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         log_string(f"learning_rate {n_epoch} {learning_rate}")
 269
 270         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 271
 272         acc_train_loss, nb_train_samples = 0, 0
 273         for mazes, policies in task.policy_batches(split="train"):
 274             output_gpt = eval_mygpt(
 275                 gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
 276             )
 277             output = model(output_gpt)
 278
 279             loss = compute_loss(mazes, output, policies, task.height, task.width)
 280             acc_train_loss += loss.item() * mazes.size(0)
 281             nb_train_samples += mazes.size(0)
 282
 283             optimizer.zero_grad()
 284             loss.backward()
 285             optimizer.step()
 286
 287         learning_rate_scheduler.update(n_epoch + 1, acc_train_loss)
 288
 289         acc_test_loss, nb_test_samples = 0, 0
 290         for mazes, policies in task.policy_batches(split="test"):
 291             output_gpt = eval_mygpt(
 292                 gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
 293             )
 294             output = model(output_gpt)
 295             loss = compute_loss(mazes, output, policies, task.height, task.width)
 296             acc_test_loss += loss.item() * mazes.size(0)
 297             nb_test_samples += mazes.size(0)
 298
 299         log_string(
 300             f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
 301         )
 302
 303         # -------------------
 304         mazes = task.test_input[:32, : task.height * task.width]
 305         policies = task.test_policies[:32]
 306         output_gpt = eval_mygpt(
 307             gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
 308         )
 309         output = model(output_gpt)
 310         if args.oneshot_output == "policy":
 311             targets = policies.permute(0, 2, 1)
 312             scores = (
 313                 (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
 314             ).float()
 315         elif args.oneshot_output == "trace":
 316             targets = maze.stationary_densities(
 317                 mazes.view(-1, task.height, task.width),
 318                 policies.view(-1, 4, task.height, task.width),
 319             ).flatten(-2)
 320             scores = output
 321         else:
 322             raise ValueError(f"{args.oneshot_output=}")
 323
 324         scores = scores.reshape(-1, task.height, task.width)
 325         mazes = mazes.reshape(-1, task.height, task.width)
 326         targets = targets.reshape(-1, task.height, task.width)
 327         filename = (
 328             f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png"
 329         )
 330         maze.save_image(
 331             os.path.join(args.result_dir, filename),
 332             mazes=mazes,
 333             score_paths=scores,
 334             score_truth=targets,
 335         )
 336         log_string(f"wrote {filename}")
 337
 338         # -------------------
 339
 340     gpt.train(t)
 341
 342
 343 ######################################################################
 344
 345
 346 class LearningRateScheduler:
 347     def get_learning_rate(self):
 348         pass
 349
 350     def update(self, nb_finished_epochs, loss):
 351         pass
 352
 353     def reset(self):
 354         pass
 355
 356     def get_state(self):
 357         return vars(self)
 358
 359     def set_state(self, state):
 360         print(f"{state=}")
 361         for k, v in state.items():
 362             setattr(self, k, v)
 363
 364
 365 class StepWiseScheduler(LearningRateScheduler):
 366     def __init__(self, schedule):
 367         self.nb_finished_epochs = 0
 368         self.schedule = schedule
 369
 370     def get_learning_rate(self):
 371         return self.schedule[self.nb_finished_epochs]
 372
 373     def update(self, nb_finished_epochs, loss):
 374         self.nb_finished_epochs = nb_finished_epochs
 375
 376     def reset(self):
 377         self.nb_finished_epochs = 0
 378
 379     def get_state(self):
 380         return {"nb_finished_epochs": self.nb_finished_epochs}
 381
 382
 383 class AutoScheduler(LearningRateScheduler):
 384     def __init__(self, learning_rate_init, growth=1.0, degrowth=0.2):
 385         self.learning_rate_init = learning_rate_init
 386         self.learning_rate = learning_rate_init
 387         self.growth = growth
 388         self.degrowth = degrowth
 389         self.pred_loss = None
 390
 391     def get_learning_rate(self):
 392         return self.learning_rate
 393
 394     def update(self, nb_finished_epochs, loss):
 395         if self.pred_loss is not None:
