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 log_string(f"cmd {' '.join(sys.argv)}")
 131
 132 for n in vars(args):
 133     log_string(f"args.{n} {getattr(args, n)}")
 134
 135 ######################################################################
 136
 137
 138 def reorder(x, order, reverse=False):  # x is NxTxD1x...xDk, order is NxT'
 139     u = x.reshape(x.size()[:2] + (-1,))
 140     order = order.unsqueeze(-1).expand(-1, -1, u.size(-1))
 141     if reverse:
 142         v = u.new(u.size()).scatter_(1, order, u)
 143     else:
 144         v = u.gather(1, order)
 145     v = v.reshape(v.size()[:2] + x.size()[2:])
 146     return v
 147
 148
 149 def shuffle(x, prompt_len):
 150     if args.random_regression_order:
 151         order = torch.rand(x.size(), device=x.device)
 152         order[:, :prompt_len] = torch.arange(-prompt_len, 0, device=x.device)
 153         order = order.sort(1).indices
 154     else:
 155         order = (
 156             torch.arange(x.size(1), device=x.device).unsqueeze(0).expand(x.size(0), -1)
 157         )
 158     return reorder(x, order), order
 159
 160
 161 def eval_mygpt(model, input, mode="standard", prompt_len=0):
 162     x, order = shuffle(input, prompt_len)
 163     x = model(mygpt.BracketedSequence(x), mode=mode, order=order).x
 164     return reorder(x, order, reverse=True)
 165
 166
 167 ######################################################################
 168
 169 # ar_mask is a Boolean matrix of same shape as input, with 1s on the
 170 # tokens that should be generated
 171
 172
 173 def masked_inplace_autoregression(model, batch_size, input, ar_mask, order=None):
 174     for input, ar_mask, order in zip(
 175         input.split(batch_size), ar_mask.split(batch_size), order.split(batch_size)
 176     ):
 177         i = (ar_mask.sum(0) > 0).nonzero()
 178         if i.min() > 0:
 179             # Needed to initialize the model's cache
 180             model(mygpt.BracketedSequence(input, 0, i.min()), order=order)
 181         for s in range(i.min(), i.max() + 1):
 182             output = model(mygpt.BracketedSequence(input, s, 1), order=order).x
 183             logits = output[:, s]
 184             if args.deterministic_synthesis:
 185                 t_next = logits.argmax(1)
 186             else:
 187                 dist = torch.distributions.categorical.Categorical(logits=logits)
 188                 t_next = dist.sample()
 189             input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 190
 191
 192 ######################################################################
 193
 194
 195 def compute_perplexity(model, task, prompt_len, split="train"):
 196     with torch.autograd.no_grad():
 197         t = model.training
 198         model.eval()
 199
 200         nb_samples, acc_loss = 0, 0.0
 201
 202         for input in task.batches(split=split):
 203             input = input.to(device)
 204             output = eval_mygpt(model, input, prompt_len=prompt_len)
 205             if args.noncausal_prompt:
 206                 d = input.size(1) // 2
 207                 loss = F.cross_entropy(output[:, d:].transpose(1, 2), input[:, d:])
 208             else:
 209                 loss = F.cross_entropy(output.transpose(1, 2), input)
 210             acc_loss += loss.item() * input.size(0)
 211             nb_samples += input.size(0)
 212
 213         model.train(t)
 214
 215         return math.exp(min(100, acc_loss / nb_samples))
 216
 217
 218 ######################################################################
 219
 220
 221 def oneshot_policy_loss(mazes, output, policies, height, width):
 222     masks = (mazes == maze.v_empty).unsqueeze(-1)
 223     targets = policies.permute(0, 2, 1) * masks
 224     output = output * masks
 225     return -(output.log_softmax(-1) * targets).sum() / masks.sum()
 226
 227
 228 def oneshot_trace_loss(mazes, output, policies, height, width):
 229     masks = mazes == maze.v_empty
 230     targets = maze.stationary_densities(
 231         mazes.view(-1, height, width), policies.view(-1, 4, height, width)
 232     ).flatten(-2)
 233     targets = targets * masks
 234     output = output.squeeze(-1) * masks
