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, 4),
 179     ).to(device)
 180
 181     print(f"{args.nb_epochs=}")
 182
 183     for n_epoch in range(args.nb_epochs):
 184         print(f"{n_epoch=}")
 185         learning_rate = learning_rate_schedule[n_epoch]
 186         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 187
 188         acc_train_loss, nb_train_samples = 0, 0
 189         for input, targets in task.policy_batches(split="train"):
 190             output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
 191             output = model(output_gpt)
 192             loss = (
 193                 -(output.log_softmax(-1) * targets).sum(-1).mean()
 194                 + targets.xlogy(targets).sum(-1).mean()
 195             )
 196             acc_train_loss += loss.item() * input.size(0)
 197             nb_train_samples += input.size(0)
 198
 199             optimizer.zero_grad()
 200             loss.backward()
 201             optimizer.step()
 202
 203         acc_test_loss, nb_test_samples = 0, 0
 204         for input, targets in task.policy_batches(split="test"):
 205             output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
 206             output = model(output_gpt)
 207             loss = (
 208                 -(output.log_softmax(-1) * targets).sum(-1).mean()
 209                 + targets.xlogy(targets).sum(-1).mean()
 210             )
 211             acc_test_loss += loss.item() * input.size(0)
 212             nb_test_samples += input.size(0)
 213
 214         log_string(
 215             f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
 216         )
 217
 218         # -------------------
 219         input, targets = next(task.policy_batches(split="test"))
 220         output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
 221         output = model(output_gpt)
 222         losses = (-output.log_softmax(-1) * targets + targets.xlogy(targets)).sum(-1)
 223         losses = losses / losses.max()
 224         print(f"{input.size()=} {losses.size()=} {losses.min()=} {losses.max()=}")
 225         losses = losses * (input == 0)
 226         losses = losses.reshape(-1, args.maze_height, args.maze_width)
 227         input = input.reshape(-1, args.maze_height, args.maze_width)
 228         maze.save_image(
 229             os.path.join(args.result_dir, f"oneshot_{n_epoch:04d}.png"),
 230             mazes=input,
 231             score_paths=losses,
 232         )
 233         # -------------------
 234
 235     gpt.train(t)
 236
 237
 238 ######################################################################
 239
 240
 241 class Task:
 242     def batches(self, split="train"):
 243         pass
 244
 245     def vocabulary_size(self):
 246         pass
 247
 248     def produce_results(self, n_epoch, model):
 249         pass
 250
 251
 252 ######################################################################
 253
 254 import maze
 255
 256
 257 class TaskMaze(Task):
 258     def map2seq(self, *m):
 259         return torch.cat([x.flatten(1) for x in m], 1)
 260
 261     def seq2map(self, s):
 262         s = s.reshape(s.size(0), -1, self.height, self.width)
 263         return (s[:, k] for k in range(s.size(1)))
 264
 265     def __init__(
 266         self,
 267         nb_train_samples,
 268         nb_test_samples,
 269         batch_size,
 270         height,
 271         width,
 272         nb_walls,
 273         device=torch.device("cpu"),
 274     ):
 275         self.batch_size = batch_size
 276         self.height = height
 277         self.width = width
 278         self.device = device
 279
 280         train_mazes, train_paths, train_policies = maze.create_maze_data(
 281             nb_train_samples,
 282             height=height,
 283             width=width,
 284             nb_walls=nb_walls,
 285             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 286         )
 287         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 288         self.train_policies = train_policies.flatten(-2).permute(0, 2, 1).to(device)
 289
 290         test_mazes, test_paths, test_policies = maze.create_maze_data(
 291             nb_test_samples,
 292             height=height,
 293             width=width,
 294             nb_walls=nb_walls,
 295             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 296         )
 297         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 298         self.test_policies = test_policies.flatten(-2).permute(0, 2, 1).to(device)
 299
 300         self.nb_codes = self.train_input.max() + 1
 301
 302     def batches(self, split="train", nb_to_use=-1):
