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("--checkpoint_name", type=str, default="checkpoint.pth")
  72
  73 ##############################
  74 # maze options
  75
  76 parser.add_argument("--maze_height", type=int, default=13)
  77
  78 parser.add_argument("--maze_width", type=int, default=21)
  79
  80 parser.add_argument("--maze_nb_walls", type=int, default=15)
  81
  82 parser.add_argument("--oneshot", action="store_true", default=False)
  83
  84 parser.add_argument("--oneshot_input", type=str, default="head")
  85
  86 parser.add_argument("--oneshot_output", type=str, default="policy")
  87
  88 ######################################################################
  89
  90 args = parser.parse_args()
  91
  92 try:
  93     os.mkdir(args.result_dir)
  94 except FileExistsError:
  95     if not args.overwrite_results:
  96         print(f"result directory {args.result_dir} already exists")
  97         exit(1)
  98
  99 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 100
 101 if args.seed >= 0:
 102     # torch.backends.cudnn.deterministic = True
 103     # torch.backends.cudnn.benchmark = False
 104     # torch.use_deterministic_algorithms(True)
 105     torch.manual_seed(args.seed)
 106     if torch.cuda.is_available():
 107         torch.cuda.manual_seed_all(args.seed)
 108
 109 ######################################################################
 110
 111
 112 def log_string(s):
 113     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 114
 115     if log_file is not None:
 116         log_file.write(t + s + "\n")
 117         log_file.flush()
 118
 119     print(t + s)
 120     sys.stdout.flush()
 121
 122
 123 for n in vars(args):
 124     log_string(f"args.{n} {getattr(args, n)}")
 125
 126 ######################################################################
 127
 128
 129 # ar_mask is a Boolean matrix of same shape as input, with 1s on the
 130 # tokens that should be generated
 131
 132
 133 def masked_inplace_autoregression(model, batch_size, input, ar_mask):
 134     for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
 135         i = (ar_mask.sum(0) > 0).nonzero()
 136         if i.min() > 0:
 137             # Needed to initialize the model's cache
 138             model(mygpt.BracketedSequence(input, 0, i.min()))
 139         for s in range(i.min(), i.max() + 1):
 140             output = model(mygpt.BracketedSequence(input, s, 1)).x
 141             logits = output[:, s]
 142             if args.deterministic_synthesis:
 143                 t_next = logits.argmax(1)
 144             else:
 145                 dist = torch.distributions.categorical.Categorical(logits=logits)
 146                 t_next = dist.sample()
 147             input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 148
 149
 150 ######################################################################
 151
 152
 153 def compute_perplexity(model, split="train"):
 154     with torch.autograd.no_grad():
 155         t = model.training
 156         model.eval()
 157
 158         nb_samples, acc_loss = 0, 0.0
 159
 160         for input in task.batches(split=split):
 161             input = input.to(device)
 162
 163             output = model(mygpt.BracketedSequence(input)).x
 164             loss = F.cross_entropy(output.transpose(1, 2), input)
 165             acc_loss += loss.item() * input.size(0)
 166             nb_samples += input.size(0)
 167
 168         model.train(t)
 169
 170         return math.exp(min(100, acc_loss / nb_samples))
 171
 172
 173 ######################################################################
 174
 175
 176 def oneshot_policy_loss(mazes, output, policies, height, width):
 177     masks = (mazes == maze.v_empty).unsqueeze(-1)
 178     targets = policies.permute(0, 2, 1) * masks
 179     output = output * masks
 180     return -(output.log_softmax(-1) * targets).sum() / masks.sum()
 181
 182
 183 def oneshot_trace_loss(mazes, output, policies, height, width):
 184     masks = mazes == maze.v_empty
 185     targets = maze.stationary_densities(
 186         mazes.view(-1, height, width), policies.view(-1, 4, height, width)
 187     ).flatten(-2)
 188     targets = targets * masks
 189     output = output.squeeze(-1) * masks
 190     return (output - targets).abs().sum() / masks.sum()
 191
 192
 193 def oneshot(gpt, task):
 194     t = gpt.training
 195     gpt.eval()
 196
 197     if args.oneshot_input == "head":
 198         dim_in = args.dim_model
 199     elif args.oneshot_input == "deep":
 200         dim_in = args.dim_model * args.nb_blocks * 2
 201     else:
 202         raise ValueError(f"{args.oneshot_input=}")
 203
 204     if args.oneshot_output == "policy":
 205         dim_out = 4
 206         compute_loss = oneshot_policy_loss
 207     elif args.oneshot_output == "trace":
 208         dim_out = 1
 209         compute_loss = oneshot_trace_loss
 210     else:
 211         raise ValueError(f"{args.oneshot_output=}")
 212
 213     model = nn.Sequential(
