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(gpt, learning_rate_scheduler, task):
 237     t = gpt.training
 238     gpt.eval()
 239
 240     if args.oneshot_input == "head":
 241         dim_in = args.dim_model
 242     elif args.oneshot_input == "deep":
 243         dim_in = args.dim_model * args.nb_blocks * 2
 244     else:
 245         raise ValueError(f"{args.oneshot_input=}")
 246
 247     if args.oneshot_output == "policy":
 248         dim_out = 4
 249         compute_loss = oneshot_policy_loss
 250     elif args.oneshot_output == "trace":
 251         dim_out = 1
 252         compute_loss = oneshot_trace_loss
 253     else:
 254         raise ValueError(f"{args.oneshot_output=}")
 255
 256     model = nn.Sequential(
 257         nn.Linear(dim_in, args.dim_model),
 258         nn.ReLU(),
 259         nn.Linear(args.dim_model, args.dim_model),
 260         nn.ReLU(),
 261         nn.Linear(args.dim_model, dim_out),
 262     ).to(device)
 263
 264     learning_rate_scheduler.reset()
 265
 266     for n_epoch in range(args.nb_epochs):
 267         learning_rate = learning_rate_scheduler.learning_rate()
 268         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 269
 270         acc_train_loss, nb_train_samples = 0, 0
 271         for mazes, policies in task.policy_batches(split="train"):
 272             output_gpt = eval_mygpt(
 273                 gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
 274             )
 275             output = model(output_gpt)
 276
 277             loss = compute_loss(mazes, output, policies, task.height, task.width)
 278             acc_train_loss += loss.item() * mazes.size(0)
 279             nb_train_samples += mazes.size(0)
 280
 281             optimizer.zero_grad()
 282             loss.backward()
 283             optimizer.step()
 284
 285         learning_rate_scheduler.update(n_epoch + 1, acc_train_loss)
 286
 287         acc_test_loss, nb_test_samples = 0, 0
 288         for mazes, policies in task.policy_batches(split="test"):
 289             output_gpt = eval_mygpt(
 290                 gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
 291             )
 292             output = model(output_gpt)
 293             loss = compute_loss(mazes, output, policies, task.height, task.width)
 294             acc_test_loss += loss.item() * mazes.size(0)
 295             nb_test_samples += mazes.size(0)
 296
 297         log_string(
 298             f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
 299         )
 300
 301         # -------------------
 302         mazes = task.test_input[:32, : task.height * task.width]
 303         policies = task.test_policies[:32]
 304         output_gpt = eval_mygpt(
 305             gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
 306         )
 307         output = model(output_gpt)
 308         if args.oneshot_output == "policy":
 309             targets = policies.permute(0, 2, 1)
 310             scores = (
 311                 (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
 312             ).float()
 313         elif args.oneshot_output == "trace":
 314             targets = maze.stationary_densities(
 315                 mazes.view(-1, task.height, task.width),
 316                 policies.view(-1, 4, task.height, task.width),
 317             ).flatten(-2)
 318             scores = output
 319         else:
 320             raise ValueError(f"{args.oneshot_output=}")
 321
 322         scores = scores.reshape(-1, task.height, task.width)
 323         mazes = mazes.reshape(-1, task.height, task.width)
 324         targets = targets.reshape(-1, task.height, task.width)
 325         filename = (
 326             f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png"
 327         )
 328         maze.save_image(
 329             os.path.join(args.result_dir, filename),
 330             mazes=mazes,
 331             score_paths=scores,
 332             score_truth=targets,
 333         )
 334         log_string(f"wrote {filename}")
 335
 336         # -------------------
 337
 338     gpt.train(t)
 339
 340
 341 ######################################################################
 342
 343
 344 class LearningRateScheduler:
 345     def learning_rate(self):
 346         pass
 347
 348     def update(self, nb_finished_epochs, loss):
 349         pass
 350
 351     def reset(self):
 352         pass
 353
 354     def get_state(self):
 355         return vars(self)
 356
 357     def set_state(self, state):
 358         for k, v in state.item():
 359             setattr(self, k, v)
 360
 361
