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(
  30     description="An implementation of GPT with cache to solve a toy geometric reasoning task."
  31 )
  32
  33 parser.add_argument("--log_filename", type=str, default="train.log")
  34
  35 parser.add_argument("--result_dir", type=str, default="results_default")
  36
  37 parser.add_argument("--seed", type=int, default=0)
  38
  39 parser.add_argument("--nb_epochs", type=int, default=25)
  40
  41 parser.add_argument("--batch_size", type=int, default=100)
  42
  43 parser.add_argument("--data_size", type=int, default=-1)
  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("--world_height", type=int, default=13)
  77
  78 parser.add_argument("--world_width", type=int, default=21)
  79
  80 parser.add_argument("--world_nb_walls", type=int, default=15)
  81
  82 ######################################################################
  83
  84 args = parser.parse_args()
  85
  86 try:
  87     os.mkdir(args.result_dir)
  88 except FileExistsError:
  89     if not args.overwrite_results:
  90         print(f"result directory {args.result_dir} already exists")
  91         exit(1)
  92
  93 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
  94
  95 if args.seed >= 0:
  96     # torch.backends.cudnn.deterministic = True
  97     # torch.backends.cudnn.benchmark = False
  98     # torch.use_deterministic_algorithms(True)
  99     torch.manual_seed(args.seed)
 100     if torch.cuda.is_available():
 101         torch.cuda.manual_seed_all(args.seed)
 102
 103 ######################################################################
 104
 105
 106 def log_string(s):
 107     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 108
 109     if log_file is not None:
 110         log_file.write(t + s + "\n")
 111         log_file.flush()
 112
 113     print(t + s)
 114     sys.stdout.flush()
 115
 116
 117 for n in vars(args):
 118     log_string(f"args.{n} {getattr(args, n)}")
 119
 120 ######################################################################
 121
 122
 123 # ar_mask is a Boolean matrix of same shape as input, with 1s on the
 124 # tokens that should be generated
 125
 126
 127 def masked_inplace_autoregression(model, batch_size, input, ar_mask):
 128
 129     for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
 130         i = (ar_mask.sum(0) > 0).nonzero()
 131         if i.min() > 0:
 132             model(
 133                 mygpt.BracketedSequence(input, 0, i.min())
 134             )  # Needed to initialize the model's cache
 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 class Task:
 150     def batches(self, split="train"):
 151         pass
 152
 153     def vocabulary_size(self):
 154         pass
 155
 156     def produce_results(self, n_epoch, model):
 157         pass
 158
 159
 160 ######################################################################
 161
 162 import maze
 163
 164
 165 class TaskMaze(Task):
 166     def map2seq(self, *m):
 167         return torch.cat([x.flatten(1) for x in m], 1)
 168
 169     def seq2map(self, s):
 170         s = s.reshape(s.size(0), -1, self.height, self.width)
 171         return (s[:, k] for k in range(s.size(1)))
 172
 173     def __init__(self, batch_size, height, width, nb_walls, device=torch.device("cpu")):
 174         self.batch_size = batch_size
 175         self.height = height
 176         self.width = width
 177         self.device = device
 178
 179         nb = args.data_size if args.data_size > 0 else 250000
 180
 181         mazes_train, paths_train = maze.create_maze_data(
 182             (4 * nb) // 5,
 183             height=height,
 184             width=width,
 185             nb_walls=nb_walls,
 186             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 187         )
 188         mazes_train, paths_train = mazes_train.to(device), paths_train.to(device)
 189         self.train_input = self.map2seq(mazes_train, paths_train)
 190         self.nb_codes = self.train_input.max() + 1
 191
 192         mazes_test, paths_test = maze.create_maze_data(
 193             nb // 5,
 194             height=height,
 195             width=width,
 196             nb_walls=nb_walls,
 197             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 198         )
 199         mazes_test, paths_test = mazes_test.to(device), paths_test.to(device)
 200         self.test_input = self.map2seq(mazes_test, paths_test)
 201
 202     def batches(self, split="train"):
 203         assert split in {"train", "test"}
 204         input = self.train_input if split == "train" else self.test_input
 205         for batch in tqdm.tqdm(
 206             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 207         ):
 208             yield batch
 209
 210     def vocabulary_size(self):
 211         return self.nb_codes
 212
 213     def compute_error(self, model, split="train"):
 214         nb_total, nb_correct = 0, 0
 215         for input in task.batches(split):
 216             result = input.clone()
 217             ar_mask = result.new_zeros(result.size())
 218             ar_mask[:, self.height * self.width :] = 1
 219             masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
 220             mazes, paths = self.seq2map(result)
 221             nb_correct += maze.path_correctness(mazes, paths).long().sum()
 222             nb_total += mazes.size(0)
 223
 224         return nb_total, nb_correct
 225
 226     def produce_results(self, n_epoch, model):
 227         train_nb_total, train_nb_correct = self.compute_error(model, "train")
 228         log_string(
 229             f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 230         )
 231
 232         test_nb_total, test_nb_correct = self.compute_error(model, "test")
 233         log_string(
 234             f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 235         )
 236
 237         input = self.test_input[:32]
 238         result = input.clone()
 239         ar_mask = result.new_zeros(result.size())
 240
 241         ar_mask[:, self.height * self.width :] = 1
 242         masked_inplace_autoregression(model, self.batch_size, result, ar_mask)
 243
 244         mazes, paths = self.seq2map(input)
 245         _, predicted_paths = self.seq2map(result)
 246         maze.save_image(f"result_{n_epoch:04d}.png", mazes, paths, predicted_paths)
 247
 248
 249 ######################################################################
 250
 251 log_string(f"device {device}")
 252
 253
 254 task = TaskMaze(
 255     batch_size=args.batch_size,
 256     height=args.world_height,
 257     width=args.world_width,
 258     nb_walls=args.world_nb_walls,
 259     device=device,
 260 )
 261
 262
 263 vocabulary_size = task.vocabulary_size()
 264
 265 log_string(f"vocabulary_size {vocabulary_size}")
 266
 267 ##############################
 268
 269 model = mygpt.MyGPT(
 270     vocabulary_size=vocabulary_size,
 271     dim_model=args.dim_model,
 272     dim_keys=args.dim_keys,
 273     dim_hidden=args.dim_hidden,
 274     nb_heads=args.nb_heads,
 275     nb_blocks=args.nb_blocks,
 276     causal=True,
 277     dropout=args.dropout,
 278 )
 279
 280 model.to(device)
 281
 282 nb_parameters = sum(p.numel() for p in model.parameters())
 283 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 284
 285 ######################################################################
 286
 287 nb_epochs_finished = 0
 288
 289 if args.no_checkpoint:
 290     log_string(f"not trying to load checkpoint.")