 214         nn.Linear(dim_in, args.dim_model),
 215         nn.ReLU(),
 216         nn.Linear(args.dim_model, args.dim_model),
 217         nn.ReLU(),
 218         nn.Linear(args.dim_model, dim_out),
 219     ).to(device)
 220
 221     for n_epoch in range(args.nb_epochs):
 222         learning_rate = learning_rate_schedule[n_epoch]
 223         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 224
 225         acc_train_loss, nb_train_samples = 0, 0
 226         for mazes, policies in task.policy_batches(split="train"):
 227             ####
 228             # print(f'{mazes.size()=} {policies.size()=}')
 229             # s = maze.stationary_densities(
 230             # exit(0)
 231             ####
 232             masks = mazes == maze.v_empty
 233             output_gpt = gpt(mygpt.BracketedSequence(mazes), mode=args.oneshot_input).x
 234             output = model(output_gpt)
 235
 236             loss = compute_loss(mazes, output, policies, task.height, task.width)
 237             acc_train_loss += loss.item() * mazes.size(0)
 238             nb_train_samples += mazes.size(0)
 239
 240             optimizer.zero_grad()
 241             loss.backward()
 242             optimizer.step()
 243
 244         acc_test_loss, nb_test_samples = 0, 0
 245         for mazes, policies in task.policy_batches(split="test"):
 246             output_gpt = gpt(mygpt.BracketedSequence(mazes), mode=args.oneshot_input).x
 247             output = model(output_gpt)
 248             loss = compute_loss(mazes, output, policies, task.height, task.width)
 249             acc_test_loss += loss.item() * mazes.size(0)
 250             nb_test_samples += mazes.size(0)
 251
 252         log_string(
 253             f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
 254         )
 255
 256         # -------------------
 257         mazes = task.test_input[:32, : task.height * task.width]
 258         policies = task.test_policies[:32]
 259         output_gpt = gpt(mygpt.BracketedSequence(mazes), mode=args.oneshot_input).x
 260         output = model(output_gpt)
 261         if args.oneshot_output == "policy":
 262             targets = policies.permute(0, 2, 1)
 263             scores = (
 264                 (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
 265             ).float()
 266         elif args.oneshot_output == "trace":
 267             targets = maze.stationary_densities(
 268                 mazes.view(-1, task.height, task.width),
 269                 policies.view(-1, 4, task.height, task.width),
 270             ).flatten(-2)
 271             scores = output.flatten(-2)
 272         else:
 273             raise ValueError(f"{args.oneshot_output=}")
 274
 275         scores = scores.reshape(-1, task.height, task.width)
 276         mazes = mazes.reshape(-1, task.height, task.width)
 277         targets = targets.reshape(-1, task.height, task.width)
 278         maze.save_image(
 279             os.path.join(
 280                 args.result_dir,
 281                 f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png",
 282             ),
 283             mazes=mazes,
 284             score_paths=scores,
 285             score_truth=targets,
 286         )
 287         # -------------------
 288
 289     gpt.train(t)
 290
 291
 292 ######################################################################
 293
 294
 295 class Task:
 296     def batches(self, split="train"):
 297         pass
 298
 299     def vocabulary_size(self):
 300         pass
 301
 302     def produce_results(self, n_epoch, model):
 303         pass
 304
 305
 306 ######################################################################
 307
 308 import maze
 309
 310
 311 class TaskMaze(Task):
 312     def map2seq(self, *m):
 313         return torch.cat([x.flatten(1) for x in m], 1)
 314
 315     def seq2map(self, s):
 316         s = s.reshape(s.size(0), -1, self.height, self.width)
 317         return (s[:, k] for k in range(s.size(1)))
 318
 319     def __init__(
 320         self,
 321         nb_train_samples,
 322         nb_test_samples,
 323         batch_size,
 324         height,
 325         width,
 326         nb_walls,
 327         device=torch.device("cpu"),
 328     ):
 329         self.batch_size = batch_size
 330         self.height = height
 331         self.width = width
 332         self.device = device
 333
 334         train_mazes, train_paths, train_policies = maze.create_maze_data(
 335             nb_train_samples,
 336             height=height,
 337             width=width,
 338             nb_walls=nb_walls,
 339             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 340         )
 341         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 342         self.train_policies = train_policies.flatten(-2).to(device)
 343
 344         test_mazes, test_paths, test_policies = maze.create_maze_data(
 345             nb_test_samples,
 346             height=height,
 347             width=width,
