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 nb_rank_error(output, targets):
 173     output = output.reshape(-1, output.size(-1))
 174     targets = targets.reshape(-1, targets.size(-1))
 175     i = outputs.argmax(1)
 176     # out=input.gather out[i][j]=input[i][index[i][j]]
 177     # u[k]=targets[k][i[k]]
 178     return output[targets.argmax(1)]
 179
 180
 181 def one_shot(gpt, task):
 182     t = gpt.training
 183     gpt.eval()
 184     model = nn.Linear(args.dim_model, 4).to(device)
 185
 186     for n_epoch in range(args.nb_epochs):
 187         optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
 188
 189         acc_train_loss, nb_train_samples = 0, 0
 190         for input, targets in task.policy_batches(split="train"):
 191             output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
 192             output = model(output_gpt)
 193             loss = -(output.log_softmax(-1) * targets).sum(-1).mean()
 194             acc_train_loss += loss.item() * input.size(0)
 195             nb_train_samples += input.size(0)
 196
 197             optimizer.zero_grad()
 198             loss.backward()
 199             optimizer.step()
 200
 201         acc_test_loss, nb_test_samples = 0, 0
 202         for input, targets in task.policy_batches(split="test"):
 203             output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
 204             output = model(output_gpt)
 205             loss = -(output.log_softmax(-1) * targets).sum(-1).mean()
 206             acc_test_loss += loss.item() * input.size(0)
 207             nb_test_samples += input.size(0)
 208
 209         print(
 210             f"{n_epoch=} {acc_train_loss/nb_train_samples=} {acc_test_loss/nb_test_samples=}"
 211         )
 212
 213     gpt.train(t)
 214
 215
 216 ######################################################################
 217
 218
 219 class Task:
 220     def batches(self, split="train"):
 221         pass
 222
 223     def vocabulary_size(self):
 224         pass
 225
 226     def produce_results(self, n_epoch, model):
 227         pass
 228
 229
 230 ######################################################################
 231
 232 import maze
 233
 234
 235 class TaskMaze(Task):
 236     def map2seq(self, *m):
 237         return torch.cat([x.flatten(1) for x in m], 1)
 238
 239     def seq2map(self, s):
 240         s = s.reshape(s.size(0), -1, self.height, self.width)
 241         return (s[:, k] for k in range(s.size(1)))
 242
 243     def __init__(
 244         self,
 245         nb_train_samples,
 246         nb_test_samples,
 247         batch_size,
 248         height,
 249         width,
 250         nb_walls,
 251         device=torch.device("cpu"),
 252     ):
 253         self.batch_size = batch_size
 254         self.height = height
 255         self.width = width
 256         self.device = device
 257
 258         train_mazes, train_paths, train_policies = maze.create_maze_data(
 259             nb_train_samples,
 260             height=height,
 261             width=width,
 262             nb_walls=nb_walls,
 263             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 264         )
 265         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 266         self.train_policies = train_policies.flatten(-2).permute(0, 2, 1).to(device)
 267
 268         test_mazes, test_paths, test_policies = maze.create_maze_data(
 269             nb_test_samples,
 270             height=height,
 271             width=width,
 272             nb_walls=nb_walls,
 273             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 274         )
 275         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 276         self.test_policies = test_policies.flatten(-2).permute(0, 2, 1).to(device)
 277
 278         self.nb_codes = self.train_input.max() + 1
 279
 280     def batches(self, split="train", nb_to_use=-1):
 281         assert split in {"train", "test"}
 282         input = self.train_input if split == "train" else self.test_input
 283         if nb_to_use > 0:
 284             input = input[:nb_to_use]
 285         for batch in tqdm.tqdm(
 286             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 287         ):
 288             yield batch
 289
 290     def policy_batches(self, split="train", nb_to_use=-1):
 291         assert split in {"train", "test"}
 292         input = self.train_input if split == "train" else self.test_input
 293         targets = self.train_policies if split == "train" else self.test_policies
 294         input = input[:, : self.height * self.width]
 295         targets = targets * (input != maze.v_wall)[:, :, None]
 296
 297         if nb_to_use > 0:
 298             input = input[:nb_to_use]
 299             targets = targets[:nb_to_use]
 300
 301         for batch in tqdm.tqdm(
 302             zip(input.split(self.batch_size), targets.split(self.batch_size)),
 303             dynamic_ncols=True,
 304             desc=f"epoch-{split}",
 305         ):
 306             yield batch
 307
 308     def vocabulary_size(self):
 309         return self.nb_codes
 310
 311     def compute_error(self, model, split="train", nb_to_use=-1):
 312         nb_total, nb_correct = 0, 0
 313         for input in task.batches(split, nb_to_use):
 314             result = input.clone()
 315             ar_mask = result.new_zeros(result.size())
 316             ar_mask[:, self.height * self.width :] = 1
 317             result *= 1 - ar_mask
 318             masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
 319             mazes, paths = self.seq2map(result)
 320             nb_correct += maze.path_correctness(mazes, paths).long().sum()
 321             nb_total += mazes.size(0)
 322
 323         return nb_total, nb_correct
 324
 325     def produce_results(self, n_epoch, model):
 326         with torch.autograd.no_grad():
 327             t = model.training
 328             model.eval()
 329
 330             train_nb_total, train_nb_correct = self.compute_error(
 331                 model, "train", nb_to_use=1000
 332             )
 333             log_string(
 334                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 335             )
 336
 337             test_nb_total, test_nb_correct = self.compute_error(
 338                 model, "test", nb_to_use=1000
 339             )
 340             log_string(
 341                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 342             )