 362 class StepWiseScheduler(LearningRateScheduler):
 363     def __init__(self, schedule):
 364         self.nb_finished_epochs = 0
 365         self.schedule = schedule
 366
 367     def learning_rate(self):
 368         return self.schedule[self.nb_finished_epochs]
 369
 370     def reset(self):
 371         self.nb_finished_epochs = 0
 372
 373
 374 ######################################################################
 375
 376
 377 class Task:
 378     def batches(self, split="train", nb_to_use=-1, desc=None):
 379         pass
 380
 381     def vocabulary_size(self):
 382         pass
 383
 384     def produce_results(self, n_epoch, model):
 385         pass
 386
 387
 388 ######################################################################
 389
 390 import maze
 391
 392
 393 class TaskMaze(Task):
 394     def map2seq(self, *m):
 395         return torch.cat([x.flatten(1) for x in m], 1)
 396
 397     def seq2map(self, s):
 398         s = s.reshape(s.size(0), -1, self.height, self.width)
 399         return (s[:, k] for k in range(s.size(1)))
 400
 401     def __init__(
 402         self,
 403         nb_train_samples,
 404         nb_test_samples,
 405         batch_size,
 406         height,
 407         width,
 408         nb_walls,
 409         device=torch.device("cpu"),
 410     ):
 411         self.batch_size = batch_size
 412         self.height = height
 413         self.width = width
 414         self.device = device
 415
 416         train_mazes, train_paths, train_policies = maze.create_maze_data(
 417             nb_train_samples,
 418             height=height,
 419             width=width,
 420             nb_walls=nb_walls,
 421             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 422         )
 423         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 424         self.train_policies = train_policies.flatten(-2).to(device)
 425
 426         test_mazes, test_paths, test_policies = maze.create_maze_data(
 427             nb_test_samples,
 428             height=height,
 429             width=width,
 430             nb_walls=nb_walls,
 431             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 432         )
 433         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 434         self.test_policies = test_policies.flatten(-2).to(device)
 435
 436         self.nb_codes = self.train_input.max() + 1
 437
 438     def batches(self, split="train", nb_to_use=-1, desc=None):
 439         assert split in {"train", "test"}
 440         input = self.train_input if split == "train" else self.test_input
 441         if nb_to_use > 0:
 442             input = input[:nb_to_use]
 443         if desc is None:
 444             desc = f"epoch-{split}"
 445         for batch in tqdm.tqdm(
 446             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 447         ):
 448             yield batch
 449
 450     def policy_batches(self, split="train", nb_to_use=-1, desc=None):
 451         assert split in {"train", "test"}
 452         input = self.train_input if split == "train" else self.test_input
 453         policies = self.train_policies if split == "train" else self.test_policies
 454         input = input[:, : self.height * self.width]
 455         policies = policies * (input != maze.v_wall)[:, None]
 456
 457         if nb_to_use > 0:
 458             input = input[:nb_to_use]
 459             policies = policies[:nb_to_use]
 460
 461         if desc is None:
 462             desc = f"epoch-{split}"
 463         for batch in tqdm.tqdm(
 464             zip(input.split(self.batch_size), policies.split(self.batch_size)),
 465             dynamic_ncols=True,
 466             desc=desc,
 467         ):
 468             yield batch
 469
 470     def vocabulary_size(self):
 471         return self.nb_codes
 472
 473     def compute_error(self, model, split="train", nb_to_use=-1):
 474         nb_total, nb_correct = 0, 0
 475         for input in task.batches(split, nb_to_use):
 476             result = input.clone()
 477             ar_mask = result.new_zeros(result.size())
 478             ar_mask[:, self.height * self.width :] = 1
 479             result *= 1 - ar_mask
 480             x, order = shuffle(result, self.height * self.width)
 481             masked_inplace_autoregression(
 482                 model, self.batch_size, x, ar_mask, order=order
 483             )
 484             result = reorder(x, order, reverse=True)
 485             mazes, paths = self.seq2map(result)
 486             nb_correct += maze.path_correctness(mazes, paths).long().sum()
 487             nb_total += mazes.size(0)