 235     return (output - targets).abs().sum() / masks.sum()
 236
 237
 238 def oneshot(model, learning_rate_scheduler, task):
 239     t = model.training
 240     model.eval()
 241     mazes = task.test_input[:48].clone()
 242     mazes[:, task.height * task.width :] = 0
 243     policies = task.test_policies[:48]
 244     targets = maze.stationary_densities(
 245         mazes[:, : task.height * task.width].view(-1, task.height, task.width),
 246         policies.view(-1, 4, task.height, task.width),
 247     ).flatten(-2)
 248     output = eval_mygpt(model, mazes, prompt_len=task.height * task.width)
 249     output = F.softmax(output, dim=2)
 250     print(f"{output.size()=}")
 251     proba_path = output[:, task.height * task.width :, 4].reshape(
 252         -1, task.height, task.width
 253     )
 254     mazes = mazes[:, : task.height * task.width].reshape(-1, task.height, task.width)
 255     targets = targets.reshape(-1, task.height, task.width)
 256     paths = task.test_input[:48, task.height * task.width :].reshape(
 257         -1, task.height, task.width
 258     )
 259     filename = f"oneshot.png"
 260     maze.save_image(
 261         os.path.join(args.result_dir, filename),
 262         mazes=mazes,
 263         # target_paths=paths,
 264         score_paths=proba_path,
 265         score_truth=targets,
 266     )
 267     log_string(f"wrote {filename}")
 268
 269
 270 def oneshot_old(gpt, learning_rate_scheduler, task):
 271     t = gpt.training
 272     gpt.eval()
 273
 274     if args.oneshot_input == "head":
 275         dim_in = args.dim_model
 276     elif args.oneshot_input == "deep":
 277         dim_in = args.dim_model * args.nb_blocks * 2
 278     else:
 279         raise ValueError(f"{args.oneshot_input=}")
 280
 281     if args.oneshot_output == "policy":
 282         dim_out = 4
 283         compute_loss = oneshot_policy_loss
 284     elif args.oneshot_output == "trace":
 285         dim_out = 1
 286         compute_loss = oneshot_trace_loss
 287     else:
 288         raise ValueError(f"{args.oneshot_output=}")
 289
 290     model = nn.Sequential(
 291         nn.Linear(dim_in, args.dim_model),
 292         nn.ReLU(),
 293         nn.Linear(args.dim_model, args.dim_model),
 294         nn.ReLU(),
 295         nn.Linear(args.dim_model, dim_out),
 296     ).to(device)
 297
 298     learning_rate_scheduler.reset()
 299
 300     for n_epoch in range(args.nb_epochs):
 301         learning_rate = learning_rate_scheduler.get_learning_rate()
 302         log_string(f"learning_rate {n_epoch} {learning_rate}")
 303
 304         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 305
 306         acc_train_loss, nb_train_samples = 0, 0
 307         for mazes, policies in task.policy_batches(split="train"):
 308             output_gpt = eval_mygpt(
 309                 gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
 310             )
 311             output = model(output_gpt)
 312
 313             loss = compute_loss(mazes, output, policies, task.height, task.width)
 314             acc_train_loss += loss.item() * mazes.size(0)
 315             nb_train_samples += mazes.size(0)
 316
 317             optimizer.zero_grad()
 318             loss.backward()
 319             optimizer.step()
 320
 321         learning_rate_scheduler.update(n_epoch + 1, acc_train_loss)
 322
 323         acc_test_loss, nb_test_samples = 0, 0
 324         for mazes, policies in task.policy_batches(split="test"):
 325             output_gpt = eval_mygpt(
 326                 gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
 327             )
 328             output = model(output_gpt)
 329             loss = compute_loss(mazes, output, policies, task.height, task.width)
 330             acc_test_loss += loss.item() * mazes.size(0)
 331             nb_test_samples += mazes.size(0)
 332
 333         log_string(
 334             f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
 335         )
 336
 337         # -------------------
 338         mazes = task.test_input[:48, : task.height * task.width]
 339         policies = task.test_policies[:48]
 340         output_gpt = eval_mygpt(
 341             gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
 342         )