 396             if loss >= self.pred_loss:
 397                 self.learning_rate *= self.degrowth
 398             else:
 399                 self.learning_rate *= self.growth
 400         self.pred_loss = loss
 401
 402     def reset(self):
 403         self.learning_rate = self.learning_rate_init
 404
 405     def get_state(self):
 406         return {
 407             "learning_rate_init": self.learning_rate_init,
 408             "pred_loss": self.pred_loss,
 409         }
 410
 411
 412 ######################################################################
 413
 414
 415 class Task:
 416     def batches(self, split="train", nb_to_use=-1, desc=None):
 417         pass
 418
 419     def vocabulary_size(self):
 420         pass
 421
 422     def produce_results(self, n_epoch, model):
 423         pass
 424
 425
 426 ######################################################################
 427
 428 import maze
 429
 430
 431 class TaskMaze(Task):
 432     def map2seq(self, *m):
 433         return torch.cat([x.flatten(1) for x in m], 1)
 434
 435     def seq2map(self, s):
 436         s = s.reshape(s.size(0), -1, self.height, self.width)
 437         return (s[:, k] for k in range(s.size(1)))
 438
 439     def __init__(
 440         self,
 441         nb_train_samples,
 442         nb_test_samples,
 443         batch_size,
 444         height,
 445         width,
 446         nb_walls,
 447         device=torch.device("cpu"),
 448     ):
 449         self.batch_size = batch_size
 450         self.height = height
 451         self.width = width
 452         self.device = device
 453
 454         train_mazes, train_paths, train_policies = maze.create_maze_data(
 455             nb_train_samples,
 456             height=height,
 457             width=width,
 458             nb_walls=nb_walls,
 459             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 460         )
 461         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 462         self.train_policies = train_policies.flatten(-2).to(device)
 463
 464         test_mazes, test_paths, test_policies = maze.create_maze_data(
 465             nb_test_samples,
 466             height=height,
 467             width=width,
 468             nb_walls=nb_walls,
 469             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 470         )
 471         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 472         self.test_policies = test_policies.flatten(-2).to(device)
 473
 474         self.nb_codes = self.train_input.max() + 1
 475
 476     def batches(self, split="train", nb_to_use=-1, desc=None):
 477         assert split in {"train", "test"}
 478         input = self.train_input if split == "train" else self.test_input
 479         if nb_to_use > 0:
 480             input = input[:nb_to_use]
 481         if desc is None:
 482             desc = f"epoch-{split}"
 483         for batch in tqdm.tqdm(
 484             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 485         ):
 486             yield batch
 487
 488     def policy_batches(self, split="train", nb_to_use=-1, desc=None):
 489         assert split in {"train", "test"}
 490         input = self.train_input if split == "train" else self.test_input
 491         policies = self.train_policies if split == "train" else self.test_policies
 492         input = input[:, : self.height * self.width]
 493         policies = policies * (input != maze.v_wall)[:, None]
 494
 495         if nb_to_use > 0:
 496             input = input[:nb_to_use]
 497             policies = policies[:nb_to_use]
 498
 499         if desc is None:
 500             desc = f"epoch-{split}"
 501         for batch in tqdm.tqdm(
 502             zip(input.split(self.batch_size), policies.split(self.batch_size)),
 503             dynamic_ncols=True,
 504             desc=desc,
 505         ):
 506             yield batch
 507
 508     def vocabulary_size(self):
 509         return self.nb_codes
 510
 511     def compute_error(self, model, split="train", nb_to_use=-1):
 512         nb_total, nb_correct = 0, 0
 513         for input in task.batches(split, nb_to_use):
 514             result = input.clone()
 515             ar_mask = result.new_zeros(result.size())
 516             ar_mask[:, self.height * self.width :] = 1
 517             result *= 1 - ar_mask
 518             x, order = shuffle(result, self.height * self.width)
 519             masked_inplace_autoregression(
 520                 model, self.batch_size, x, ar_mask, order=order
 521             )
 522             result = reorder(x, order, reverse=True)
 523             mazes, paths = self.seq2map(result)
 524             nb_correct += maze.path_correctness(mazes, paths).long().sum()