 291
 292 else:
 293     try:
 294         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 295         checkpoint = torch.load(checkpoint_name)
 296         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 297         model.load_state_dict(checkpoint["model_state"])
 298         torch.set_rng_state(checkpoint["rng_state"])
 299         if torch.cuda.is_available():
 300             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 301
 302         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 303
 304     except FileNotFoundError:
 305         log_string("starting from scratch.")
 306
 307     except:
 308         log_string("error when loading the checkpoint.")
 309         exit(1)
 310
 311 ######################################################################
 312
 313 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 314
 315 token_count = 0
 316 for input in task.batches(split="train"):
 317     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 318 token_probas = token_count / token_count.sum()
 319 entropy = -torch.xlogy(token_probas, token_probas).sum()
 320 train_set_perplexity = math.exp(entropy)
 321
 322 ##############################
 323
 324 if args.learning_rate_schedule == "cos":
 325     learning_rate_schedule = {}
 326     for n_epoch in range(args.nb_epochs):
 327         u = n_epoch / args.nb_epochs * math.pi
 328         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 329 else:
 330     u = {
 331         int(k): float(v)
 332         for k, v in [
 333             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 334         ]
 335     }
 336
 337     learning_rate_schedule = {}
 338     learning_rate = args.learning_rate
 339     for n_epoch in range(args.nb_epochs):
 340         if n_epoch in u:
 341             learning_rate = u[n_epoch]
 342         learning_rate_schedule[n_epoch] = learning_rate
 343
 344 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 345
 346 ##############################
 347
 348 nb_samples_seen = 0
 349
 350 if nb_epochs_finished >= nb_epochs:
 351     task.produce_results(nb_epochs_finished, model)
 352
 353 for n_epoch in range(nb_epochs_finished, nb_epochs):
 354
 355     learning_rate = learning_rate_schedule[n_epoch]
 356
 357     log_string(f"learning_rate {learning_rate}")
 358
 359     if args.optim == "sgd":
 360         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 361     elif args.optim == "adam":
 362         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 363     elif args.optim == "adamw":
 364         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 365     else:
 366         raise ValueError(f"Unknown optimizer {args.optim}.")
 367
 368     model.train()
 369
 370     nb_train_samples, acc_train_loss = 0, 0.0
 371
 372     for input in task.batches(split="train"):
 373         input = input.to(device)
 374         output = model(mygpt.BracketedSequence(input)).x
 375         loss = F.cross_entropy(output.transpose(1, 2), input)
 376         acc_train_loss += loss.item() * input.size(0)
 377         nb_train_samples += input.size(0)
 378         nb_samples_seen += input.size(0)
 379
 380         optimizer.zero_grad()
 381         loss.backward()
 382         optimizer.step()
 383
 384     with torch.autograd.no_grad():
 385
 386         model.eval()
 387
 388         nb_test_samples, acc_test_loss = 0, 0.0
 389
 390         for input in task.batches(split="test"):
 391             input = input.to(device)
 392
 393             # input, loss_masks, true_images = task.excise_last_image(input)
 394             # input, loss_masks = task.add_true_image(input, true_images, loss_masks)
 395
 396             output = model(mygpt.BracketedSequence(input)).x
 397             loss = F.cross_entropy(output.transpose(1, 2), input)
 398             acc_test_loss += loss.item() * input.size(0)
 399             nb_test_samples += input.size(0)
 400
 401         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 402         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 403
 404         log_string(
 405             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 406         )
 407
 408         task.produce_results(n_epoch, model)
 409
 410     checkpoint = {
 411         "nb_epochs_finished": n_epoch + 1,
 412         "model_state": model.state_dict(),
 413         "rng_state": torch.get_rng_state(),
 414     }
 415
 416     if torch.cuda.is_available():
 417         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 418
 419     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 420     torch.save(checkpoint, checkpoint_name)
 421     log_string(f"saved checkpoint {checkpoint_name}")
 422
 423 ######################################################################