 343         output = model(output_gpt)
 344         if args.oneshot_output == "policy":
 345             targets = policies.permute(0, 2, 1)
 346             scores = (
 347                 (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
 348             ).float()
 349         elif args.oneshot_output == "trace":
 350             targets = maze.stationary_densities(
 351                 mazes.view(-1, task.height, task.width),
 352                 policies.view(-1, 4, task.height, task.width),
 353             ).flatten(-2)
 354             scores = output
 355         else:
 356             raise ValueError(f"{args.oneshot_output=}")
 357
 358         scores = scores.reshape(-1, task.height, task.width)
 359         mazes = mazes.reshape(-1, task.height, task.width)
 360         targets = targets.reshape(-1, task.height, task.width)
 361         filename = (
 362             f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png"
 363         )
 364         maze.save_image(
 365             os.path.join(args.result_dir, filename),
 366             mazes=mazes,
 367             score_paths=scores,
 368             score_truth=targets,
 369         )
 370         log_string(f"wrote {filename}")
 371
 372         # -------------------
 373
 374     gpt.train(t)
 375
 376
 377 ######################################################################
 378
 379
 380 class LearningRateScheduler:
 381     def get_learning_rate(self):
 382         pass
 383
 384     def update(self, nb_finished_epochs, loss):
 385         pass
 386
 387     def reset(self):
 388         pass
 389
 390     def get_state(self):
 391         return vars(self)
 392
 393     def set_state(self, state):
 394         print(f"{state=}")
 395         for k, v in state.items():
 396             setattr(self, k, v)
 397
 398
 399 class StepWiseScheduler(LearningRateScheduler):
 400     def __init__(self, schedule):
 401         self.nb_finished_epochs = 0
 402         self.schedule = schedule
 403
 404     def get_learning_rate(self):
 405         return self.schedule[self.nb_finished_epochs]
 406
 407     def update(self, nb_finished_epochs, loss):
 408         self.nb_finished_epochs = nb_finished_epochs
 409
 410     def reset(self):
 411         self.nb_finished_epochs = 0
 412
 413     def get_state(self):
 414         return {"nb_finished_epochs": self.nb_finished_epochs}
 415
 416
 417 class AutoScheduler(LearningRateScheduler):
 418     def __init__(self, learning_rate_init, growth=1.0, degrowth=0.2):
 419         self.learning_rate_init = learning_rate_init
 420         self.learning_rate = learning_rate_init
 421         self.growth = growth
 422         self.degrowth = degrowth
 423         self.pred_loss = None
 424
 425     def get_learning_rate(self):
 426         return self.learning_rate
 427
 428     def update(self, nb_finished_epochs, loss):
 429         if self.pred_loss is not None:
 430             if loss >= self.pred_loss:
 431                 self.learning_rate *= self.degrowth
 432             else:
 433                 self.learning_rate *= self.growth
 434         self.pred_loss = loss
 435
 436     def reset(self):
 437         self.learning_rate = self.learning_rate_init
 438
 439     def get_state(self):
 440         return {
 441             "learning_rate_init": self.learning_rate_init,
 442             "pred_loss": self.pred_loss,
 443         }
 444
 445
 446 ######################################################################
 447
 448
 449 class Task:
 450     def batches(self, split="train", nb_to_use=-1, desc=None):
 451         pass
 452
 453     def vocabulary_size(self):
 454         pass
 455
 456     def produce_results(self, n_epoch, model):
 457         pass
 458
 459
 460 ######################################################################
 461
 462 import maze
 463
 464
 465 class TaskMaze(Task):
 466     def map2seq(self, *m):
 467         return torch.cat([x.flatten(1) for x in m], 1)
 468
 469     def seq2map(self, s):
 470         s = s.reshape(s.size(0), -1, self.height, self.width)
 471         return (s[:, k] for k in range(s.size(1)))
 472
 473     def __init__(
 474         self,
 475         nb_train_samples,
