beaver.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(description="A maze shortest path solving with a GPT.")
  30
  31 parser.add_argument("--log_filename", type=str, default="train.log")
  32
  33 parser.add_argument("--result_dir", type=str, default="results_default")
  34
  35 parser.add_argument("--seed", type=int, default=0)
  36
  37 parser.add_argument("--nb_epochs", type=int, default=25)
  38
  39 parser.add_argument("--nb_train_samples", type=int, default=200000)
  40
  41 parser.add_argument("--nb_test_samples", type=int, default=50000)
  42
  43 parser.add_argument("--batch_size", type=int, default=25)
  44
  45 parser.add_argument("--optim", type=str, default="adam")
  46
  47 parser.add_argument("--learning_rate", type=float, default=1e-3)
  48
  49 parser.add_argument(
  50     "--learning_rate_schedule", type=str, default="10: 2e-4,20: 4e-5,30: 8e-6"
  51 )
  52
  53 parser.add_argument("--dim_model", type=int, default=512)
  54
  55 parser.add_argument("--dim_keys", type=int, default=64)
  56
  57 parser.add_argument("--dim_hidden", type=int, default=2048)
  58
  59 parser.add_argument("--nb_heads", type=int, default=8)
  60
  61 parser.add_argument("--nb_blocks", type=int, default=12)
  62
  63 parser.add_argument("--dropout", type=float, default=0.1)
  64
  65 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  66
  67 parser.add_argument("--random_regression_order", action="store_true", default=False)
  68
  69 parser.add_argument("--noncausal_prompt", action="store_true", default=False)
  70
  71 parser.add_argument("--no_checkpoint", action="store_true", default=False)
  72
  73 parser.add_argument("--overwrite_results", action="store_true", default=False)
  74
  75 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
  76
  77 ##############################
  78 # maze options
  79
  80 parser.add_argument("--maze_height", type=int, default=13)
  81
  82 parser.add_argument("--maze_width", type=int, default=21)
  83
  84 parser.add_argument("--maze_nb_walls", type=int, default=15)
  85
  86 ##############################
  87 # one-shot prediction
  88
  89 parser.add_argument("--oneshot", action="store_true", default=False)
  90
  91 parser.add_argument("--oneshot_input", type=str, default="head")
  92
  93 parser.add_argument("--oneshot_output", type=str, default="trace")
  94
  95 ######################################################################
  96
  97 args = parser.parse_args()
  98
  99 try:
 100     os.mkdir(args.result_dir)
 101 except FileExistsError:
 102     if not args.overwrite_results:
 103         print(f"result directory {args.result_dir} already exists")
 104         exit(1)
 105
 106 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 107
 108 if args.seed >= 0:
 109     # torch.backends.cudnn.deterministic = True
 110     # torch.backends.cudnn.benchmark = False
 111     # torch.use_deterministic_algorithms(True)
 112     torch.manual_seed(args.seed)
 113     if torch.cuda.is_available():
 114         torch.cuda.manual_seed_all(args.seed)
 115
 116 ######################################################################
 117
 118
 119 def log_string(s):
 120     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 121
 122     if log_file is not None:
 123         log_file.write(t + s + "\n")
 124         log_file.flush()
 125
 126     print(t + s)
 127     sys.stdout.flush()
 128
 129
 130 for n in vars(args):
 131     log_string(f"args.{n} {getattr(args, n)}")
 132
 133 ######################################################################
 134
 135
 136 def reorder(x, order, reverse=False):  # x is NxTxD1x...xDk, order is NxT'
 137     u = x.reshape(x.size()[:2] + (-1,))
 138     order = order.unsqueeze(-1).expand(-1, -1, u.size(-1))
 139     if reverse:
 140         v = u.new(u.size()).scatter_(1, order, u)
 141     else:
 142         v = u.gather(1, order)
 143     v = v.reshape(v.size()[:2] + x.size()[2:])
 144     return v
 145
 146
 147 def shuffle(x, prompt_len):
 148     if args.random_regression_order:
 149         order = torch.rand(x.size(), device=x.device)
 150         order[:, :prompt_len] = torch.arange(-prompt_len, 0, device=x.device)
 151         order = order.sort(1).indices
 152     else:
 153         order = (
 154             torch.arange(x.size(1), device=x.device).unsqueeze(0).expand(x.size(0), -1)
 155         )
 156     return reorder(x, order), order
 157
 158
 159 def eval_mygpt(model, input, mode="standard", prompt_len=0):
 160     x, order = shuffle(input, prompt_len)
 161     x = model(mygpt.BracketedSequence(x), mode=mode, order=order).x
 162     return reorder(x, order, reverse=True)
 163
 164
 165 ######################################################################
 166
 167 # ar_mask is a Boolean matrix of same shape as input, with 1s on the
