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.get_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 get_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         print(f"{state=}")
 359         for k, v in state.items():
 360             setattr(self, k, v)
 361
 362
 363 class StepWiseScheduler(LearningRateScheduler):
 364     def __init__(self, schedule):
 365         self.nb_finished_epochs = 0
 366         self.schedule = schedule
 367
 368     def get_learning_rate(self):
 369         return self.schedule[self.nb_finished_epochs]
 370
 371     def update(self, nb_finished_epochs, loss):
 372         self.nb_finished_epochs = nb_finished_epochs
 373
 374     def reset(self):
 375         self.nb_finished_epochs = 0
 376
 377     def get_state(self):
 378         return {"nb_finished_epochs": self.nb_finished_epochs}
 379
 380
 381 class AutoScheduler(LearningRateScheduler):
 382     def __init__(self, learning_rate_init, growth=1.0, degrowth=0.2):
 383         self.learning_rate_init = learning_rate_init
 384         self.learning_rate = learning_rate_init
 385         self.growth = growth
 386         self.degrowth = degrowth
 387         self.pred_loss = None
 388
 389     def get_learning_rate(self):
 390         return self.learning_rate
 391
 392     def update(self, nb_finished_epochs, loss):
 393         if self.pred_loss is not None:
 394             if loss >= self.pred_loss:
 395                 self.learning_rate *= self.degrowth
 396             else:
 397                 self.learning_rate *= self.growth
 398         self.pred_loss = loss
 399
 400     def reset(self):
 401         self.learning_rate = self.learning_rate_init
 402
 403     def get_state(self):
 404         return {
 405             "learning_rate_init": self.learning_rate_init,
 406             "pred_loss": self.pred_loss,
 407         }
 408
 409
 410 ######################################################################
 411
 412
 413 class Task:
 414     def batches(self, split="train", nb_to_use=-1, desc=None):
 415         pass
 416
 417     def vocabulary_size(self):
 418         pass
 419
 420     def produce_results(self, n_epoch, model):
 421         pass
 422
 423
 424 ######################################################################
 425
 426 import maze
 427
 428
 429 class TaskMaze(Task):
 430     def map2seq(self, *m):
 431         return torch.cat([x.flatten(1) for x in m], 1)
 432
 433     def seq2map(self, s):
 434         s = s.reshape(s.size(0), -1, self.height, self.width)
 435         return (s[:, k] for k in range(s.size(1)))
 436
 437     def __init__(
 438         self,
 439         nb_train_samples,
 440         nb_test_samples,
 441         batch_size,
 442         height,
 443         width,
 444         nb_walls,
 445         device=torch.device("cpu"),
 446     ):
 447         self.batch_size = batch_size
 448         self.height = height
 449         self.width = width
 450         self.device = device
 451
 452         train_mazes, train_paths, train_policies = maze.create_maze_data(
 453             nb_train_samples,
 454             height=height,
 455             width=width,
 456             nb_walls=nb_walls,
 457             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 458         )
 459         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 460         self.train_policies = train_policies.flatten(-2).to(device)
 461
 462         test_mazes, test_paths, test_policies = maze.create_maze_data(
 463             nb_test_samples,
 464             height=height,
 465             width=width,
 466             nb_walls=nb_walls,
 467             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 468         )
 469         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 470         self.test_policies = test_policies.flatten(-2).to(device)
 471
 472         self.nb_codes = self.train_input.max() + 1
 473
 474     def batches(self, split="train", nb_to_use=-1, desc=None):
 475         assert split in {"train", "test"}
 476         input = self.train_input if split == "train" else self.test_input
 477         if nb_to_use > 0:
 478             input = input[:nb_to_use]
 479         if desc is None:
 480             desc = f"epoch-{split}"
 481         for batch in tqdm.tqdm(
 482             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 483         ):
 484             yield batch
 485
 486     def policy_batches(self, split="train", nb_to_use=-1, desc=None):
 487         assert split in {"train", "test"}
 488         input = self.train_input if split == "train" else self.test_input
 489         policies = self.train_policies if split == "train" else self.test_policies
 490         input = input[:, : self.height * self.width]
 491         policies = policies * (input != maze.v_wall)[:, None]
 492
 493         if nb_to_use > 0:
 494             input = input[:nb_to_use]
 495             policies = policies[:nb_to_use]
 496
 497         if desc is None:
 498             desc = f"epoch-{split}"
 499         for batch in tqdm.tqdm(
 500             zip(input.split(self.batch_size), policies.split(self.batch_size)),
 501             dynamic_ncols=True,
 502             desc=desc,
 503         ):
 504             yield batch
 505
 506     def vocabulary_size(self):
 507         return self.nb_codes
 508
 509     def compute_error(self, model, split="train", nb_to_use=-1):