 476         nb_test_samples,
 477         batch_size,
 478         height,
 479         width,
 480         nb_walls,
 481         device=torch.device("cpu"),
 482     ):
 483         self.batch_size = batch_size
 484         self.height = height
 485         self.width = width
 486         self.device = device
 487
 488         train_mazes, train_paths, train_policies = maze.create_maze_data(
 489             nb_train_samples,
 490             height=height,
 491             width=width,
 492             nb_walls=nb_walls,
 493             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 494         )
 495         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 496         self.train_policies = train_policies.flatten(-2).to(device)
 497
 498         test_mazes, test_paths, test_policies = maze.create_maze_data(
 499             nb_test_samples,
 500             height=height,
 501             width=width,
 502             nb_walls=nb_walls,
 503             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 504         )
 505         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 506         self.test_policies = test_policies.flatten(-2).to(device)
 507
 508         self.nb_codes = self.train_input.max() + 1
 509
 510     def batches(self, split="train", nb_to_use=-1, desc=None):
 511         assert split in {"train", "test"}
 512         input = self.train_input if split == "train" else self.test_input
 513         if nb_to_use > 0:
 514             input = input[:nb_to_use]
 515         if desc is None:
 516             desc = f"epoch-{split}"
 517         for batch in tqdm.tqdm(
 518             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 519         ):
 520             yield batch
 521
 522     def policy_batches(self, split="train", nb_to_use=-1, desc=None):
 523         assert split in {"train", "test"}
 524         input = self.train_input if split == "train" else self.test_input
 525         policies = self.train_policies if split == "train" else self.test_policies
 526         input = input[:, : self.height * self.width]
 527         policies = policies * (input != maze.v_wall)[:, None]
 528
 529         if nb_to_use > 0:
 530             input = input[:nb_to_use]
 531             policies = policies[:nb_to_use]
 532
 533         if desc is None:
 534             desc = f"epoch-{split}"
 535         for batch in tqdm.tqdm(
 536             zip(input.split(self.batch_size), policies.split(self.batch_size)),
 537             dynamic_ncols=True,
 538             desc=desc,
 539         ):
 540             yield batch
 541
 542     def vocabulary_size(self):
 543         return self.nb_codes
 544
 545     def compute_error(self, model, split="train", nb_to_use=-1):
 546         nb_total, nb_correct = 0, 0
 547         for input in task.batches(split, nb_to_use):
 548             result = input.clone()
 549             ar_mask = result.new_zeros(result.size())
 550             ar_mask[:, self.height * self.width :] = 1
 551             result *= 1 - ar_mask
 552             x, order = shuffle(result, self.height * self.width)
 553             masked_inplace_autoregression(
 554                 model, self.batch_size, x, ar_mask, order=order
 555             )
 556             result = reorder(x, order, reverse=True)
 557             mazes, paths = self.seq2map(result)
 558             nb_correct += maze.path_correctness(mazes, paths).long().sum()
 559             nb_total += mazes.size(0)
 560
 561         return nb_total, nb_correct
 562
 563     def produce_results(self, n_epoch, model):
 564         with torch.autograd.no_grad():
 565             t = model.training
 566             model.eval()
 567
 568             train_nb_total, train_nb_correct = self.compute_error(
 569                 model, "train", nb_to_use=1000
 570             )
 571             log_string(
 572                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 573             )
 574
 575             test_nb_total, test_nb_correct = self.compute_error(
 576                 model, "test", nb_to_use=1000
 577             )
 578             log_string(
 579                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 580             )