 168 # tokens that should be generated
 169
 170
 171 def masked_inplace_autoregression(model, batch_size, input, ar_mask, order=None):
 172     for input, ar_mask, order in zip(
 173         input.split(batch_size), ar_mask.split(batch_size), order.split(batch_size)
 174     ):
 175         i = (ar_mask.sum(0) > 0).nonzero()
 176         if i.min() > 0:
 177             # Needed to initialize the model's cache
 178             model(mygpt.BracketedSequence(input, 0, i.min()), order=order)
 179         for s in range(i.min(), i.max() + 1):
 180             output = model(mygpt.BracketedSequence(input, s, 1), order=order).x
 181             logits = output[:, s]
 182             if args.deterministic_synthesis:
 183                 t_next = logits.argmax(1)
 184             else:
 185                 dist = torch.distributions.categorical.Categorical(logits=logits)
 186                 t_next = dist.sample()
 187             input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 188
 189
 190 ######################################################################
 191
 192
 193 def compute_perplexity(model, task, prompt_len, split="train"):
 194     with torch.autograd.no_grad():
 195         t = model.training
 196         model.eval()
 197
 198         nb_samples, acc_loss = 0, 0.0
 199
 200         for input in task.batches(split=split):
 201             input = input.to(device)
 202             output = eval_mygpt(model, input, prompt_len=prompt_len)
 203             if args.noncausal_prompt:
 204                 d = input.size(1) // 2
 205                 loss = F.cross_entropy(output[:, d:].transpose(1, 2), input[:, d:])
 206             else:
 207                 loss = F.cross_entropy(output.transpose(1, 2), input)
 208             acc_loss += loss.item() * input.size(0)
 209             nb_samples += input.size(0)
 210
 211         model.train(t)
 212
 213         return math.exp(min(100, acc_loss / nb_samples))
 214
 215
 216 ######################################################################
 217
 218
 219 def oneshot_policy_loss(mazes, output, policies, height, width):
 220     masks = (mazes == maze.v_empty).unsqueeze(-1)
 221     targets = policies.permute(0, 2, 1) * masks
 222     output = output * masks
 223     return -(output.log_softmax(-1) * targets).sum() / masks.sum()
 224
 225
 226 def oneshot_trace_loss(mazes, output, policies, height, width):
 227     masks = mazes == maze.v_empty
 228     targets = maze.stationary_densities(
 229         mazes.view(-1, height, width), policies.view(-1, 4, height, width)
 230     ).flatten(-2)
 231     targets = targets * masks
 232     output = output.squeeze(-1) * masks
 233     return (output - targets).abs().sum() / masks.sum()
 234
 235
 236 def oneshot(model, learning_rate_scheduler, task):
 237     t = model.training
 238     model.eval()
 239     mazes = task.test_input[:32].clone()
 240     mazes[:, task.height * task.width :] = 0
 241     output = eval_mygpt(model, mazes, prompt_len=task.height * task.width)
 242     output = F.softmax(output, dim=2)
 243     print(f"{output.size()=}")
 244     proba_path = output[:, task.height * task.width :, 4].reshape(
 245         -1, task.height, task.width
 246     )
 247     mazes = mazes[:, : task.height * task.width].reshape(-1, task.height, task.width)
 248     # targets = targets.reshape(-1, task.height, task.width)
 249     filename = f"oneshot.png"
 250     maze.save_image(
 251         os.path.join(args.result_dir, filename),
 252         mazes=mazes,
 253         score_paths=proba_path,
 254         # score_truth=targets,
 255     )
 256     log_string(f"wrote {filename}")
 257
 258
 259 def oneshot_old(gpt, learning_rate_scheduler, task):
 260     t = gpt.training
 261     gpt.eval()
 262
 263     if args.oneshot_input == "head":
 264         dim_in = args.dim_model
 265     elif args.oneshot_input == "deep":
 266         dim_in = args.dim_model * args.nb_blocks * 2
 267     else:
 268         raise ValueError(f"{args.oneshot_input=}")
 269
 270     if args.oneshot_output == "policy":
 271         dim_out = 4
 272         compute_loss = oneshot_policy_loss
 273     elif args.oneshot_output == "trace":
 274         dim_out = 1
 275         compute_loss = oneshot_trace_loss
 276     else:
 277         raise ValueError(f"{args.oneshot_output=}")
 278
 279     model = nn.Sequential(
 280         nn.Linear(dim_in, args.dim_model),
 281         nn.ReLU(),
 282         nn.Linear(args.dim_model, args.dim_model),
 283         nn.ReLU(),
 284         nn.Linear(args.dim_model, dim_out),
 285     ).to(device)
 286
 287     learning_rate_scheduler.reset()
 288
 289     for n_epoch in range(args.nb_epochs):
 290         learning_rate = learning_rate_scheduler.get_learning_rate()