 510         nb_total, nb_correct = 0, 0
 511         for input in task.batches(split, nb_to_use):
 512             result = input.clone()
 513             ar_mask = result.new_zeros(result.size())
 514             ar_mask[:, self.height * self.width :] = 1
 515             result *= 1 - ar_mask
 516             x, order = shuffle(result, self.height * self.width)
 517             masked_inplace_autoregression(
 518                 model, self.batch_size, x, ar_mask, order=order
 519             )
 520             result = reorder(x, order, reverse=True)
 521             mazes, paths = self.seq2map(result)
 522             nb_correct += maze.path_correctness(mazes, paths).long().sum()
 523             nb_total += mazes.size(0)
 524
 525         return nb_total, nb_correct
 526
 527     def produce_results(self, n_epoch, model):
 528         with torch.autograd.no_grad():
 529             t = model.training
 530             model.eval()
 531
 532             train_nb_total, train_nb_correct = self.compute_error(
 533                 model, "train", nb_to_use=1000
 534             )
 535             log_string(
 536                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 537             )
 538
 539             test_nb_total, test_nb_correct = self.compute_error(
 540                 model, "test", nb_to_use=1000
 541             )
 542             log_string(
 543                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 544             )
 545
 546             input = self.test_input[:32]
 547             result = input.clone()
 548             ar_mask = result.new_zeros(result.size())
 549             ar_mask[:, self.height * self.width :] = 1
 550             result *= 1 - ar_mask
 551             x, order = shuffle(result, self.height * self.width)
 552             masked_inplace_autoregression(
 553                 model, self.batch_size, x, ar_mask, order=order
 554             )
 555             result = reorder(x, order, reverse=True)
 556
 557             mazes, paths = self.seq2map(input)
 558             _, predicted_paths = self.seq2map(result)
 559             filename = f"result_{n_epoch:04d}.png"
 560             maze.save_image(
 561                 os.path.join(args.result_dir, filename),
 562                 mazes=mazes,
 563                 target_paths=paths,
 564                 predicted_paths=predicted_paths,
 565                 path_correct=maze.path_correctness(mazes, predicted_paths),
 566             )
 567             log_string(f"wrote {filename}")
 568
 569             model.train(t)
 570
 571
 572 ######################################################################
 573
 574 log_string(f"device {device}")
 575
 576
 577 task = TaskMaze(
 578     nb_train_samples=args.nb_train_samples,
 579     nb_test_samples=args.nb_test_samples,
 580     batch_size=args.batch_size,
 581     height=args.maze_height,
 582     width=args.maze_width,
 583     nb_walls=args.maze_nb_walls,
 584     device=device,
 585 )
 586
 587
 588 vocabulary_size = task.vocabulary_size()
 589
 590 log_string(f"vocabulary_size {vocabulary_size}")
 591
 592 ##############################
 593
 594
 595 def noncausal_prompt_amm_generator(d):
 596     q = torch.arange(d)[:, None]
 597     k = torch.arange(d)[None, :]
 598     s = args.maze_height * args.maze_width
 599     #    return torch.logical_and(q < k, torch.logical_or(q >= s, k >= s))
 600     return q < k
 601
 602
 603 amm_generator = None
 604
 605 if args.noncausal_prompt:
 606     amm_generator = noncausal_prompt_amm_generator
 607
 608 model = mygpt.MyGPT(
 609     vocabulary_size=vocabulary_size,
 610     dim_model=args.dim_model,
 611     dim_keys=args.dim_keys,
 612     dim_hidden=args.dim_hidden,
 613     nb_heads=args.nb_heads,
 614     nb_blocks=args.nb_blocks,
 615     causal=True,
 616     dropout=args.dropout,
 617     amm_generator=amm_generator,
 618 )
 619
 620 model.to(device)
 621
 622 nb_parameters = sum(p.numel() for p in model.parameters())
 623 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 624
 625 ######################################################################
 626
 627 if args.learning_rate_schedule == "auto":
 628     learning_rate_scheduler = AutoScheduler(args.learning_rate)
 629
 630 elif args.learning_rate_schedule == "cos":
 631     schedule = {}
 632     for n_epoch in range(args.nb_epochs):
 633         u = n_epoch / args.nb_epochs * math.pi
 634         schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 635     learning_rate_scheduler = StepWiseScheduler(schedule)
 636     log_string(f"learning_rate_schedule {schedule}")
 637
 638 else:
 639     u = {
 640         int(k): float(v)
 641         for k, v in [
 642             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 643         ]
 644     }
 645
 646     schedule = {}
 647     learning_rate = args.learning_rate
 648     for n_epoch in range(args.nb_epochs):
 649         if n_epoch in u:
 650             learning_rate = u[n_epoch]
 651         schedule[n_epoch] = learning_rate
 652     learning_rate_scheduler = StepWiseScheduler(schedule)
 653     log_string(f"learning_rate_schedule {schedule}")
 654
 655 ######################################################################
 656
 657 nb_epochs_finished = 0
 658
 659 if args.no_checkpoint:
 660     log_string(f"not trying to load checkpoint.")