 525             nb_total += mazes.size(0)
 526
 527         return nb_total, nb_correct
 528
 529     def produce_results(self, n_epoch, model):
 530         with torch.autograd.no_grad():
 531             t = model.training
 532             model.eval()
 533
 534             train_nb_total, train_nb_correct = self.compute_error(
 535                 model, "train", nb_to_use=1000
 536             )
 537             log_string(
 538                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 539             )
 540
 541             test_nb_total, test_nb_correct = self.compute_error(
 542                 model, "test", nb_to_use=1000
 543             )
 544             log_string(
 545                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 546             )
 547
 548             input = self.test_input[:32]
 549             result = input.clone()
 550             ar_mask = result.new_zeros(result.size())
 551             ar_mask[:, self.height * self.width :] = 1
 552             result *= 1 - ar_mask
 553             x, order = shuffle(result, self.height * self.width)
 554             masked_inplace_autoregression(
 555                 model, self.batch_size, x, ar_mask, order=order
 556             )
 557             result = reorder(x, order, reverse=True)
 558
 559             mazes, paths = self.seq2map(input)
 560             _, predicted_paths = self.seq2map(result)
 561             filename = f"result_{n_epoch:04d}.png"
 562             maze.save_image(
 563                 os.path.join(args.result_dir, filename),
 564                 mazes=mazes,
 565                 target_paths=paths,
 566                 predicted_paths=predicted_paths,
 567                 path_correct=maze.path_correctness(mazes, predicted_paths),
 568             )
 569             log_string(f"wrote {filename}")
 570
 571             model.train(t)
 572
 573
 574 ######################################################################
 575
 576 log_string(f"device {device}")
 577
 578
 579 task = TaskMaze(
 580     nb_train_samples=args.nb_train_samples,
 581     nb_test_samples=args.nb_test_samples,
 582     batch_size=args.batch_size,
 583     height=args.maze_height,
 584     width=args.maze_width,
 585     nb_walls=args.maze_nb_walls,
 586     device=device,
 587 )
 588
 589
 590 vocabulary_size = task.vocabulary_size()
 591
 592 log_string(f"vocabulary_size {vocabulary_size}")
 593
 594 ##############################
 595
 596
 597 def noncausal_prompt_amm_generator(d):
 598     q = torch.arange(d)[:, None]
 599     k = torch.arange(d)[None, :]
 600     s = args.maze_height * args.maze_width
 601     #    return torch.logical_and(q < k, torch.logical_or(q >= s, k >= s))
 602     return q < k
 603
 604
 605 amm_generator = None
 606
 607 if args.noncausal_prompt:
 608     amm_generator = noncausal_prompt_amm_generator
 609
 610 model = mygpt.MyGPT(
 611     vocabulary_size=vocabulary_size,
 612     dim_model=args.dim_model,
 613     dim_keys=args.dim_keys,
 614     dim_hidden=args.dim_hidden,
 615     nb_heads=args.nb_heads,
 616     nb_blocks=args.nb_blocks,
 617     causal=True,
 618     dropout=args.dropout,
 619     amm_generator=amm_generator,
 620 )
 621
 622 model.to(device)
 623
 624 nb_parameters = sum(p.numel() for p in model.parameters())
 625 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 626
 627 ######################################################################
 628
 629 if args.learning_rate_schedule == "auto":
 630     learning_rate_scheduler = AutoScheduler(args.learning_rate)
 631
 632 elif args.learning_rate_schedule == "cos":
 633     schedule = {}
 634     for n_epoch in range(args.nb_epochs):
 635         u = n_epoch / args.nb_epochs * math.pi
 636         schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 637     learning_rate_scheduler = StepWiseScheduler(schedule)
 638     log_string(f"learning_rate_schedule {schedule}")
 639
 640 else:
 641     u = {
 642         int(k): float(v)
 643         for k, v in [
 644             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 645         ]
 646     }
 647
 648     schedule = {}
 649     learning_rate = args.learning_rate
 650     for n_epoch in range(args.nb_epochs):
 651         if n_epoch in u:
 652             learning_rate = u[n_epoch]
 653         schedule[n_epoch] = learning_rate
 654     learning_rate_scheduler = StepWiseScheduler(schedule)
 655     log_string(f"learning_rate_schedule {schedule}")
 656
 657 ######################################################################
 658
 659 nb_epochs_finished = 0
 660
 661 if args.no_checkpoint:
 662     log_string(f"not trying to load checkpoint.")