 303         assert split in {"train", "test"}
 304         input = self.train_input if split == "train" else self.test_input
 305         if nb_to_use > 0:
 306             input = input[:nb_to_use]
 307         for batch in tqdm.tqdm(
 308             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 309         ):
 310             yield batch
 311
 312     def policy_batches(self, split="train", nb_to_use=-1):
 313         assert split in {"train", "test"}
 314         input = self.train_input if split == "train" else self.test_input
 315         targets = self.train_policies if split == "train" else self.test_policies
 316         input = input[:, : self.height * self.width]
 317         targets = targets * (input != maze.v_wall)[:, :, None]
 318
 319         if nb_to_use > 0:
 320             input = input[:nb_to_use]
 321             targets = targets[:nb_to_use]
 322
 323         for batch in tqdm.tqdm(
 324             zip(input.split(self.batch_size), targets.split(self.batch_size)),
 325             dynamic_ncols=True,
 326             desc=f"epoch-{split}",
 327         ):
 328             yield batch
 329
 330     def vocabulary_size(self):
 331         return self.nb_codes
 332
 333     def compute_error(self, model, split="train", nb_to_use=-1):
 334         nb_total, nb_correct = 0, 0
 335         for input in task.batches(split, nb_to_use):
 336             result = input.clone()
 337             ar_mask = result.new_zeros(result.size())
 338             ar_mask[:, self.height * self.width :] = 1
 339             result *= 1 - ar_mask
 340             masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
 341             mazes, paths = self.seq2map(result)
 342             nb_correct += maze.path_correctness(mazes, paths).long().sum()
 343             nb_total += mazes.size(0)
 344
 345         return nb_total, nb_correct
 346
 347     def produce_results(self, n_epoch, model):
 348         with torch.autograd.no_grad():
 349             t = model.training
 350             model.eval()
 351
 352             train_nb_total, train_nb_correct = self.compute_error(
 353                 model, "train", nb_to_use=1000
 354             )
 355             log_string(
 356                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 357             )
 358
 359             test_nb_total, test_nb_correct = self.compute_error(
 360                 model, "test", nb_to_use=1000
 361             )
 362             log_string(
 363                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 364             )
 365
 366             input = self.test_input[:32]
 367             result = input.clone()
 368             ar_mask = result.new_zeros(result.size())
 369             ar_mask[:, self.height * self.width :] = 1
 370             result *= 1 - ar_mask
 371             masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
 372
 373             mazes, paths = self.seq2map(input)
 374             _, predicted_paths = self.seq2map(result)
 375             maze.save_image(
 376                 os.path.join(args.result_dir, f"result_{n_epoch:04d}.png"),
 377                 mazes=mazes,
 378                 target_paths=paths,
 379                 predicted_paths=predicted_paths,
 380                 path_correct=maze.path_correctness(mazes, predicted_paths),
 381             )
 382
 383             model.train(t)
 384
 385
 386 ######################################################################
 387
 388 log_string(f"device {device}")
 389
 390
 391 task = TaskMaze(
 392     nb_train_samples=args.nb_train_samples,
 393     nb_test_samples=args.nb_test_samples,
 394     batch_size=args.batch_size,
 395     height=args.maze_height,
 396     width=args.maze_width,
 397     nb_walls=args.maze_nb_walls,
 398     device=device,
 399 )
 400
 401
 402 vocabulary_size = task.vocabulary_size()
 403
 404 log_string(f"vocabulary_size {vocabulary_size}")
 405
 406 ##############################
 407
 408 model = mygpt.MyGPT(
 409     vocabulary_size=vocabulary_size,
 410     dim_model=args.dim_model,
 411     dim_keys=args.dim_keys,
 412     dim_hidden=args.dim_hidden,
 413     nb_heads=args.nb_heads,
 414     nb_blocks=args.nb_blocks,
 415     causal=True,
 416     dropout=args.dropout,
 417 )
 418
 419 model.to(device)
 420
 421 nb_parameters = sum(p.numel() for p in model.parameters())
 422 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 423
 424 ######################################################################
 425
 426 nb_epochs_finished = 0
 427
 428 if args.no_checkpoint:
 429     log_string(f"not trying to load checkpoint.")