 488
 489         return nb_total, nb_correct
 490
 491     def produce_results(self, n_epoch, model):
 492         with torch.autograd.no_grad():
 493             t = model.training
 494             model.eval()
 495
 496             train_nb_total, train_nb_correct = self.compute_error(
 497                 model, "train", nb_to_use=1000
 498             )
 499             log_string(
 500                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 501             )
 502
 503             test_nb_total, test_nb_correct = self.compute_error(
 504                 model, "test", nb_to_use=1000
 505             )
 506             log_string(
 507                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 508             )
 509
 510             input = self.test_input[:32]
 511             result = input.clone()
 512             ar_mask = result.new_zeros(result.size())
 513             ar_mask[:, self.height * self.width :] = 1
 514             result *= 1 - ar_mask
 515             x, order = shuffle(result, self.height * self.width)
 516             masked_inplace_autoregression(
 517                 model, self.batch_size, x, ar_mask, order=order
 518             )
 519             result = reorder(x, order, reverse=True)
 520
 521             mazes, paths = self.seq2map(input)
 522             _, predicted_paths = self.seq2map(result)
 523             filename = f"result_{n_epoch:04d}.png"
 524             maze.save_image(
 525                 os.path.join(args.result_dir, filename),
 526                 mazes=mazes,
 527                 target_paths=paths,
 528                 predicted_paths=predicted_paths,
 529                 path_correct=maze.path_correctness(mazes, predicted_paths),
 530             )
 531             log_string(f"wrote {filename}")
 532
 533             model.train(t)
 534
 535
 536 ######################################################################
 537
 538 log_string(f"device {device}")
 539
 540
 541 task = TaskMaze(
 542     nb_train_samples=args.nb_train_samples,
 543     nb_test_samples=args.nb_test_samples,
 544     batch_size=args.batch_size,
 545     height=args.maze_height,
 546     width=args.maze_width,
 547     nb_walls=args.maze_nb_walls,
 548     device=device,
 549 )
 550
 551
 552 vocabulary_size = task.vocabulary_size()
 553
 554 log_string(f"vocabulary_size {vocabulary_size}")
 555
 556 ##############################
 557
 558
 559 def noncausal_prompt_amm_generator(d):
 560     q = torch.arange(d)[:, None]
 561     k = torch.arange(d)[None, :]
 562     s = args.maze_height * args.maze_width
 563     #    return torch.logical_and(q < k, torch.logical_or(q >= s, k >= s))
 564     return q < k
 565
 566
 567 amm_generator = None
 568
 569 if args.noncausal_prompt:
 570     amm_generator = noncausal_prompt_amm_generator
 571
 572 model = mygpt.MyGPT(
 573     vocabulary_size=vocabulary_size,
 574     dim_model=args.dim_model,
 575     dim_keys=args.dim_keys,
 576     dim_hidden=args.dim_hidden,
 577     nb_heads=args.nb_heads,
 578     nb_blocks=args.nb_blocks,
 579     causal=True,
 580     dropout=args.dropout,
 581     amm_generator=amm_generator,
 582 )
 583
 584 model.to(device)
 585
 586 nb_parameters = sum(p.numel() for p in model.parameters())
 587 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 588
 589 ######################################################################
 590
 591 if args.learning_rate_schedule == "auto":
 592     pass
 593
 594 elif args.learning_rate_schedule == "cos":
 595     schedule = {}
 596     for n_epoch in range(args.nb_epochs):
 597         u = n_epoch / args.nb_epochs * math.pi
 598         schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 599     learning_rate_scheduler = StepWiseScheduler(schedule)
 600     log_string(f"learning_rate_schedule {schedule}")
 601
 602 else:
 603     u = {
 604         int(k): float(v)
 605         for k, v in [
 606             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 607         ]
 608     }
 609
 610     schedule = {}
 611     learning_rate = args.learning_rate
 612     for n_epoch in range(args.nb_epochs):
 613         if n_epoch in u:
 614             learning_rate = u[n_epoch]
 615         schedule[n_epoch] = learning_rate
 616     learning_rate_scheduler = StepWiseScheduler(schedule)
 617     log_string(f"learning_rate_schedule {schedule}")
 618
 619 ######################################################################
 620
 621 nb_epochs_finished = 0
 622
 623 if args.no_checkpoint:
 624     log_string(f"not trying to load checkpoint.")
