main.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, 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.",
  31     formatter_class=argparse.ArgumentDefaultsHelpFormatter,
  32 )
  33
  34 parser.add_argument("--task", type=str, default="picoclvr")
  35
  36 parser.add_argument("--log_filename", type=str, default="train.log")
  37
  38 parser.add_argument("--result_dir", type=str, default="results_default")
  39
  40 parser.add_argument("--seed", type=int, default=0)
  41
  42 parser.add_argument("--nb_epochs", type=int, default=None)
  43
  44 parser.add_argument("--batch_size", type=int, default=None)
  45
  46 parser.add_argument("--nb_train_samples", type=int, default=250000)
  47
  48 parser.add_argument("--nb_test_samples", type=int, default=10000)
  49
  50 parser.add_argument("--optim", type=str, default="adam")
  51
  52 parser.add_argument("--learning_rate", type=float, default=1e-4)
  53
  54 parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
  55
  56 parser.add_argument("--dim_model", type=int, default=512)
  57
  58 parser.add_argument("--dim_keys", type=int, default=64)
  59
  60 parser.add_argument("--dim_hidden", type=int, default=2048)
  61
  62 parser.add_argument("--nb_heads", type=int, default=8)
  63
  64 parser.add_argument("--nb_blocks", type=int, default=12)
  65
  66 parser.add_argument("--dropout", type=float, default=0.1)
  67
  68 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  69
  70 parser.add_argument("--no_checkpoint", action="store_true", default=False)
  71
  72 parser.add_argument("--overwrite_results", action="store_true", default=False)
  73
  74 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
  75
  76 ##############################
  77 # picoclvr options
  78
  79 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
  80
  81 parser.add_argument("--picoclvr_height", type=int, default=12)
  82
  83 parser.add_argument("--picoclvr_width", type=int, default=16)
  84
  85 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
  86
  87 ##############################
  88 # Maze options
  89
  90 parser.add_argument("--maze_height", type=int, default=13)
  91
  92 parser.add_argument("--maze_width", type=int, default=21)
  93
  94 parser.add_argument("--maze_nb_walls", type=int, default=15)
  95
  96 ##############################
  97 # Snake options
  98
  99 parser.add_argument("--snake_height", type=int, default=6)
 100
 101 parser.add_argument("--snake_width", type=int, default=8)
 102
 103 parser.add_argument("--snake_nb_colors", type=int, default=5)
 104
 105 parser.add_argument("--snake_length", type=int, default=200)
 106
 107 ######################################################################
 108
 109 args = parser.parse_args()
 110
 111 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
 112
 113 try:
 114     os.mkdir(args.result_dir)
 115 except FileExistsError:
 116     if not args.overwrite_results:
 117         print(f"result directory {args.result_dir} already exists")
 118         exit(1)
 119
 120 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 121
 122 if args.seed >= 0:
 123     # torch.backends.cudnn.deterministic = True
 124     # torch.backends.cudnn.benchmark = False
 125     # torch.use_deterministic_algorithms(True)
 126     torch.manual_seed(args.seed)
 127     if torch.cuda.is_available():
 128         torch.cuda.manual_seed_all(args.seed)
 129
 130 ######################################################################
 131
 132 default_args = {
 133     "picoclvr": {
 134         "nb_epochs": 25,
 135         "batch_size": 25,
 136     },
 137     "mnist": {
 138         "nb_epochs": 25,
 139         "batch_size": 10,
 140     },
 141     "maze": {
 142         "nb_epochs": 25,
 143         "batch_size": 25,
 144     },
 145     "snake": {
 146         "nb_epochs": 5,
 147         "batch_size": 25,
 148     },
 149 }
 150
 151 if args.task in default_args:
 152     for k, v in default_args[args.task].items():
 153         if getattr(args, k) is None:
 154             setattr(args, k, v)
 155
 156 ######################################################################
 157
 158
 159 def log_string(s):
 160     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 161
 162     if log_file is not None:
 163         log_file.write(t + s + "\n")
 164         log_file.flush()
 165
 166     print(t + s)
 167     sys.stdout.flush()
 168
 169
 170 for n in vars(args):
 171     log_string(f"args.{n} {getattr(args, n)}")
 172
 173 ######################################################################
 174
 175
 176 # ra_mask is boolean, with 1s on the values to generate
 177
 178
 179 def masked_inplace_autoregression(
