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("--no_checkpoint", action="store_true", default=False)
  68
  69 parser.add_argument("--overwrite_results", action="store_true", default=False)
  70
  71 parser.add_argument("--one_shot", action="store_true", default=False)
  72
  73 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
  74
  75 ##############################
  76 # maze options
  77
  78 parser.add_argument("--maze_height", type=int, default=13)
  79
  80 parser.add_argument("--maze_width", type=int, default=21)
  81
  82 parser.add_argument("--maze_nb_walls", type=int, default=15)
  83
  84 parser.add_argument("--oneshot_mode", type=str, default="head")
  85
  86 ######################################################################
  87
  88 args = parser.parse_args()
  89
  90 try:
  91     os.mkdir(args.result_dir)
  92 except FileExistsError:
  93     if not args.overwrite_results:
  94         print(f"result directory {args.result_dir} already exists")
  95         exit(1)
  96
  97 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
  98
  99 if args.seed >= 0:
 100     # torch.backends.cudnn.deterministic = True
 101     # torch.backends.cudnn.benchmark = False
 102     # torch.use_deterministic_algorithms(True)
 103     torch.manual_seed(args.seed)
 104     if torch.cuda.is_available():
 105         torch.cuda.manual_seed_all(args.seed)
 106
 107 ######################################################################
 108
 109
 110 def log_string(s):
 111     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 112
 113     if log_file is not None:
 114         log_file.write(t + s + "\n")
 115         log_file.flush()
 116
 117     print(t + s)
 118     sys.stdout.flush()
 119
 120
 121 for n in vars(args):
 122     log_string(f"args.{n} {getattr(args, n)}")
 123
 124 ######################################################################
 125
 126
 127 # ar_mask is a Boolean matrix of same shape as input, with 1s on the
 128 # tokens that should be generated
 129
 130
 131 def masked_inplace_autoregression(model, batch_size, input, ar_mask):
 132     for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
 133         i = (ar_mask.sum(0) > 0).nonzero()
 134         if i.min() > 0:
 135             # Needed to initialize the model's cache
 136             model(mygpt.BracketedSequence(input, 0, i.min()))
 137         for s in range(i.min(), i.max() + 1):
 138             output = model(mygpt.BracketedSequence(input, s, 1)).x
 139             logits = output[:, s]
 140             if args.deterministic_synthesis:
 141                 t_next = logits.argmax(1)
 142             else:
 143                 dist = torch.distributions.categorical.Categorical(logits=logits)
 144                 t_next = dist.sample()
 145             input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 146
 147
 148 ######################################################################
 149
 150
 151 def compute_perplexity(model, split="train"):
 152     with torch.autograd.no_grad():
 153         t = model.training
 154         model.eval()
 155
 156         nb_samples, acc_loss = 0, 0.0
 157
 158         for input in task.batches(split=split):
 159             input = input.to(device)
 160
 161             output = model(mygpt.BracketedSequence(input)).x
 162             loss = F.cross_entropy(output.transpose(1, 2), input)
 163             acc_loss += loss.item() * input.size(0)
 164             nb_samples += input.size(0)
 165
 166         model.train(t)
 167
 168         return math.exp(min(100, acc_loss / nb_samples))
 169
 170
 171 ######################################################################
 172
 173
 174 def one_shot(gpt, task):
 175     t = gpt.training
 176     gpt.eval()
 177     dim_in = args.dim_model * (args.nb_blocks * 2 if args.oneshot_mode == "deep" else 1)
 178     model = nn.Sequential(
 179         nn.Linear(dim_in, args.dim_model),
 180         nn.ReLU(),
 181         nn.Linear(args.dim_model, args.dim_model),
 182         nn.ReLU(),
 183         nn.Linear(args.dim_model, 4),
 184     ).to(device)
 185
 186     for n_epoch in range(args.nb_epochs):
 187         learning_rate = learning_rate_schedule[n_epoch]
 188         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 189
 190         acc_train_loss, nb_train_samples = 0, 0
 191         for input, policies in task.policy_batches(split="train"):
 192             ####
 193             # print(f'{input.size()=} {policies.size()=}')
 194             # s = maze.stationary_densities(
 195             # exit(0)
 196             ####
 197             mask = input.unsqueeze(-1) == maze.v_empty
