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     policies = task.test_policies[:32]
 242     targets = maze.stationary_densities(
 243         mazes[:, : task.height * task.width].view(-1, task.height, task.width),
 244         policies.view(-1, 4, task.height, task.width),
 245     ).flatten(-2)
 246     output = eval_mygpt(model, mazes, prompt_len=task.height * task.width)
 247     output = F.softmax(output, dim=2)
 248     print(f"{output.size()=}")
 249     proba_path = output[:, task.height * task.width :, 4].reshape(
 250         -1, task.height, task.width
 251     )
 252     mazes = mazes[:, : task.height * task.width].reshape(-1, task.height, task.width)
 253     targets = targets.reshape(-1, task.height, task.width)
 254     paths = task.test_input[:32, task.height * task.width :].reshape(
 255         -1, task.height, task.width
 256     )
 257     filename = f"oneshot.png"
 258     maze.save_image(
 259         os.path.join(args.result_dir, filename),
 260         mazes=mazes,
 261         # target_paths=paths,
 262         score_paths=proba_path,
 263         score_truth=targets,
 264     )
 265     log_string(f"wrote {filename}")
 266
 267
 268 def oneshot_old(gpt, learning_rate_scheduler, task):
 269     t = gpt.training
 270     gpt.eval()
 271
 272     if args.oneshot_input == "head":
 273         dim_in = args.dim_model
 274     elif args.oneshot_input == "deep":
 275         dim_in = args.dim_model * args.nb_blocks * 2
 276     else:
 277         raise ValueError(f"{args.oneshot_input=}")
 278
 279     if args.oneshot_output == "policy":
 280         dim_out = 4
 281         compute_loss = oneshot_policy_loss
 282     elif args.oneshot_output == "trace":
 283         dim_out = 1
 284         compute_loss = oneshot_trace_loss
 285     else:
 286         raise ValueError(f"{args.oneshot_output=}")
 287
 288     model = nn.Sequential(
 289         nn.Linear(dim_in, args.dim_model),
 290         nn.ReLU(),
 291         nn.Linear(args.dim_model, args.dim_model),
 292         nn.ReLU(),
 293         nn.Linear(args.dim_model, dim_out),
 294     ).to(device)
 295
 296     learning_rate_scheduler.reset()
 297
 298     for n_epoch in range(args.nb_epochs):
 299         learning_rate = learning_rate_scheduler.get_learning_rate()
 300         log_string(f"learning_rate {n_epoch} {learning_rate}")
 301
 302         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 303
 304         acc_train_loss, nb_train_samples = 0, 0
 305         for mazes, policies in task.policy_batches(split="train"):
 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
 311             loss = compute_loss(mazes, output, policies, task.height, task.width)
 312             acc_train_loss += loss.item() * mazes.size(0)
 313             nb_train_samples += mazes.size(0)
 314
 315             optimizer.zero_grad()
 316             loss.backward()
 317             optimizer.step()
 318
 319         learning_rate_scheduler.update(n_epoch + 1, acc_train_loss)
 320
 321         acc_test_loss, nb_test_samples = 0, 0
 322         for mazes, policies in task.policy_batches(split="test"):
 323             output_gpt = eval_mygpt(
 324                 gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
 325             )
 326             output = model(output_gpt)
 327             loss = compute_loss(mazes, output, policies, task.height, task.width)
 328             acc_test_loss += loss.item() * mazes.size(0)
 329             nb_test_samples += mazes.size(0)
 330
 331         log_string(
 332             f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
 333         )
 334
 335         # -------------------
 336         mazes = task.test_input[:32, : task.height * task.width]
 337         policies = task.test_policies[:32]
 338         output_gpt = eval_mygpt(
 339             gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
 340         )
 341         output = model(output_gpt)
 342         if args.oneshot_output == "policy":
 343             targets = policies.permute(0, 2, 1)
 344             scores = (
 345                 (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
 346             ).float()
 347         elif args.oneshot_output == "trace":
 348             targets = maze.stationary_densities(
 349                 mazes.view(-1, task.height, task.width),
 350                 policies.view(-1, 4, task.height, task.width),
 351             ).flatten(-2)
 352             scores = output
 353         else:
 354             raise ValueError(f"{args.oneshot_output=}")
 355
 356         scores = scores.reshape(-1, task.height, task.width)
 357         mazes = mazes.reshape(-1, task.height, task.width)
 358         targets = targets.reshape(-1, task.height, task.width)
 359         filename = (
 360             f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png"
 361         )
 362         maze.save_image(
 363             os.path.join(args.result_dir, filename),
 364             mazes=mazes,
 365             score_paths=scores,
 366             score_truth=targets,
 367         )
 368         log_string(f"wrote {filename}")
 369
 370         # -------------------
 371
 372     gpt.train(t)
 373
 374
 375 ######################################################################
 376
 377
 378 class LearningRateScheduler:
 379     def get_learning_rate(self):
 380         pass
 381
 382     def update(self, nb_finished_epochs, loss):
 383         pass
 384
 385     def reset(self):
 386         pass
 387
 388     def get_state(self):