 661
 662 else:
 663     try:
 664         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 665         checkpoint = torch.load(checkpoint_name)
 666         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 667         model.load_state_dict(checkpoint["model_state"])
 668         learning_rate_scheduler.set_state(checkpoint["learning_rate_scheduler_state"])
 669         torch.set_rng_state(checkpoint["rng_state"])
 670         if torch.cuda.is_available():
 671             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 672
 673         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 674
 675     except FileNotFoundError:
 676         log_string("starting from scratch.")
 677
 678     # except:
 679     # log_string("error when loading the checkpoint.")
 680     # exit(1)
 681
 682 ######################################################################
 683
 684 token_count = 0
 685 for input in task.batches(split="train"):
 686     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 687 token_probas = token_count / token_count.sum()
 688 entropy = -torch.xlogy(token_probas, token_probas).sum()
 689 train_set_perplexity = math.exp(entropy)
 690
 691 ##############################
 692
 693 if nb_epochs_finished >= args.nb_epochs:
 694     n_epoch = nb_epochs_finished
 695     train_perplexity = compute_perplexity(
 696         model, task, prompt_len=task.height * task.width, split="train"
 697     )
 698     test_perplexity = compute_perplexity(
 699         model, task, prompt_len=task.height * task.width, split="test"
 700     )
 701
 702     log_string(
 703         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 704     )
 705
 706     task.produce_results(n_epoch, model)
 707
 708 ##############################
 709
 710 learning_rate_scheduler.reset()
 711
 712 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 713     learning_rate = learning_rate_scheduler.get_learning_rate()
 714
 715     log_string(f"learning_rate {learning_rate}")
 716
 717     if args.optim == "sgd":
 718         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 719     elif args.optim == "adam":
 720         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 721     elif args.optim == "adamw":
 722         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 723     else:
 724         raise ValueError(f"{args.optim=}")
 725
 726     model.train()
 727
 728     nb_train_samples, acc_train_loss = 0, 0.0
 729
 730     for input in task.batches(split="train"):
 731         input = input.to(device)
 732         output = eval_mygpt(model, input, prompt_len=task.height * task.width)
 733         if args.noncausal_prompt:
 734             d = input.size(1) // 2
 735             loss = F.cross_entropy(output[:, d:].transpose(1, 2), input[:, d:])
 736         else:
 737             loss = F.cross_entropy(output.transpose(1, 2), input)
 738         acc_train_loss += loss.item() * input.size(0)
 739         nb_train_samples += input.size(0)
 740
 741         optimizer.zero_grad()
 742         loss.backward()
 743         optimizer.step()
 744
 745     learning_rate_scheduler.update(n_epoch + 1, acc_train_loss)
 746
 747     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 748     test_perplexity = compute_perplexity(
 749         model, task, prompt_len=task.height * task.width, split="test"
 750     )
 751
 752     log_string(
 753         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 754     )
 755
 756     task.produce_results(n_epoch, model)
 757
 758     checkpoint = {
 759         "nb_epochs_finished": n_epoch + 1,
 760         "model_state": model.state_dict(),
 761         "learning_rate_scheduler_state": learning_rate_scheduler.get_state(),
 762         "rng_state": torch.get_rng_state(),
 763     }
 764
 765     if torch.cuda.is_available():
 766         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 767
 768     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 769     torch.save(checkpoint, checkpoint_name)
 770     log_string(f"saved checkpoint {checkpoint_name}")
 771
 772 ######################################################################
 773
 774 if args.oneshot:
 775     oneshot(model, learning_rate_scheduler, task)
 776
 777 ######################################################################