 198             output_gpt = gpt(mygpt.BracketedSequence(input), mode=args.oneshot_mode).x
 199             output = model(output_gpt)
 200             targets = policies.permute(0, 2, 1) * mask
 201             output = output * mask
 202             # loss = (output.softmax(-1) - targets).abs().max(-1).values.mean()
 203             loss = -(output.log_softmax(-1) * targets).sum() / mask.sum()
 204             acc_train_loss += loss.item() * input.size(0)
 205             nb_train_samples += input.size(0)
 206
 207             optimizer.zero_grad()
 208             loss.backward()
 209             optimizer.step()
 210
 211         acc_test_loss, nb_test_samples = 0, 0
 212         for input, policies in task.policy_batches(split="test"):
 213             mask = input.unsqueeze(-1) == maze.v_empty
 214             output_gpt = gpt(mygpt.BracketedSequence(input), mode=args.oneshot_mode).x
 215             output = model(output_gpt)
 216             targets = policies.permute(0, 2, 1) * mask
 217             output = output * mask
 218             # loss = (output.softmax(-1) - targets).abs().max(-1).values.mean()
 219             loss = -(output.log_softmax(-1) * targets).sum() / mask.sum()
 220             acc_test_loss += loss.item() * input.size(0)
 221             nb_test_samples += input.size(0)
 222
 223         log_string(
 224             f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
 225         )
 226
 227         # -------------------
 228         input = task.test_input[:32, : task.height * task.width]
 229         targets = task.test_policies[:32].permute(0, 2, 1)
 230         output_gpt = gpt(mygpt.BracketedSequence(input), mode=args.oneshot_mode).x
 231         output = model(output_gpt)
 232         # losses = (-output.log_softmax(-1) * targets + targets.xlogy(targets)).sum(-1)
 233         # losses = losses * mask
 234         # losses = losses / losses.max()
 235         # losses = (output.softmax(-1) - targets).abs().max(-1).values
 236         # losses = (losses >= 0.05).float()
 237         losses = (
 238             (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
 239         ).float()
 240         losses = losses.reshape(-1, args.maze_height, args.maze_width)
 241         input = input.reshape(-1, args.maze_height, args.maze_width)
 242         maze.save_image(
 243             os.path.join(
 244                 args.result_dir, f"oneshot_{args.oneshot_mode}_{n_epoch:04d}.png"
 245             ),
 246             mazes=input,
 247             score_paths=losses,
 248         )
 249         # -------------------
 250
 251     gpt.train(t)
 252
 253
 254 ######################################################################
 255
 256
 257 class Task:
 258     def batches(self, split="train"):
 259         pass
 260
 261     def vocabulary_size(self):
 262         pass
 263
 264     def produce_results(self, n_epoch, model):
 265         pass
 266
 267
 268 ######################################################################
 269
 270 import maze
 271
 272
 273 class TaskMaze(Task):
 274     def map2seq(self, *m):
 275         return torch.cat([x.flatten(1) for x in m], 1)
 276
 277     def seq2map(self, s):
 278         s = s.reshape(s.size(0), -1, self.height, self.width)
 279         return (s[:, k] for k in range(s.size(1)))
 280
 281     def __init__(
 282         self,
 283         nb_train_samples,
 284         nb_test_samples,
 285         batch_size,
 286         height,
 287         width,
 288         nb_walls,
 289         device=torch.device("cpu"),
 290     ):
 291         self.batch_size = batch_size
 292         self.height = height
 293         self.width = width
 294         self.device = device
 295
 296         train_mazes, train_paths, train_policies = maze.create_maze_data(
 297             nb_train_samples,
 298             height=height,
 299             width=width,
 300             nb_walls=nb_walls,
 301             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 302         )
 303         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 304         self.train_policies = train_policies.flatten(-2).to(device)
 305
 306         test_mazes, test_paths, test_policies = maze.create_maze_data(
 307             nb_test_samples,
 308             height=height,
 309             width=width,
 310             nb_walls=nb_walls,
 311             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 312         )
 313         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 314         self.test_policies = test_policies.flatten(-2).to(device)
 315
 316         self.nb_codes = self.train_input.max() + 1
 317