 343
 344             input = self.test_input[:32]
 345             result = input.clone()
 346             ar_mask = result.new_zeros(result.size())
 347             ar_mask[:, self.height * self.width :] = 1
 348             result *= 1 - ar_mask
 349             masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
 350
 351             mazes, paths = self.seq2map(input)
 352             _, predicted_paths = self.seq2map(result)
 353             maze.save_image(
 354                 os.path.join(args.result_dir, f"result_{n_epoch:04d}.png"),
 355                 mazes,
 356                 paths,
 357                 predicted_paths,
 358                 maze.path_correctness(mazes, predicted_paths),
 359             )
 360
 361             model.train(t)
 362
 363
 364 ######################################################################
 365
 366 log_string(f"device {device}")
 367
 368
 369 task = TaskMaze(
 370     nb_train_samples=args.nb_train_samples,
 371     nb_test_samples=args.nb_test_samples,
 372     batch_size=args.batch_size,
 373     height=args.maze_height,
 374     width=args.maze_width,
 375     nb_walls=args.maze_nb_walls,
 376     device=device,
 377 )
 378
 379
 380 vocabulary_size = task.vocabulary_size()
 381
 382 log_string(f"vocabulary_size {vocabulary_size}")
 383
 384 ##############################
 385
 386 model = mygpt.MyGPT(
 387     vocabulary_size=vocabulary_size,
 388     dim_model=args.dim_model,
 389     dim_keys=args.dim_keys,
 390     dim_hidden=args.dim_hidden,
 391     nb_heads=args.nb_heads,
 392     nb_blocks=args.nb_blocks,
 393     causal=True,
 394     dropout=args.dropout,
 395 )
 396
 397 model.to(device)
 398
 399 nb_parameters = sum(p.numel() for p in model.parameters())
 400 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 401
 402 ######################################################################
 403
 404 nb_epochs_finished = 0
 405
 406 if args.no_checkpoint:
 407     log_string(f"not trying to load checkpoint.")
 408
 409 else:
 410     try:
 411         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 412         checkpoint = torch.load(checkpoint_name)
 413         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 414         model.load_state_dict(checkpoint["model_state"])
 415         torch.set_rng_state(checkpoint["rng_state"])
 416         if torch.cuda.is_available():
 417             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 418
 419         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 420
 421     except FileNotFoundError:
 422         log_string("starting from scratch.")
 423
 424     except:
 425         log_string("error when loading the checkpoint.")
 426         exit(1)
 427
 428 ######################################################################
 429
 430 token_count = 0
 431 for input in task.batches(split="train"):
 432     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 433 token_probas = token_count / token_count.sum()
 434 entropy = -torch.xlogy(token_probas, token_probas).sum()
 435 train_set_perplexity = math.exp(entropy)
 436
 437 ##############################
 438
 439 if args.learning_rate_schedule == "cos":
 440     learning_rate_schedule = {}
 441     for n_epoch in range(args.nb_epochs):
 442         u = n_epoch / args.nb_epochs * math.pi
 443         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 444 else:
 445     u = {
 446         int(k): float(v)
 447         for k, v in [
 448             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 449         ]
 450     }
 451
 452     learning_rate_schedule = {}
 453     learning_rate = args.learning_rate
 454     for n_epoch in range(args.nb_epochs):
 455         if n_epoch in u:
 456             learning_rate = u[n_epoch]
 457         learning_rate_schedule[n_epoch] = learning_rate
 458
 459 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 460
 461 ##############################
 462
 463 if args.one_shot:
 464     one_shot(model, task)
 465     exit(0)
 466
 467 ##############################
 468
 469 if nb_epochs_finished >= args.nb_epochs:
 470     n_epoch = nb_epochs_finished
 471     train_perplexity = compute_perplexity(model, split="train")
 472     test_perplexity = compute_perplexity(model, split="test")
 473
 474     log_string(
 475         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 476     )
 477
 478     task.produce_results(n_epoch, model)
 479
 480     exit(0)
 481
 482 ##############################
 483
 484 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 485     learning_rate = learning_rate_schedule[n_epoch]
 486
 487     log_string(f"learning_rate {learning_rate}")
 488
 489     if args.optim == "sgd":
 490         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 491     elif args.optim == "adam":
 492         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 493     elif args.optim == "adamw":
 494         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 495     else:
 496         raise ValueError(f"Unknown optimizer {args.optim}.")
 497
 498     model.train()
 499
 500     nb_train_samples, acc_train_loss = 0, 0.0
 501
 502     for input in task.batches(split="train"):
 503         input = input.to(device)
 504         output = model(mygpt.BracketedSequence(input)).x
 505         loss = F.cross_entropy(output.transpose(1, 2), input)
 506         acc_train_loss += loss.item() * input.size(0)
 507         nb_train_samples += input.size(0)
 508
 509         optimizer.zero_grad()
 510         loss.backward()
 511         optimizer.step()
 512
 513     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 514     test_perplexity = compute_perplexity(model, split="test")
 515
 516     log_string(
 517         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 518     )
 519
 520     task.produce_results(n_epoch, model)
 521
 522     checkpoint = {
 523         "nb_epochs_finished": n_epoch + 1,
 524         "model_state": model.state_dict(),
 525         "rng_state": torch.get_rng_state(),
 526     }
 527
 528     if torch.cuda.is_available():
 529         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 530
 531     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 532     torch.save(checkpoint, checkpoint_name)
 533     log_string(f"saved checkpoint {checkpoint_name}")
 534
 535 ######################################################################