 581
 582             input = self.test_input[:48]
 583             result = input.clone()
 584             ar_mask = result.new_zeros(result.size())
 585             ar_mask[:, self.height * self.width :] = 1
 586             result *= 1 - ar_mask
 587             x, order = shuffle(result, self.height * self.width)
 588             masked_inplace_autoregression(
 589                 model, self.batch_size, x, ar_mask, order=order
 590             )
 591             result = reorder(x, order, reverse=True)
 592
 593             mazes, paths = self.seq2map(input)
 594             _, predicted_paths = self.seq2map(result)
 595             filename = f"result_{n_epoch:04d}.png"
 596             maze.save_image(
 597                 os.path.join(args.result_dir, filename),
 598                 mazes=mazes,
 599                 target_paths=paths,
 600                 predicted_paths=predicted_paths,
 601                 path_correct=maze.path_correctness(mazes, predicted_paths),
 602             )
 603             log_string(f"wrote {filename}")
 604
 605             model.train(t)
 606
 607
 608 ######################################################################
 609
 610 log_string(f"device {device}")
 611
 612
 613 task = TaskMaze(
 614     nb_train_samples=args.nb_train_samples,
 615     nb_test_samples=args.nb_test_samples,
 616     batch_size=args.batch_size,
 617     height=args.maze_height,
 618     width=args.maze_width,
 619     nb_walls=args.maze_nb_walls,
 620     device=device,
 621 )
 622
 623
 624 vocabulary_size = task.vocabulary_size()
 625
 626 log_string(f"vocabulary_size {vocabulary_size}")
 627
 628 ##############################
 629
 630
 631 def noncausal_prompt_amm_generator(d):
 632     q = torch.arange(d)[:, None]
 633     k = torch.arange(d)[None, :]
 634     s = args.maze_height * args.maze_width
 635     return torch.logical_and(q < k, torch.logical_or(q >= s, k >= s))
 636     # return q < k
 637
 638
 639 def noncausal_prompt_oneshot_amm_generator(d):
 640     q = torch.arange(d)[:, None]
 641     k = torch.arange(d)[None, :]
 642     s = args.maze_height * args.maze_width
 643     return k >= s
 644     # return q < k
 645
 646
 647 if args.oneshot:
 648     amm_generator = noncausal_prompt_oneshot_amm_generator
 649 elif args.noncausal_prompt:
 650     amm_generator = noncausal_prompt_amm_generator
 651 else:
 652     amm_generator = None
 653
 654 model = mygpt.MyGPT(
 655     vocabulary_size=vocabulary_size,
 656     dim_model=args.dim_model,
 657     dim_keys=args.dim_keys,
 658     dim_hidden=args.dim_hidden,
 659     nb_heads=args.nb_heads,
 660     nb_blocks=args.nb_blocks,
 661     causal=True,
 662     dropout=args.dropout,
 663     amm_generator=amm_generator,
 664 )
 665
 666 model.to(device)
 667
 668 nb_parameters = sum(p.numel() for p in model.parameters())
 669 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 670
 671 ######################################################################
 672
 673 if args.learning_rate_schedule == "auto":
 674     learning_rate_scheduler = AutoScheduler(args.learning_rate)
 675
 676 elif args.learning_rate_schedule == "cos":
 677     schedule = {}
 678     for n_epoch in range(args.nb_epochs):
 679         u = n_epoch / args.nb_epochs * math.pi
 680         schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 681     learning_rate_scheduler = StepWiseScheduler(schedule)
 682     log_string(f"learning_rate_schedule {schedule}")
 683
 684 else:
 685     u = {
 686         int(k): float(v)
 687         for k, v in [
 688             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 689         ]
 690     }
 691
 692     schedule = {}
 693     learning_rate = args.learning_rate
 694     for n_epoch in range(args.nb_epochs):
 695         if n_epoch in u:
 696             learning_rate = u[n_epoch]
 697         schedule[n_epoch] = learning_rate
 698     learning_rate_scheduler = StepWiseScheduler(schedule)
 699     log_string(f"learning_rate_schedule {schedule}")
 700
 701 ######################################################################
 702
 703 nb_epochs_finished = 0
 704
 705 if args.no_checkpoint:
 706     log_string(f"not trying to load checkpoint.")