 180     model,
 181     batch_size,
 182     input,
 183     ar_mask,
 184     forbidden_tokens=None,
 185     progress_bar_desc="autoregression",
 186     device=torch.device("cpu"),
 187 ):
 188     batches = zip(input.split(batch_size), ar_mask.split(batch_size))
 189     if progress_bar_desc is not None:
 190         tqdm.tqdm(
 191             batches,
 192             dynamic_ncols=True,
 193             desc=progress_bar_desc,
 194             total=input.size(0) // batch_size,
 195         )
 196     for input, ar_mask in batches:
 197         i = (ar_mask.sum(0) > 0).nonzero()
 198         if i.min() > 0:
 199             model(
 200                 mygpt.BracketedSequence(input, 0, i.min())
 201             )  # Needed to initialize the model's cache
 202         for s in range(i.min(), i.max() + 1):
 203             output = model(mygpt.BracketedSequence(input, s, 1)).x
 204             logits = output[:, s]
 205             if forbidden_tokens is not None:
 206                 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
 207             if args.deterministic_synthesis:
 208                 t_next = logits.argmax(1)
 209             else:
 210                 dist = torch.distributions.categorical.Categorical(logits=logits)
 211                 t_next = dist.sample()
 212             input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 213
 214
 215 ######################################################################
 216
 217
 218 class Task:
 219     def batches(self, split="train"):
 220         pass
 221
 222     def vocabulary_size(self):
 223         pass
 224
 225     def produce_results(self, n_epoch, model):
 226         pass
 227
 228
 229 ######################################################################
 230
 231 import picoclvr
 232
 233
 234 class TaskPicoCLVR(Task):
 235     # Make a tensor from a list of strings
 236     def tensorize(self, descr):
 237         token_descr = [s.strip().split(" ") for s in descr]
 238         l = max([len(s) for s in token_descr])
 239         token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
 240         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
 241         return torch.tensor(id_descr, device=self.device)
 242
 243     # Make a list of strings from a tensor
 244     def detensorize(self, x):
 245         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
 246
 247     # trim all the tensors in the tuple z to remove as much token from
 248     # left and right in the first tensor. If z is a tuple, all its
 249     # elements are trimed according to the triming for the first
 250     def trim(self, z, token="<nul>"):
 251         n = self.token2id[token]
 252         if type(z) == tuple:
 253             x = z[0]
 254             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 255             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 256             return tuple([t[:, a:b] for t in z])
 257         else:
 258             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 259             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 260             return z[:, a:b]
 261
 262     ######################
 263     # Not the cleanest part of the code
 264
 265     # Extract the last image of each sequence, from the last <img>
 266     # included, and set to <nul> all the tokens from the beginning of
 267     # that image to the end
 268     def excise_last_image(self, input):
 269         t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 270         nb_img_tokens = self.height * self.width + 1
 271
 272         input = input.clone()
 273         t = (input == t_img).long()
 274         tail_masks = (t.cumsum(dim=1) == t.sum(dim=1, keepdim=True)).long()
 275         i = (t * tail_masks).nonzero(as_tuple=True)
 276         j = (
 277             i[0][:, None],
 278             i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
 279         )
 280         images = self.trim(input[j])
 281         input[j] = t_nul
 282         loss_masks = 1 - tail_masks
 283         input, loss_masks = self.trim((input, loss_masks))
 284         return input, loss_masks, images
 285
 286     def add_true_image(self, input, images, loss_masks):
 287         t_nul = self.token2id["<nul>"]
 288         nb_img_tokens = self.height * self.width + 1
 289         input = F.pad(input, (0, nb_img_tokens), value=t_nul)
 290         loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
 291         t = (input == t_nul).long()
 292         i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
 293         j = (
 294             i[0][:, None],
 295             i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