 389         return vars(self)
 390
 391     def set_state(self, state):
 392         print(f"{state=}")
 393         for k, v in state.items():
 394             setattr(self, k, v)
 395
 396
 397 class StepWiseScheduler(LearningRateScheduler):
 398     def __init__(self, schedule):
 399         self.nb_finished_epochs = 0
 400         self.schedule = schedule
 401
 402     def get_learning_rate(self):
 403         return self.schedule[self.nb_finished_epochs]
 404
 405     def update(self, nb_finished_epochs, loss):
 406         self.nb_finished_epochs = nb_finished_epochs
 407
 408     def reset(self):
 409         self.nb_finished_epochs = 0
 410
 411     def get_state(self):
 412         return {"nb_finished_epochs": self.nb_finished_epochs}
 413
 414
 415 class AutoScheduler(LearningRateScheduler):
 416     def __init__(self, learning_rate_init, growth=1.0, degrowth=0.2):
 417         self.learning_rate_init = learning_rate_init
 418         self.learning_rate = learning_rate_init
 419         self.growth = growth
 420         self.degrowth = degrowth
 421         self.pred_loss = None
 422
 423     def get_learning_rate(self):
 424         return self.learning_rate
 425
 426     def update(self, nb_finished_epochs, loss):
 427         if self.pred_loss is not None:
 428             if loss >= self.pred_loss:
 429                 self.learning_rate *= self.degrowth
 430             else:
 431                 self.learning_rate *= self.growth
 432         self.pred_loss = loss
 433
 434     def reset(self):
 435         self.learning_rate = self.learning_rate_init
 436
 437     def get_state(self):
 438         return {
 439             "learning_rate_init": self.learning_rate_init,
 440             "pred_loss": self.pred_loss,
 441         }
 442
 443
 444 ######################################################################
 445
 446
 447 class Task:
 448     def batches(self, split="train", nb_to_use=-1, desc=None):
 449         pass
 450
 451     def vocabulary_size(self):
 452         pass
 453
 454     def produce_results(self, n_epoch, model):
 455         pass
 456
 457
 458 ######################################################################
 459
 460 import maze
 461
 462
 463 class TaskMaze(Task):
 464     def map2seq(self, *m):
 465         return torch.cat([x.flatten(1) for x in m], 1)
 466
 467     def seq2map(self, s):
 468         s = s.reshape(s.size(0), -1, self.height, self.width)
 469         return (s[:, k] for k in range(s.size(1)))
 470
 471     def __init__(
 472         self,
 473         nb_train_samples,
 474         nb_test_samples,
 475         batch_size,
 476         height,
 477         width,
 478         nb_walls,
 479         device=torch.device("cpu"),
 480     ):
 481         self.batch_size = batch_size
 482         self.height = height
 483         self.width = width
 484         self.device = device
 485
 486         train_mazes, train_paths, train_policies = maze.create_maze_data(
 487             nb_train_samples,
 488             height=height,
 489             width=width,
 490             nb_walls=nb_walls,
 491             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 492         )
 493         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 494         self.train_policies = train_policies.flatten(-2).to(device)
 495
 496         test_mazes, test_paths, test_policies = maze.create_maze_data(
 497             nb_test_samples,
 498             height=height,
 499             width=width,
 500             nb_walls=nb_walls,
 501             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 502         )
 503         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 504         self.test_policies = test_policies.flatten(-2).to(device)
 505
 506         self.nb_codes = self.train_input.max() + 1
 507
 508     def batches(self, split="train", nb_to_use=-1, desc=None):
 509         assert split in {"train", "test"}
 510         input = self.train_input if split == "train" else self.test_input
 511         if nb_to_use > 0:
 512             input = input[:nb_to_use]
 513         if desc is None:
 514             desc = f"epoch-{split}"
 515         for batch in tqdm.tqdm(
 516             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 517         ):
 518             yield batch
 519
 520     def policy_batches(self, split="train", nb_to_use=-1, desc=None):
 521         assert split in {"train", "test"}
 522         input = self.train_input if split == "train" else self.test_input
 523         policies = self.train_policies if split == "train" else self.test_policies
 524         input = input[:, : self.height * self.width]
 525         policies = policies * (input != maze.v_wall)[:, None]
 526
 527         if nb_to_use > 0:
 528             input = input[:nb_to_use]
 529             policies = policies[:nb_to_use]
 530
 531         if desc is None:
 532             desc = f"epoch-{split}"
 533         for batch in tqdm.tqdm(
 534             zip(input.split(self.batch_size), policies.split(self.batch_size)),
 535             dynamic_ncols=True,
 536             desc=desc,
 537         ):
 538             yield batch
 539
 540     def vocabulary_size(self):
 541         return self.nb_codes
 542
 543     def compute_error(self, model, split="train", nb_to_use=-1):
 544         nb_total, nb_correct = 0, 0
 545         for input in task.batches(split, nb_to_use):
 546             result = input.clone()
 547             ar_mask = result.new_zeros(result.size())
