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