 348             nb_walls=nb_walls,
 349             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 350         )
 351         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 352         self.test_policies = test_policies.flatten(-2).to(device)
 353
 354         self.nb_codes = self.train_input.max() + 1
 355
 356     def batches(self, split="train", nb_to_use=-1):
 357         assert split in {"train", "test"}
 358         input = self.train_input if split == "train" else self.test_input
 359         if nb_to_use > 0:
 360             input = input[:nb_to_use]
 361         for batch in tqdm.tqdm(
 362             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 363         ):
 364             yield batch
 365
 366     def policy_batches(self, split="train", nb_to_use=-1):
 367         assert split in {"train", "test"}
 368         input = self.train_input if split == "train" else self.test_input
 369         policies = self.train_policies if split == "train" else self.test_policies
 370         input = input[:, : self.height * self.width]
 371         policies = policies * (input != maze.v_wall)[:, None]
 372
 373         if nb_to_use > 0:
 374             input = input[:nb_to_use]
 375             policies = policies[:nb_to_use]
 376
 377         for batch in tqdm.tqdm(
 378             zip(input.split(self.batch_size), policies.split(self.batch_size)),
 379             dynamic_ncols=True,
 380             desc=f"epoch-{split}",
 381         ):
 382             yield batch
 383
 384     def vocabulary_size(self):
 385         return self.nb_codes
 386
 387     def compute_error(self, model, split="train", nb_to_use=-1):
 388         nb_total, nb_correct = 0, 0
 389         for input in task.batches(split, nb_to_use):
 390             result = input.clone()
 391             ar_mask = result.new_zeros(result.size())
 392             ar_mask[:, self.height * self.width :] = 1
 393             result *= 1 - ar_mask
 394             masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
 395             mazes, paths = self.seq2map(result)
 396             nb_correct += maze.path_correctness(mazes, paths).long().sum()
 397             nb_total += mazes.size(0)
 398
 399         return nb_total, nb_correct
 400
 401     def produce_results(self, n_epoch, model):
 402         with torch.autograd.no_grad():
 403             t = model.training
 404             model.eval()
 405
 406             train_nb_total, train_nb_correct = self.compute_error(
 407                 model, "train", nb_to_use=1000
 408             )
 409             log_string(
 410                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 411             )
 412
 413             test_nb_total, test_nb_correct = self.compute_error(
 414                 model, "test", nb_to_use=1000
 415             )
 416             log_string(
 417                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 418             )
 419
 420             input = self.test_input[:32]
 421             result = input.clone()
 422             ar_mask = result.new_zeros(result.size())
 423             ar_mask[:, self.height * self.width :] = 1
 424             result *= 1 - ar_mask
 425             masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
 426
 427             mazes, paths = self.seq2map(input)
 428             _, predicted_paths = self.seq2map(result)
 429             maze.save_image(
 430                 os.path.join(args.result_dir, f"result_{n_epoch:04d}.png"),
 431                 mazes=mazes,
 432                 target_paths=paths,
 433                 predicted_paths=predicted_paths,
 434                 path_correct=maze.path_correctness(mazes, predicted_paths),
 435             )
 436
 437             model.train(t)
 438
 439
 440 ######################################################################
 441
 442 log_string(f"device {device}")
 443
 444
 445 task = TaskMaze(
 446     nb_train_samples=args.nb_train_samples,
 447     nb_test_samples=args.nb_test_samples,
 448     batch_size=args.batch_size,
 449     height=args.maze_height,
 450     width=args.maze_width,
 451     nb_walls=args.maze_nb_walls,
 452     device=device,
 453 )
 454
 455
 456 vocabulary_size = task.vocabulary_size()
 457
 458 log_string(f"vocabulary_size {vocabulary_size}")
 459
 460 ##############################
 461
 462 model = mygpt.MyGPT(
 463     vocabulary_size=vocabulary_size,
 464     dim_model=args.dim_model,
 465     dim_keys=args.dim_keys,
 466     dim_hidden=args.dim_hidden,
 467     nb_heads=args.nb_heads,
 468     nb_blocks=args.nb_blocks,
 469     causal=True,
 470     dropout=args.dropout,
 471 )
 472
 473 model.to(device)
 474
 475 nb_parameters = sum(p.numel() for p in model.parameters())
 476 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 477
 478 ######################################################################
 479
 480 nb_epochs_finished = 0
 481
 482 if args.no_checkpoint:
 483     log_string(f"not trying to load checkpoint.")