 296         )
 297         input[j] = images
 298         loss_masks[j] = 1
 299         input, loss_masks = self.trim((input, loss_masks))
 300         return input, loss_masks
 301
 302     def add_generated_image(self, input, loss_masks, model):
 303         t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 304         nb_img_tokens = self.height * self.width + 1
 305
 306         input = F.pad(input, (0, nb_img_tokens), value=t_nul)
 307         loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
 308         t = (input == t_nul).long()
 309         i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
 310         input[i] = t_img
 311
 312         j = (
 313             i[0][:, None],
 314             i[1][:, None]
 315             + 1
 316             + torch.arange(nb_img_tokens - 1, device=input.device)[None, :],
 317         )
 318         ar_masks = input.new_zeros(input.size(), dtype=torch.int64)
 319         ar_masks[j] = 1
 320         forbidden_tokens = (
 321             torch.arange(self.vocabulary_size(), device=input.device) == t_nul
 322         )
 323         with torch.autograd.no_grad():
 324             t = model.training
 325             model.eval()
 326             masked_inplace_autoregression(
 327                 model,
 328                 self.batch_size,
 329                 input,
 330                 ar_masks,
 331                 forbidden_tokens,
 332                 progress_bar_desc=None,
 333                 device=self.device,
 334             )
 335             model.train(t)
 336
 337         input, loss_masks = self.trim((input, loss_masks))
 338
 339         return input, loss_masks
 340
 341     ######################
 342
 343     def __init__(
 344         self,
 345         nb_train_samples,
 346         nb_test_samples,
 347         batch_size,
 348         height,
 349         width,
 350         nb_colors=5,
 351         device=torch.device("cpu"),
 352         pruner_train=None,
 353         pruner_eval=None,
 354     ):
 355         def generate_descr(nb, cache_suffix, pruner):
 356             return picoclvr.generate(
 357                 nb,
 358                 height=self.height,
 359                 width=self.width,
 360                 nb_colors=nb_colors,
 361                 pruner=pruner,
 362             )
 363
 364         self.height = height
 365         self.width = width
 366         self.batch_size = batch_size
 367         self.device = device
 368         self.pruner_train = pruner_train
 369         self.pruner_eval = pruner_eval
 370
 371         param = {
 372             "nb_train_samples": nb_train_samples,
 373             "nb_test_samples": nb_test_samples,
 374             "height": height,
 375             "width": width,
 376             "nb_colors": nb_colors,
 377             "batch_size": batch_size,
 378             "rng_state": list(torch.get_rng_state()),
 379         }
 380
 381         log_string(
 382             f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
 383         )
 384         self.train_descr = generate_descr(
 385             nb_train_samples, "train", pruner=self.pruner_train
 386         )
 387         self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
 388
 389         # Build the tokenizer
 390         tokens = {"<nul>", "<img>"}
 391         for d in [self.train_descr, self.test_descr]:
 392             for s in d:
 393                 for t in s.strip().split(" "):
 394                     tokens.add(t)
 395         # make this set a sorted list to get the same tensors given
 396         # the same descr
 397         tokens = list(tokens)
 398         tokens.sort()
 399         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
 400         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
 401
 402         # Tokenize the train and test sets
 403         self.train_input = self.tensorize(self.train_descr)
 404         self.test_input = self.tensorize(self.test_descr)
 405
 406     def batches(self, split="train"):
 407         assert split in {"train", "test"}
 408         input = self.train_input if split == "train" else self.test_input
 409         for batch in tqdm.tqdm(
 410             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 411         ):
 412             yield self.trim(batch)
 413
 414     def vocabulary_size(self):
 415         return len(self.token2id)
 416
 417     def compute_missing_properties(self, n_epoch, model, pruner=None):
 418         acc_nb_requested_properties = []
 419         acc_nb_missing_properties = []
 420         acc_nb_results = 0
 421
 422         for input in tqdm.tqdm(
 423             self.test_input.split(self.batch_size),