 430
 431 else:
 432     try:
 433         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 434         checkpoint = torch.load(checkpoint_name)
 435         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 436         model.load_state_dict(checkpoint["model_state"])
 437         torch.set_rng_state(checkpoint["rng_state"])
 438         if torch.cuda.is_available():
 439             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 440
 441         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 442
 443     except FileNotFoundError:
 444         log_string("starting from scratch.")
 445
 446     except:
 447         log_string("error when loading the checkpoint.")
 448         exit(1)
 449
 450 ######################################################################
 451
 452 token_count = 0
 453 for input in task.batches(split="train"):
 454     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 455 token_probas = token_count / token_count.sum()
 456 entropy = -torch.xlogy(token_probas, token_probas).sum()
 457 train_set_perplexity = math.exp(entropy)
 458
 459 ##############################
 460
 461 if args.learning_rate_schedule == "cos":
 462     learning_rate_schedule = {}
 463     for n_epoch in range(args.nb_epochs):
 464         u = n_epoch / args.nb_epochs * math.pi
 465         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 466 else:
 467     u = {
 468         int(k): float(v)
 469         for k, v in [
 470             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 471         ]
 472     }
 473
 474     learning_rate_schedule = {}
 475     learning_rate = args.learning_rate
 476     for n_epoch in range(args.nb_epochs):
 477         if n_epoch in u:
 478             learning_rate = u[n_epoch]
 479         learning_rate_schedule[n_epoch] = learning_rate
 480
 481 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 482
 483 ##############################
 484
 485 if args.one_shot:
 486     one_shot(model, task)
 487     exit(0)
 488
 489 ##############################
 490
 491 if nb_epochs_finished >= args.nb_epochs:
 492     n_epoch = nb_epochs_finished
 493     train_perplexity = compute_perplexity(model, split="train")
 494     test_perplexity = compute_perplexity(model, split="test")
 495
 496     log_string(
 497         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 498     )
 499
 500     task.produce_results(n_epoch, model)
 501
 502     exit(0)
 503
 504 ##############################
 505
 506 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 507     learning_rate = learning_rate_schedule[n_epoch]
 508
 509     log_string(f"learning_rate {learning_rate}")
 510
 511     if args.optim == "sgd":
 512         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 513     elif args.optim == "adam":
 514         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 515     elif args.optim == "adamw":
 516         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 517     else:
 518         raise ValueError(f"Unknown optimizer {args.optim}.")
 519
 520     model.train()
 521
 522     nb_train_samples, acc_train_loss = 0, 0.0
 523
 524     for input in task.batches(split="train"):
 525         input = input.to(device)
 526         output = model(mygpt.BracketedSequence(input)).x
 527         loss = F.cross_entropy(output.transpose(1, 2), input)
 528         acc_train_loss += loss.item() * input.size(0)
 529         nb_train_samples += input.size(0)
 530
 531         optimizer.zero_grad()
 532         loss.backward()
 533         optimizer.step()
 534
 535     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 536     test_perplexity = compute_perplexity(model, split="test")
 537
 538     log_string(
 539         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 540     )
 541
 542     task.produce_results(n_epoch, model)
 543
 544     checkpoint = {
 545         "nb_epochs_finished": n_epoch + 1,
 546         "model_state": model.state_dict(),
 547         "rng_state": torch.get_rng_state(),
 548     }
 549
 550     if torch.cuda.is_available():
 551         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 552
 553     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 554     torch.save(checkpoint, checkpoint_name)
 555     log_string(f"saved checkpoint {checkpoint_name}")
 556
 557 ######################################################################