 548             ar_mask[:, self.height * self.width :] = 1
 549             result *= 1 - ar_mask
 550             x, order = shuffle(result, self.height * self.width)
 551             masked_inplace_autoregression(
 552                 model, self.batch_size, x, ar_mask, order=order
 553             )
 554             result = reorder(x, order, reverse=True)
 555             mazes, paths = self.seq2map(result)
 556             nb_correct += maze.path_correctness(mazes, paths).long().sum()
 557             nb_total += mazes.size(0)
 558
 559         return nb_total, nb_correct
 560
 561     def produce_results(self, n_epoch, model):
 562         with torch.autograd.no_grad():
 563             t = model.training
 564             model.eval()
 565
 566             train_nb_total, train_nb_correct = self.compute_error(
 567                 model, "train", nb_to_use=1000
 568             )
 569             log_string(
 570                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 571             )
 572
 573             test_nb_total, test_nb_correct = self.compute_error(
 574                 model, "test", nb_to_use=1000
 575             )
 576             log_string(
 577                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 578             )
 579
 580             input = self.test_input[:32]
 581             result = input.clone()
 582             ar_mask = result.new_zeros(result.size())
 583             ar_mask[:, self.height * self.width :] = 1
 584             result *= 1 - ar_mask
 585             x, order = shuffle(result, self.height * self.width)
 586             masked_inplace_autoregression(
 587                 model, self.batch_size, x, ar_mask, order=order
 588             )
 589             result = reorder(x, order, reverse=True)
 590
 591             mazes, paths = self.seq2map(input)
 592             _, predicted_paths = self.seq2map(result)
 593             filename = f"result_{n_epoch:04d}.png"
 594             maze.save_image(
 595                 os.path.join(args.result_dir, filename),
 596                 mazes=mazes,
 597                 target_paths=paths,
 598                 predicted_paths=predicted_paths,
 599                 path_correct=maze.path_correctness(mazes, predicted_paths),
 600             )
 601             log_string(f"wrote {filename}")
 602
 603             model.train(t)
 604
 605
 606 ######################################################################
 607
 608 log_string(f"device {device}")
 609
 610
 611 task = TaskMaze(
 612     nb_train_samples=args.nb_train_samples,
 613     nb_test_samples=args.nb_test_samples,
 614     batch_size=args.batch_size,
 615     height=args.maze_height,
 616     width=args.maze_width,
 617     nb_walls=args.maze_nb_walls,
 618     device=device,
 619 )
 620
 621
 622 vocabulary_size = task.vocabulary_size()
 623
 624 log_string(f"vocabulary_size {vocabulary_size}")
 625
 626 ##############################
 627
 628
 629 def noncausal_prompt_amm_generator(d):
 630     q = torch.arange(d)[:, None]
 631     k = torch.arange(d)[None, :]
 632     s = args.maze_height * args.maze_width
 633     return torch.logical_and(q < k, torch.logical_or(q >= s, k >= s))
 634     # return q < k
 635
 636
 637 def noncausal_prompt_oneshot_amm_generator(d):
 638     q = torch.arange(d)[:, None]
 639     k = torch.arange(d)[None, :]
 640     s = args.maze_height * args.maze_width
 641     return k >= s
 642     # return q < k
 643
 644
 645 if args.oneshot:
 646     amm_generator = noncausal_prompt_oneshot_amm_generator
 647 elif args.noncausal_prompt:
 648     amm_generator = noncausal_prompt_amm_generator
 649 else:
 650     amm_generator = None
 651
 652 model = mygpt.MyGPT(
 653     vocabulary_size=vocabulary_size,
 654     dim_model=args.dim_model,
 655     dim_keys=args.dim_keys,
 656     dim_hidden=args.dim_hidden,
 657     nb_heads=args.nb_heads,
 658     nb_blocks=args.nb_blocks,
 659     causal=True,
 660     dropout=args.dropout,
 661     amm_generator=amm_generator,
 662 )
 663
 664 model.to(device)
 665
 666 nb_parameters = sum(p.numel() for p in model.parameters())
 667 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 668
 669 ######################################################################
 670
 671 if args.learning_rate_schedule == "auto":
 672     learning_rate_scheduler = AutoScheduler(args.learning_rate)
 673
 674 elif args.learning_rate_schedule == "cos":
 675     schedule = {}
 676     for n_epoch in range(args.nb_epochs):
 677         u = n_epoch / args.nb_epochs * math.pi
 678         schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 679     learning_rate_scheduler = StepWiseScheduler(schedule)
 680     log_string(f"learning_rate_schedule {schedule}")
 681
 682 else:
 683     u = {
 684         int(k): float(v)
 685         for k, v in [
 686             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 687         ]
 688     }
 689
 690     schedule = {}
 691     learning_rate = args.learning_rate
 692     for n_epoch in range(args.nb_epochs):
 693         if n_epoch in u:
 694             learning_rate = u[n_epoch]
 695         schedule[n_epoch] = learning_rate
 696     learning_rate_scheduler = StepWiseScheduler(schedule)
 697     log_string(f"learning_rate_schedule {schedule}")
 698
 699 ######################################################################
 700
 701 nb_epochs_finished = 0
 702
 703 if args.no_checkpoint:
 704     log_string(f"not trying to load checkpoint.")