 318     def batches(self, split="train", nb_to_use=-1):
 319         assert split in {"train", "test"}
 320         input = self.train_input if split == "train" else self.test_input
 321         if nb_to_use > 0:
 322             input = input[:nb_to_use]
 323         for batch in tqdm.tqdm(
 324             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 325         ):
 326             yield batch
 327
 328     def policy_batches(self, split="train", nb_to_use=-1):
 329         assert split in {"train", "test"}
 330         input = self.train_input if split == "train" else self.test_input
 331         policies = self.train_policies if split == "train" else self.test_policies
 332         input = input[:, : self.height * self.width]
 333         policies = policies * (input != maze.v_wall)[:, None]
 334
 335         if nb_to_use > 0:
 336             input = input[:nb_to_use]
 337             policies = policies[:nb_to_use]
 338
 339         for batch in tqdm.tqdm(
 340             zip(input.split(self.batch_size), policies.split(self.batch_size)),
 341             dynamic_ncols=True,
 342             desc=f"epoch-{split}",
 343         ):
 344             yield batch
 345
 346     def vocabulary_size(self):
 347         return self.nb_codes
 348
 349     def compute_error(self, model, split="train", nb_to_use=-1):
 350         nb_total, nb_correct = 0, 0
 351         for input in task.batches(split, nb_to_use):
 352             result = input.clone()
 353             ar_mask = result.new_zeros(result.size())
 354             ar_mask[:, self.height * self.width :] = 1
 355             result *= 1 - ar_mask
 356             masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
 357             mazes, paths = self.seq2map(result)
 358             nb_correct += maze.path_correctness(mazes, paths).long().sum()
 359             nb_total += mazes.size(0)
 360
 361         return nb_total, nb_correct
 362
 363     def produce_results(self, n_epoch, model):
 364         with torch.autograd.no_grad():
 365             t = model.training
 366             model.eval()
 367
 368             train_nb_total, train_nb_correct = self.compute_error(
 369                 model, "train", nb_to_use=1000
 370             )
 371             log_string(
 372                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 373             )
 374
 375             test_nb_total, test_nb_correct = self.compute_error(
 376                 model, "test", nb_to_use=1000
 377             )
 378             log_string(
 379                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 380             )
 381
 382             input = self.test_input[:32]
 383             result = input.clone()
 384             ar_mask = result.new_zeros(result.size())
 385             ar_mask[:, self.height * self.width :] = 1
 386             result *= 1 - ar_mask
 387             masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
 388
 389             mazes, paths = self.seq2map(input)
 390             _, predicted_paths = self.seq2map(result)
 391             maze.save_image(
 392                 os.path.join(args.result_dir, f"result_{n_epoch:04d}.png"),
 393                 mazes=mazes,
 394                 target_paths=paths,
 395                 predicted_paths=predicted_paths,
 396                 path_correct=maze.path_correctness(mazes, predicted_paths),
 397             )
 398
 399             model.train(t)
 400
 401
 402 ######################################################################
 403
 404 log_string(f"device {device}")
 405
 406
 407 task = TaskMaze(
 408     nb_train_samples=args.nb_train_samples,
 409     nb_test_samples=args.nb_test_samples,
 410     batch_size=args.batch_size,
 411     height=args.maze_height,
 412     width=args.maze_width,
 413     nb_walls=args.maze_nb_walls,
 414     device=device,
 415 )
 416
 417
 418 vocabulary_size = task.vocabulary_size()
 419
 420 log_string(f"vocabulary_size {vocabulary_size}")
 421
 422 ##############################
 423
 424 model = mygpt.MyGPT(
 425     vocabulary_size=vocabulary_size,
 426     dim_model=args.dim_model,
 427     dim_keys=args.dim_keys,
 428     dim_hidden=args.dim_hidden,
 429     nb_heads=args.nb_heads,
 430     nb_blocks=args.nb_blocks,
 431     causal=True,
 432     dropout=args.dropout,
 433 )
 434
 435 model.to(device)
 436
 437 nb_parameters = sum(p.numel() for p in model.parameters())
 438 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 439
 440 ######################################################################
 441
 442 nb_epochs_finished = 0
 443
 444 if args.no_checkpoint:
 445     log_string(f"not trying to load checkpoint.")