 291         log_string(f"learning_rate {n_epoch} {learning_rate}")
 292
 293         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 294
 295         acc_train_loss, nb_train_samples = 0, 0
 296         for mazes, policies in task.policy_batches(split="train"):
 297             output_gpt = eval_mygpt(
 298                 gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
 299             )
 300             output = model(output_gpt)
 301
 302             loss = compute_loss(mazes, output, policies, task.height, task.width)
 303             acc_train_loss += loss.item() * mazes.size(0)
 304             nb_train_samples += mazes.size(0)
 305
 306             optimizer.zero_grad()
 307             loss.backward()
 308             optimizer.step()
 309
 310         learning_rate_scheduler.update(n_epoch + 1, acc_train_loss)
 311
 312         acc_test_loss, nb_test_samples = 0, 0
 313         for mazes, policies in task.policy_batches(split="test"):
 314             output_gpt = eval_mygpt(
 315                 gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
 316             )
 317             output = model(output_gpt)
 318             loss = compute_loss(mazes, output, policies, task.height, task.width)
 319             acc_test_loss += loss.item() * mazes.size(0)
 320             nb_test_samples += mazes.size(0)
 321
 322         log_string(
 323             f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
 324         )
 325
 326         # -------------------
 327         mazes = task.test_input[:32, : task.height * task.width]
 328         policies = task.test_policies[:32]
 329         output_gpt = eval_mygpt(
 330             gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
 331         )
 332         output = model(output_gpt)
 333         if args.oneshot_output == "policy":
 334             targets = policies.permute(0, 2, 1)
 335             scores = (
 336                 (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
 337             ).float()
 338         elif args.oneshot_output == "trace":
 339             targets = maze.stationary_densities(
 340                 mazes.view(-1, task.height, task.width),
 341                 policies.view(-1, 4, task.height, task.width),
 342             ).flatten(-2)
 343             scores = output
 344         else:
 345             raise ValueError(f"{args.oneshot_output=}")
 346
 347         scores = scores.reshape(-1, task.height, task.width)
 348         mazes = mazes.reshape(-1, task.height, task.width)
 349         targets = targets.reshape(-1, task.height, task.width)
 350         filename = (
 351             f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png"
 352         )
 353         maze.save_image(
 354             os.path.join(args.result_dir, filename),
 355             mazes=mazes,
 356             score_paths=scores,
 357             score_truth=targets,
 358         )
 359         log_string(f"wrote {filename}")
 360
 361         # -------------------
 362
 363     gpt.train(t)
 364
 365
 366 ######################################################################
 367
 368
 369 class LearningRateScheduler:
 370     def get_learning_rate(self):
 371         pass
 372
 373     def update(self, nb_finished_epochs, loss):
 374         pass
 375
 376     def reset(self):
 377         pass
 378
 379     def get_state(self):
 380         return vars(self)
 381
 382     def set_state(self, state):
 383         print(f"{state=}")
 384         for k, v in state.items():
 385             setattr(self, k, v)
 386
 387
 388 class StepWiseScheduler(LearningRateScheduler):
 389     def __init__(self, schedule):
 390         self.nb_finished_epochs = 0
 391         self.schedule = schedule
 392
 393     def get_learning_rate(self):
 394         return self.schedule[self.nb_finished_epochs]
 395
 396     def update(self, nb_finished_epochs, loss):
 397         self.nb_finished_epochs = nb_finished_epochs
 398
 399     def reset(self):
 400         self.nb_finished_epochs = 0
 401
 402     def get_state(self):
 403         return {"nb_finished_epochs": self.nb_finished_epochs}
 404
 405
 406 class AutoScheduler(LearningRateScheduler):
 407     def __init__(self, learning_rate_init, growth=1.0, degrowth=0.2):
 408         self.learning_rate_init = learning_rate_init
 409         self.learning_rate = learning_rate_init
 410         self.growth = growth
 411         self.degrowth = degrowth
 412         self.pred_loss = None
 413
 414     def get_learning_rate(self):
 415         return self.learning_rate
 416
 417     def update(self, nb_finished_epochs, loss):
 418         if self.pred_loss is not None:
 419             if loss >= self.pred_loss:
 420                 self.learning_rate *= self.degrowth
 421             else:
 422                 self.learning_rate *= self.growth
 423         self.pred_loss = loss
 424
 425     def reset(self):
 426         self.learning_rate = self.learning_rate_init
 427
 428     def get_state(self):
 429         return {
 430             "learning_rate_init": self.learning_rate_init,
 431             "pred_loss": self.pred_loss,
 432         }
 433
 434
 435 ######################################################################
 436
 437
 438 class Task:
 439     def batches(self, split="train", nb_to_use=-1, desc=None):
 440         pass
 441
 442     def vocabulary_size(self):
 443         pass
 444
 445     def produce_results(self, n_epoch, model):
 446         pass
 447
 448
 449 ######################################################################
 450
 451 import maze
 452
 453
 454 class TaskMaze(Task):
 455     def map2seq(self, *m):
 456         return torch.cat([x.flatten(1) for x in m], 1)
 457
 458     def seq2map(self, s):
 459         s = s.reshape(s.size(0), -1, self.height, self.width)
 460         return (s[:, k] for k in range(s.size(1)))
 461
 462     def __init__(
 463         self,
 464         nb_train_samples,
 465         nb_test_samples,
 466         batch_size,
 467         height,
 468         width,
 469         nb_walls,
 470         device=torch.device("cpu"),
 471     ):
 472         self.batch_size = batch_size
 473         self.height = height
 474         self.width = width
 475         self.device = device
 476
 477         train_mazes, train_paths, train_policies = maze.create_maze_data(
 478             nb_train_samples,
 479             height=height,
 480             width=width,
 481             nb_walls=nb_walls,
 482             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 483         )
 484         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 485         self.train_policies = train_policies.flatten(-2).to(device)
 486
 487         test_mazes, test_paths, test_policies = maze.create_maze_data(
 488             nb_test_samples,
 489             height=height,
 490             width=width,
 491             nb_walls=nb_walls,
 492             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 493         )
 494         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 495         self.test_policies = test_policies.flatten(-2).to(device)
 496
 497         self.nb_codes = self.train_input.max() + 1
 498
 499     def batches(self, split="train", nb_to_use=-1, desc=None):
 500         assert split in {"train", "test"}
 501         input = self.train_input if split == "train" else self.test_input
 502         if nb_to_use > 0:
 503             input = input[:nb_to_use]
 504         if desc is None:
 505             desc = f"epoch-{split}"
 506         for batch in tqdm.tqdm(
 507             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 508         ):
 509             yield batch
 510
 511     def policy_batches(self, split="train", nb_to_use=-1, desc=None):
 512         assert split in {"train", "test"}
 513         input = self.train_input if split == "train" else self.test_input
 514         policies = self.train_policies if split == "train" else self.test_policies
 515         input = input[:, : self.height * self.width]
 516         policies = policies * (input != maze.v_wall)[:, None]
 517
 518         if nb_to_use > 0:
 519             input = input[:nb_to_use]
 520             policies = policies[:nb_to_use]
 521
 522         if desc is None:
 523             desc = f"epoch-{split}"
 524         for batch in tqdm.tqdm(
 525             zip(input.split(self.batch_size), policies.split(self.batch_size)),
 526             dynamic_ncols=True,
 527             desc=desc,
 528         ):
 529             yield batch
 530
 531     def vocabulary_size(self):
 532         return self.nb_codes
 533
 534     def compute_error(self, model, split="train", nb_to_use=-1):
 535         nb_total, nb_correct = 0, 0
 536         for input in task.batches(split, nb_to_use):
 537             result = input.clone()
 538             ar_mask = result.new_zeros(result.size())
 539             ar_mask[:, self.height * self.width :] = 1
 540             result *= 1 - ar_mask
 541             x, order = shuffle(result, self.height * self.width)
 542             masked_inplace_autoregression(
 543                 model, self.batch_size, x, ar_mask, order=order
 544             )
 545             result = reorder(x, order, reverse=True)
 546             mazes, paths = self.seq2map(result)
 547             nb_correct += maze.path_correctness(mazes, paths).long().sum()
 548             nb_total += mazes.size(0)
 549
 550         return nb_total, nb_correct
 551
 552     def produce_results(self, n_epoch, model):
 553         with torch.autograd.no_grad():
 554             t = model.training