 424             dynamic_ncols=True,
 425             desc=f"test-properties",
 426         ):
 427             tape, loss_masks, _ = self.excise_last_image(input)
 428             tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
 429             result_descr = self.detensorize(tape)
 430             np = picoclvr.nb_properties(
 431                 result_descr,
 432                 height=self.height,
 433                 width=self.width,
 434                 pruner=pruner,
 435             )
 436             nb_requested_properties, _, nb_missing_properties = zip(*np)
 437             acc_nb_requested_properties += nb_requested_properties
 438             acc_nb_missing_properties += nb_missing_properties
 439             acc_nb_results += len(result_descr)
 440
 441         nb_requested_properties = sum(acc_nb_requested_properties)
 442         nb_missing_properties = sum(acc_nb_missing_properties)
 443
 444         prefix = "" if pruner is None else "pruned_"
 445         log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 446         log_string(
 447             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 448         )
 449         log_string(
 450             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 451         )
 452
 453     ######################################################################
 454
 455     def produce_results(self, n_epoch, model):
 456         self.compute_missing_properties(n_epoch, model)
 457
 458         if self.pruner_eval is not None:
 459             self.compute_missing_properties(n_epoch, model, self.pruner_eval)
 460
 461         nb_tokens_to_generate = self.height * self.width + 3
 462         result_descr = []
 463         nb_per_primer = 8
 464         primer = []
 465
 466         for primer_descr in [
 467             "red above green <sep> green top <sep> blue right of red",
 468             "there is red <sep> there is yellow <sep> there is blue",
 469             "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
 470             "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
 471         ]:
 472             primer += [primer_descr] * nb_per_primer
 473
 474         tape = self.tensorize(primer)
 475         loss_masks = 1 - (tape == self.token2id["<nul>"]).long()
 476         tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
 477         result_descr = self.detensorize(tape)
 478
 479         np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
 480
 481         acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
 482         acc_nb_results = len(result_descr)
 483
 484         nb_requested_properties = sum(acc_nb_requested_properties)
 485         nb_missing_properties = sum(acc_nb_missing_properties)
 486
 487         prefix = "demo_"
 488         log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 489         log_string(
 490             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 491         )
 492         log_string(
 493             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 494         )
 495
 496         img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
 497
 498         if img.dim() == 5:
 499             if img.size(1) == 1:
 500                 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
 501             else:
 502                 img = torch.cat(
 503                     [
 504                         torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
 505                         for x in img
 506                     ],
 507                     0,
 508                 )
 509
 510         image_name = os.path.join(args.result_dir, f"picoclvr_result_{n_epoch:04d}.png")
 511         torchvision.utils.save_image(
 512             img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=1.0
 513         )
 514         log_string(f"wrote {image_name}")
 515
 516
 517 ######################################################################
 518
 519
 520 class TaskMNIST(Task):
 521     def __init__(self, batch_size, device=torch.device("cpu")):
 522         self.device = device
 523         self.batch_size = batch_size
 524
 525     def batches(self, split="train"):
 526         assert split in {"train", "test"}
 527         data_set = torchvision.datasets.MNIST(
 528             root="./data", train=(split == "train"), download=True
 529         )
 530         data_input = data_set.data.view(-1, 28 * 28).long()
 531         if args.nb_train_samples is not None:
 532             data_input = data_input[: args.nb_train_samples]
 533         for batch in tqdm.tqdm(
 534             data_input.split(self.batch_size), desc=f"epoch-{split}"
