequence(input)).x
 964         loss = F.cross_entropy(output.transpose(1, 2), input)
 965         acc_train_loss += loss.item() * input.size(0)
 966         nb_train_samples += input.size(0)
 967         nb_samples_seen += input.size(0)
 968
 969         optimizer.zero_grad()
 970         loss.backward()
 971         optimizer.step()
 972
 973     with torch.autograd.no_grad():
 974         model.eval()
 975
 976         nb_test_samples, acc_test_loss = 0, 0.0
 977
 978         for input in task.batches(split="test"):
 979             input = input.to(device)
 980
 981             # input, loss_masks, true_images = task.excise_last_image(input)
 982             # input, loss_masks = task.add_true_image(input, true_images, loss_masks)
 983
 984             output = model(mygpt.BracketedSequence(input)).x
 985             loss = F.cross_entropy(output.transpose(1, 2), input)
 986             acc_test_loss += loss.item() * input.size(0)
 987             nb_test_samples += input.size(0)
 988
 989         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 990         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 991
 992         log_string(
 993             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 994         )
 995
 996         task.produce_results(n_epoch, model)
 997
 998     checkpoint = {
 999         "nb_epochs_finished": n_epoch + 1,
1000         "model_state": model.state_dict(),
1001         "rng_state": torch.get_rng_state(),
1002     }
1003
1004     if torch.cuda.is_available():
1005         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
1006
1007     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
1008     torch.save(checkpoint, checkpoint_name)
1009     log_string(f"saved checkpoint {checkpoint_name}")
1010
1011 ######################################################################