 555             model.eval()
 556
 557             train_nb_total, train_nb_correct = self.compute_error(
 558                 model, "train", nb_to_use=1000
 559             )
 560             log_string(
 561                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 562             )
 563
 564             test_nb_total, test_nb_correct = self.compute_error(
 565                 model, "test", nb_to_use=1000
 566             )
 567             log_string(
 568                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 569             )
 570
 571             input = self.test_input[:32]
 572             result = input.clone()
 573             ar_mask = result.new_zeros(result.size())
 574             ar_mask[:, self.height * self.width :] = 1
 575             result *= 1 - ar_mask
 576             x, order = shuffle(result, self.height * self.width)
 577             masked_inplace_autoregression(
 578                 model, self.batch_size, x, ar_mask, order=order
 579             )
 580             result = reorder(x, order, reverse=True)
 581
 582             mazes, paths = self.seq2map(input)
 583             _, predicted_paths = self.seq2map(result)
 584             filename = f"result_{n_epoch:04d}.png"
 585             maze.save_image(
 586                 os.path.join(args.result_dir, filename),
 587                 mazes=mazes,
 588                 target_paths=paths,
 589                 predicted_paths=predicted_paths,
 590                 path_correct=maze.path_correctness(mazes, predicted_paths),
 591             )
 592             log_string(f"wrote {filename}")
 593
 594             model.train(t)
 595
 596
 597 ######################################################################
 598
 599 log_string(f"device {device}")
 600
 601
 602 task = TaskMaze(
 603     nb_train_samples=args.nb_train_samples,
 604     nb_test_samples=args.nb_test_samples,
 605     batch_size=args.batch_size,
 606     height=args.maze_height,
 607     width=args.maze_width,
 608     nb_walls=args.maze_nb_walls,
 609     device=device,
 610 )
 611
 612
 613 vocabulary_size = task.vocabulary_size()
 614
 615 log_string(f"vocabulary_size {vocabulary_size}")
 616
 617 ##############################
 618
 619
 620 def noncausal_prompt_amm_generator(d):
 621     q = torch.arange(d)[:, None]
 622     k = torch.arange(d)[None, :]
 623     s = args.maze_height * args.maze_width
 624     return torch.logical_and(q < k, torch.logical_or(q >= s, k >= s))
 625     # return q < k
 626
 627
 628 def noncausal_prompt_oneshot_amm_generator(d):
 629     q = torch.arange(d)[:, None]
 630     k = torch.arange(d)[None, :]
 631     s = args.maze_height * args.maze_width
 632     return k >= s
 633     # return q < k
 634
 635
 636 if args.oneshot:
 637     amm_generator = noncausal_prompt_oneshot_amm_generator
 638 elif args.noncausal_prompt:
 639     amm_generator = noncausal_prompt_amm_generator
 640 else:
 641     amm_generator = None
 642
 643 model = mygpt.MyGPT(
 644     vocabulary_size=vocabulary_size,
 645     dim_model=args.dim_model,
 646     dim_keys=args.dim_keys,
 647     dim_hidden=args.dim_hidden,
 648     nb_heads=args.nb_heads,
 649     nb_blocks=args.nb_blocks,
 650     causal=True,
 651     dropout=args.dropout,
 652     amm_generator=amm_generator,
 653 )
 654
 655 model.to(device)
 656
 657 nb_parameters = sum(p.numel() for p in model.parameters())
 658 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 659
 660 ######################################################################
 661
 662 if args.learning_rate_schedule == "auto":
 663     learning_rate_scheduler = AutoScheduler(args.learning_rate)
 664
 665 elif args.learning_rate_schedule == "cos":
 666     schedule = {}
 667     for n_epoch in range(args.nb_epochs):
 668         u = n_epoch / args.nb_epochs * math.pi
 669         schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 670     learning_rate_scheduler = StepWiseScheduler(schedule)
 671     log_string(f"learning_rate_schedule {schedule}")
 672
 673 else:
 674     u = {
 675         int(k): float(v)
 676         for k, v in [
 677             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 678         ]
 679     }
 680
 681     schedule = {}
 682     learning_rate = args.learning_rate
 683     for n_epoch in range(args.nb_epochs):
 684         if n_epoch in u:
 685             learning_rate = u[n_epoch]
 686         schedule[n_epoch] = learning_rate
 687     learning_rate_scheduler = StepWiseScheduler(schedule)
 688     log_string(f"learning_rate_schedule {schedule}")
 689
 690 ######################################################################
 691
 692 nb_epochs_finished = 0
 693
 694 if args.no_checkpoint:
 695     log_string(f"not trying to load checkpoint.")