 625
 626 else:
 627     try:
 628         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 629         checkpoint = torch.load(checkpoint_name)
 630         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 631         model.load_state_dict(checkpoint["model_state"])
 632         torch.set_rng_state(checkpoint["rng_state"])
 633         if torch.cuda.is_available():
 634             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 635
 636         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 637
 638     except FileNotFoundError:
 639         log_string("starting from scratch.")
 640
 641     except:
 642         log_string("error when loading the checkpoint.")
 643         exit(1)
 644
 645 ######################################################################
 646
 647 token_count = 0
 648 for input in task.batches(split="train"):
 649     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 650 token_probas = token_count / token_count.sum()
 651 entropy = -torch.xlogy(token_probas, token_probas).sum()
 652 train_set_perplexity = math.exp(entropy)
 653
 654 ##############################
 655
 656 if nb_epochs_finished >= args.nb_epochs:
 657     n_epoch = nb_epochs_finished
 658     train_perplexity = compute_perplexity(
 659         model, task, prompt_len=task.height * task.width, split="train"
 660     )
 661     test_perplexity = compute_perplexity(
 662         model, task, prompt_len=task.height * task.width, split="test"
 663     )
 664
 665     log_string(
 666         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 667     )
 668
 669     task.produce_results(n_epoch, model)
 670
 671 ##############################
 672
 673 learning_rate_scheduler.reset()
 674
 675 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 676     learning_rate = learning_rate_scheduler.learning_rate()
 677
 678     log_string(f"learning_rate {learning_rate}")
 679
 680     if args.optim == "sgd":
 681         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 682     elif args.optim == "adam":
 683         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 684     elif args.optim == "adamw":
 685         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 686     else:
 687         raise ValueError(f"{args.optim=}")
 688
 689     model.train()
 690
 691     nb_train_samples, acc_train_loss = 0, 0.0
 692
 693     for input in task.batches(split="train"):
 694         input = input.to(device)
 695         output = eval_mygpt(model, input, prompt_len=task.height * task.width)
 696         if args.noncausal_prompt:
 697             d = input.size(1) // 2
 698             loss = F.cross_entropy(output[:, d:].transpose(1, 2), input[:, d:])
 699         else:
 700             loss = F.cross_entropy(output.transpose(1, 2), input)
 701         acc_train_loss += loss.item() * input.size(0)
 702         nb_train_samples += input.size(0)
 703
 704         optimizer.zero_grad()
 705         loss.backward()
 706         optimizer.step()
 707
 708     learning_rate_scheduler.update(n_epoch + 1, acc_train_loss)
 709
 710     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 711     test_perplexity = compute_perplexity(
 712         model, task, prompt_len=task.height * task.width, split="test"
 713     )
 714
 715     log_string(
 716         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 717     )
 718
 719     task.produce_results(n_epoch, model)
 720
 721     checkpoint = {
 722         "nb_epochs_finished": n_epoch + 1,
 723         "model_state": model.state_dict(),
 724         "rng_state": torch.get_rng_state(),
 725     }
 726
 727     if torch.cuda.is_available():
 728         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 729
 730     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 731     torch.save(checkpoint, checkpoint_name)
 732     log_string(f"saved checkpoint {checkpoint_name}")
 733
 734 ######################################################################
 735
 736 if args.oneshot:
 737     oneshot(model, learning_rate_scheduler, task)
 738
 739 ######################################################################