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