 696
 697 else:
 698     try:
 699         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 700         checkpoint = torch.load(checkpoint_name)
 701         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 702         model.load_state_dict(checkpoint["model_state"])
 703         learning_rate_scheduler.set_state(checkpoint["learning_rate_scheduler_state"])
 704         torch.set_rng_state(checkpoint["rng_state"])
 705         if torch.cuda.is_available():
 706             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 707
 708         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 709
 710     except FileNotFoundError:
 711         log_string("starting from scratch.")
 712
 713     # except:
 714     # log_string("error when loading the checkpoint.")
 715     # exit(1)
 716
 717 ######################################################################
 718
 719 if args.oneshot:
 720     oneshot(model, learning_rate_scheduler, task)
 721     exit(0)
 722
 723 ######################################################################
 724
 725 token_count = 0
 726 for input in task.batches(split="train"):
 727     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 728 token_probas = token_count / token_count.sum()
 729 entropy = -torch.xlogy(token_probas, token_probas).sum()
 730 train_set_perplexity = math.exp(entropy)
 731
 732 ##############################
 733
 734 if nb_epochs_finished >= args.nb_epochs:
 735     n_epoch = nb_epochs_finished
 736     train_perplexity = compute_perplexity(
 737         model, task, prompt_len=task.height * task.width, split="train"
 738     )
 739     test_perplexity = compute_perplexity(
 740         model, task, prompt_len=task.height * task.width, split="test"
 741     )
 742
 743     log_string(
 744         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 745     )
 746
 747     task.produce_results(n_epoch, model)
 748
 749 ##############################
 750
 751 learning_rate_scheduler.reset()
 752
 753 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 754     learning_rate = learning_rate_scheduler.get_learning_rate()
 755     log_string(f"learning_rate {n_epoch} {learning_rate}")
 756
 757     if args.optim == "sgd":
 758         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 759     elif args.optim == "adam":
 760         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 761     elif args.optim == "adamw":
 762         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 763     else:
 764         raise ValueError(f"{args.optim=}")
 765
 766     model.train()
 767
 768     nb_train_samples, acc_train_loss = 0, 0.0
 769
 770     for input in task.batches(split="train"):
 771         input = input.to(device)
 772         output = eval_mygpt(model, input, prompt_len=task.height * task.width)
 773         if args.noncausal_prompt:
 774             d = input.size(1) // 2
 775             loss = F.cross_entropy(output[:, d:].transpose(1, 2), input[:, d:])
 776         else:
 777             loss = F.cross_entropy(output.transpose(1, 2), input)
 778         acc_train_loss += loss.item() * input.size(0)
 779         nb_train_samples += input.size(0)
 780
 781         optimizer.zero_grad()
 782         loss.backward()
 783         optimizer.step()
 784
 785     learning_rate_scheduler.update(n_epoch + 1, acc_train_loss)
 786
 787     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 788     test_perplexity = compute_perplexity(
 789         model, task, prompt_len=task.height * task.width, split="test"
 790     )
 791
 792     log_string(
 793         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 794     )
 795
 796     task.produce_results(n_epoch, model)
 797
 798     checkpoint = {
 799         "nb_epochs_finished": n_epoch + 1,
 800         "model_state": model.state_dict(),
 801         "learning_rate_scheduler_state": learning_rate_scheduler.get_state(),
 802         "rng_state": torch.get_rng_state(),
 803     }
 804
 805     if torch.cuda.is_available():
 806         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 807
 808     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 809     torch.save(checkpoint, checkpoint_name)
 810     log_string(f"saved checkpoint {checkpoint_name}")
 811
 812 ######################################################################