 663
 664 else:
 665     try:
 666         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 667         checkpoint = torch.load(checkpoint_name)
 668         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 669         model.load_state_dict(checkpoint["model_state"])
 670         learning_rate_scheduler.set_state(checkpoint["learning_rate_scheduler_state"])
 671         torch.set_rng_state(checkpoint["rng_state"])
 672         if torch.cuda.is_available():
 673             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 674
 675         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 676
 677     except FileNotFoundError:
 678         log_string("starting from scratch.")
 679
 680     # except:
 681     # log_string("error when loading the checkpoint.")
 682     # exit(1)
 683
 684 ######################################################################
 685
 686 token_count = 0
 687 for input in task.batches(split="train"):
 688     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 689 token_probas = token_count / token_count.sum()
 690 entropy = -torch.xlogy(token_probas, token_probas).sum()
 691 train_set_perplexity = math.exp(entropy)
 692
 693 ##############################
 694
 695 if nb_epochs_finished >= args.nb_epochs:
 696     n_epoch = nb_epochs_finished
 697     train_perplexity = compute_perplexity(
 698         model, task, prompt_len=task.height * task.width, split="train"
 699     )
 700     test_perplexity = compute_perplexity(
 701         model, task, prompt_len=task.height * task.width, split="test"
 702     )
 703
 704     log_string(
 705         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 706     )
 707
 708     task.produce_results(n_epoch, model)
 709
 710 ##############################
 711
 712 learning_rate_scheduler.reset()
 713
 714 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 715     learning_rate = learning_rate_scheduler.get_learning_rate()
 716     log_string(f"learning_rate {n_epoch} {learning_rate}")
 717
 718     if args.optim == "sgd":
 719         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 720     elif args.optim == "adam":
 721         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 722     elif args.optim == "adamw":
 723         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 724     else:
 725         raise ValueError(f"{args.optim=}")
 726
 727     model.train()
 728
 729     nb_train_samples, acc_train_loss = 0, 0.0
 730
 731     for input in task.batches(split="train"):
 732         input = input.to(device)
 733         output = eval_mygpt(model, input, prompt_len=task.height * task.width)
 734         if args.noncausal_prompt:
 735             d = input.size(1) // 2
 736             loss = F.cross_entropy(output[:, d:].transpose(1, 2), input[:, d:])
 737         else:
 738             loss = F.cross_entropy(output.transpose(1, 2), input)
 739         acc_train_loss += loss.item() * input.size(0)
 740         nb_train_samples += input.size(0)
 741
 742         optimizer.zero_grad()
 743         loss.backward()
 744         optimizer.step()
 745
 746     learning_rate_scheduler.update(n_epoch + 1, acc_train_loss)
 747
 748     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 749     test_perplexity = compute_perplexity(
 750         model, task, prompt_len=task.height * task.width, split="test"
 751     )
 752
 753     log_string(
 754         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 755     )
 756
 757     task.produce_results(n_epoch, model)
 758
 759     checkpoint = {
 760         "nb_epochs_finished": n_epoch + 1,
 761         "model_state": model.state_dict(),
 762         "learning_rate_scheduler_state": learning_rate_scheduler.get_state(),
 763         "rng_state": torch.get_rng_state(),
 764     }
 765
 766     if torch.cuda.is_available():
 767         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 768
 769     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 770     torch.save(checkpoint, checkpoint_name)
 771     log_string(f"saved checkpoint {checkpoint_name}")
 772
 773 ######################################################################
 774
 775 if args.oneshot:
 776     oneshot(model, learning_rate_scheduler, task)
 777
 778 ######################################################################