 705
 706 else:
 707     try:
 708         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 709         checkpoint = torch.load(checkpoint_name)
 710         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 711         model.load_state_dict(checkpoint["model_state"])
 712         learning_rate_scheduler.set_state(checkpoint["learning_rate_scheduler_state"])
 713         torch.set_rng_state(checkpoint["rng_state"])
 714         if torch.cuda.is_available():
 715             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 716
 717         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 718
 719     except FileNotFoundError:
 720         log_string("starting from scratch.")
 721
 722     # except:
 723     # log_string("error when loading the checkpoint.")
 724     # exit(1)
 725
 726 ######################################################################
 727
 728 if args.oneshot:
 729     oneshot(model, learning_rate_scheduler, task)
 730     exit(0)
 731
 732 ######################################################################
 733
 734 token_count = 0
 735 for input in task.batches(split="train"):
 736     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 737 token_probas = token_count / token_count.sum()
 738 entropy = -torch.xlogy(token_probas, token_probas).sum()
 739 train_set_perplexity = math.exp(entropy)
 740
 741 ##############################
 742
 743 if nb_epochs_finished >= args.nb_epochs:
 744     n_epoch = nb_epochs_finished
 745     train_perplexity = compute_perplexity(
 746         model, task, prompt_len=task.height * task.width, split="train"
 747     )
 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 ##############################
 759
 760 learning_rate_scheduler.reset()
 761
 762 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 763     learning_rate = learning_rate_scheduler.get_learning_rate()
 764     log_string(f"learning_rate {n_epoch} {learning_rate}")
 765
 766     if args.optim == "sgd":
 767         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 768     elif args.optim == "adam":
 769         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 770     elif args.optim == "adamw":
 771         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 772     else:
 773         raise ValueError(f"{args.optim=}")
 774
 775     model.train()
 776
 777     nb_train_samples, acc_train_loss = 0, 0.0
 778
 779     for input in task.batches(split="train"):
 780         input = input.to(device)
 781         output = eval_mygpt(model, input, prompt_len=task.height * task.width)
 782         if args.noncausal_prompt:
 783             d = input.size(1) // 2
 784             loss = F.cross_entropy(output[:, d:].transpose(1, 2), input[:, d:])
 785         else:
 786             loss = F.cross_entropy(output.transpose(1, 2), input)
 787         acc_train_loss += loss.item() * input.size(0)
 788         nb_train_samples += input.size(0)
 789
 790         optimizer.zero_grad()
 791         loss.backward()
 792         optimizer.step()
 793
 794     learning_rate_scheduler.update(n_epoch + 1, acc_train_loss)
 795
 796     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 797     test_perplexity = compute_perplexity(
 798         model, task, prompt_len=task.height * task.width, split="test"
 799     )
 800
 801     log_string(
 802         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 803     )
 804
 805     task.produce_results(n_epoch, model)
 806
 807     checkpoint = {
 808         "nb_epochs_finished": n_epoch + 1,
 809         "model_state": model.state_dict(),
 810         "learning_rate_scheduler_state": learning_rate_scheduler.get_state(),
 811         "rng_state": torch.get_rng_state(),
 812     }
 813
 814     if torch.cuda.is_available():
 815         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 816
 817     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 818     torch.save(checkpoint, checkpoint_name)
 819     log_string(f"saved checkpoint {checkpoint_name}")
 820
 821 ######################################################################