 484
 485 else:
 486     try:
 487         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 488         checkpoint = torch.load(checkpoint_name)
 489         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 490         model.load_state_dict(checkpoint["model_state"])
 491         torch.set_rng_state(checkpoint["rng_state"])
 492         if torch.cuda.is_available():
 493             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 494
 495         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 496
 497     except FileNotFoundError:
 498         log_string("starting from scratch.")
 499
 500     except:
 501         log_string("error when loading the checkpoint.")
 502         exit(1)
 503
 504 ######################################################################
 505
 506 token_count = 0
 507 for input in task.batches(split="train"):
 508     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 509 token_probas = token_count / token_count.sum()
 510 entropy = -torch.xlogy(token_probas, token_probas).sum()
 511 train_set_perplexity = math.exp(entropy)
 512
 513 ##############################
 514
 515 if args.learning_rate_schedule == "cos":
 516     learning_rate_schedule = {}
 517     for n_epoch in range(args.nb_epochs):
 518         u = n_epoch / args.nb_epochs * math.pi
 519         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 520 else:
 521     u = {
 522         int(k): float(v)
 523         for k, v in [
 524             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 525         ]
 526     }
 527
 528     learning_rate_schedule = {}
 529     learning_rate = args.learning_rate
 530     for n_epoch in range(args.nb_epochs):
 531         if n_epoch in u:
 532             learning_rate = u[n_epoch]
 533         learning_rate_schedule[n_epoch] = learning_rate
 534
 535 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 536
 537 ##############################
 538
 539 if args.oneshot:
 540     oneshot(model, task)
 541     exit(0)
 542
 543 ##############################
 544
 545 if nb_epochs_finished >= args.nb_epochs:
 546     n_epoch = nb_epochs_finished
 547     train_perplexity = compute_perplexity(model, split="train")
 548     test_perplexity = compute_perplexity(model, split="test")
 549
 550     log_string(
 551         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 552     )
 553
 554     task.produce_results(n_epoch, model)
 555
 556     exit(0)
 557
 558 ##############################
 559
 560 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 561     learning_rate = learning_rate_schedule[n_epoch]
 562
 563     log_string(f"learning_rate {learning_rate}")
 564
 565     if args.optim == "sgd":
 566         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 567     elif args.optim == "adam":
 568         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 569     elif args.optim == "adamw":
 570         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 571     else:
 572         raise ValueError(f"{args.optim=}")
 573
 574     model.train()
 575
 576     nb_train_samples, acc_train_loss = 0, 0.0
 577
 578     for input in task.batches(split="train"):
 579         input = input.to(device)
 580         output = model(mygpt.BracketedSequence(input)).x
 581         loss = F.cross_entropy(output.transpose(1, 2), input)
 582         acc_train_loss += loss.item() * input.size(0)
 583         nb_train_samples += input.size(0)
 584
 585         optimizer.zero_grad()
 586         loss.backward()
 587         optimizer.step()
 588
 589     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 590     test_perplexity = compute_perplexity(model, split="test")
 591
 592     log_string(
 593         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 594     )
 595
 596     task.produce_results(n_epoch, model)
 597
 598     checkpoint = {
 599         "nb_epochs_finished": n_epoch + 1,
 600         "model_state": model.state_dict(),
 601         "rng_state": torch.get_rng_state(),
 602     }
 603
 604     if torch.cuda.is_available():
 605         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 606
 607     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 608     torch.save(checkpoint, checkpoint_name)
 609     log_string(f"saved checkpoint {checkpoint_name}")
 610
 611 ######################################################################