 707
 708 else:
 709     try:
 710         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 711         checkpoint = torch.load(checkpoint_name)
 712         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 713         model.load_state_dict(checkpoint["model_state"])
 714         learning_rate_scheduler.set_state(checkpoint["learning_rate_scheduler_state"])
 715         torch.set_rng_state(checkpoint["rng_state"])
 716         if torch.cuda.is_available():
 717             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 718
 719         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 720
 721     except FileNotFoundError:
 722         log_string("starting from scratch.")
 723
 724     # except:
 725     # log_string("error when loading the checkpoint.")
 726     # exit(1)
 727
 728 ######################################################################
 729
 730 if args.oneshot:
 731     oneshot(model, learning_rate_scheduler, task)
 732     exit(0)
 733
 734 ######################################################################
 735
 736 token_count = 0
 737 for input in task.batches(split="train"):
 738     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 739 token_probas = token_count / token_count.sum()
 740 entropy = -torch.xlogy(token_probas, token_probas).sum()
 741 train_set_perplexity = math.exp(entropy)
 742
 743 ##############################
 744
 745 if nb_epochs_finished >= args.nb_epochs:
 746     n_epoch = nb_epochs_finished
 747     train_perplexity = compute_perplexity(
 748         model, task, prompt_len=task.height * task.width, split="train"
 749     )
 750     test_perplexity = compute_perplexity(
 751         model, task, prompt_len=task.height * task.width, split="test"
 752     )
 753
 754     log_string(
 755         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 756     )
 757
 758     task.produce_results(n_epoch, model)
 759
 760 ##############################
 761
 762 learning_rate_scheduler.reset()
 763
 764 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 765     learning_rate = learning_rate_scheduler.get_learning_rate()
 766     log_string(f"learning_rate {n_epoch} {learning_rate}")
 767
 768     if args.optim == "sgd":
 769         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 770     elif args.optim == "adam":
 771         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 772     elif args.optim == "adamw":
 773         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 774     else:
 775         raise ValueError(f"{args.optim=}")
 776
 777     model.train()
 778
 779     nb_train_samples, acc_train_loss = 0, 0.0
 780
 781     for input in task.batches(split="train"):
 782         input = input.to(device)
 783         output = eval_mygpt(model, input, prompt_len=task.height * task.width)
 784         if args.noncausal_prompt:
 785             d = input.size(1) // 2
 786             loss = F.cross_entropy(output[:, d:].transpose(1, 2), input[:, d:])
 787         else:
 788             loss = F.cross_entropy(output.transpose(1, 2), input)
 789         acc_train_loss += loss.item() * input.size(0)
 790         nb_train_samples += input.size(0)
 791
 792         optimizer.zero_grad()
 793         loss.backward()
 794         optimizer.step()
 795
 796     learning_rate_scheduler.update(n_epoch + 1, acc_train_loss)
 797
 798     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 799     test_perplexity = compute_perplexity(
 800         model, task, prompt_len=task.height * task.width, split="test"
 801     )
 802
 803     log_string(
 804         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 805     )
 806
 807     task.produce_results(n_epoch, model)
 808
 809     checkpoint = {
 810         "nb_epochs_finished": n_epoch + 1,
 811         "model_state": model.state_dict(),
 812         "learning_rate_scheduler_state": learning_rate_scheduler.get_state(),
 813         "rng_state": torch.get_rng_state(),
 814     }
 815
 816     if torch.cuda.is_available():
 817         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 818
 819     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 820     torch.save(checkpoint, checkpoint_name)
 821     log_string(f"saved checkpoint {checkpoint_name}")
 822
 823 ######################################################################