 446
 447 else:
 448     try:
 449         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 450         checkpoint = torch.load(checkpoint_name)
 451         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 452         model.load_state_dict(checkpoint["model_state"])
 453         torch.set_rng_state(checkpoint["rng_state"])
 454         if torch.cuda.is_available():
 455             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 456
 457         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 458
 459     except FileNotFoundError:
 460         log_string("starting from scratch.")
 461
 462     except:
 463         log_string("error when loading the checkpoint.")
 464         exit(1)
 465
 466 ######################################################################
 467
 468 token_count = 0
 469 for input in task.batches(split="train"):
 470     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 471 token_probas = token_count / token_count.sum()
 472 entropy = -torch.xlogy(token_probas, token_probas).sum()
 473 train_set_perplexity = math.exp(entropy)
 474
 475 ##############################
 476
 477 if args.learning_rate_schedule == "cos":
 478     learning_rate_schedule = {}
 479     for n_epoch in range(args.nb_epochs):
 480         u = n_epoch / args.nb_epochs * math.pi
 481         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 482 else:
 483     u = {
 484         int(k): float(v)
 485         for k, v in [
 486             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 487         ]
 488     }
 489
 490     learning_rate_schedule = {}
 491     learning_rate = args.learning_rate
 492     for n_epoch in range(args.nb_epochs):
 493         if n_epoch in u:
 494             learning_rate = u[n_epoch]
 495         learning_rate_schedule[n_epoch] = learning_rate
 496
 497 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 498
 499 ##############################
 500
 501 if args.one_shot:
 502     one_shot(model, task)
 503     exit(0)
 504
 505 ##############################
 506
 507 if nb_epochs_finished >= args.nb_epochs:
 508     n_epoch = nb_epochs_finished
 509     train_perplexity = compute_perplexity(model, split="train")
 510     test_perplexity = compute_perplexity(model, split="test")
 511
 512     log_string(
 513         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 514     )
 515
 516     task.produce_results(n_epoch, model)
 517
 518     exit(0)
 519
 520 ##############################
 521
 522 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 523     learning_rate = learning_rate_schedule[n_epoch]
 524
 525     log_string(f"learning_rate {learning_rate}")
 526
 527     if args.optim == "sgd":
 528         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 529     elif args.optim == "adam":
 530         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 531     elif args.optim == "adamw":
 532         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 533     else:
 534         raise ValueError(f"Unknown optimizer {args.optim}.")
 535
 536     model.train()
 537
 538     nb_train_samples, acc_train_loss = 0, 0.0
 539
 540     for input in task.batches(split="train"):
 541         input = input.to(device)
 542         output = model(mygpt.BracketedSequence(input)).x
 543         loss = F.cross_entropy(output.transpose(1, 2), input)
 544         acc_train_loss += loss.item() * input.size(0)
 545         nb_train_samples += input.size(0)
 546
 547         optimizer.zero_grad()
 548         loss.backward()
 549         optimizer.step()
 550
 551     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 552     test_perplexity = compute_perplexity(model, split="test")
 553
 554     log_string(
 555         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 556     )
 557
 558     task.produce_results(n_epoch, model)
 559
 560     checkpoint = {
 561         "nb_epochs_finished": n_epoch + 1,
 562         "model_state": model.state_dict(),
 563         "rng_state": torch.get_rng_state(),
 564     }
 565
 566     if torch.cuda.is_available():
 567         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 568
 569     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 570     torch.save(checkpoint, checkpoint_name)
 571     log_string(f"saved checkpoint {checkpoint_name}")
 572
 573 ######################################################################