 535         ):
 536             yield batch
 537
 538     def vocabulary_size(self):
 539         return 256
 540
 541     def produce_results(self, n_epoch, model):
 542         results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
 543         ar_mask = torch.full_like(results, 1)
 544         masked_inplace_autoregression(
 545             model, self.batch_size, results, ar_mask, device=self.device
 546         )
 547         image_name = os.path.join(args.result_dir, f"mnist_result_{n_epoch:04d}.png")
 548         torchvision.utils.save_image(
 549             1 - results.reshape(-1, 1, 28, 28) / 255.0,
 550             image_name,
 551             nrow=16,
 552             pad_value=0.8,
 553         )
 554         log_string(f"wrote {image_name}")
 555
 556
 557 ######################################################################
 558
 559 import maze
 560
 561
 562 class TaskMaze(Task):
 563     def map2seq(self, *m):
 564         return torch.cat([x.flatten(1) for x in m], 1)
 565
 566     def seq2map(self, s):
 567         s = s.reshape(s.size(0), -1, self.height, self.width)
 568         return (s[:, k] for k in range(s.size(1)))
 569
 570     def __init__(
 571         self,
 572         nb_train_samples,
 573         nb_test_samples,
 574         batch_size,
 575         height,
 576         width,
 577         nb_walls,
 578         device=torch.device("cpu"),
 579     ):
 580         self.batch_size = batch_size
 581         self.height = height
 582         self.width = width
 583         self.device = device
 584
 585         train_mazes, train_paths, _ = maze.create_maze_data(
 586             nb_train_samples,
 587             height=height,
 588             width=width,
 589             nb_walls=nb_walls,
 590             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 591         )
 592         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 593
 594         test_mazes, test_paths, _ = maze.create_maze_data(
 595             nb_test_samples,
 596             height=height,
 597             width=width,
 598             nb_walls=nb_walls,
 599             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 600         )
 601         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 602
 603         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 604
 605     def batches(self, split="train", nb_to_use=-1, desc=None):
 606         assert split in {"train", "test"}
 607         input = self.train_input if split == "train" else self.test_input
 608         if nb_to_use > 0:
 609             input = input[:nb_to_use]
 610         if desc is None:
 611             desc = f"epoch-{split}"
 612         for batch in tqdm.tqdm(
 613             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 614         ):
 615             yield batch
 616
 617     def vocabulary_size(self):
 618         return self.nb_codes
 619
 620     def compute_error(self, model, split="train", nb_to_use=-1):
 621         nb_total, nb_correct = 0, 0
 622         for input in task.batches(split, nb_to_use):
 623             result = input.clone()
 624             ar_mask = result.new_zeros(result.size())
 625             ar_mask[:, self.height * self.width :] = 1
 626             result *= 1 - ar_mask
 627             masked_inplace_autoregression(
 628                 model, self.batch_size, result, ar_mask, device=self.device
 629             )
 630             mazes, paths = self.seq2map(result)
 631             nb_correct += maze.path_correctness(mazes, paths).long().sum()
 632             nb_total += mazes.size(0)
 633
 634         return nb_total, nb_correct
 635
 636     def produce_results(self, n_epoch, model):
 637         with torch.autograd.no_grad():
 638             t = model.training
 639             model.eval()
 640
 641             train_nb_total, train_nb_correct = self.compute_error(
 642                 model, "train", nb_to_use=1000
 643             )
 644             log_string(
 645                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 646             )
 647
 648             test_nb_total, test_nb_correct = self.compute_error(
 649                 model, "test", nb_to_use=1000
 650             )
 651             log_string(
 652                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 653             )
 654
 655             input = self.test_input[:48]
 656             result = input.clone()
 657             ar_mask = result.new_zeros(result.size())
 658             ar_mask[:, self.height * self.width :] = 1
 659             result *= 1 - ar_mask
 660             masked_inplace_autoregression(
 661                 model, self.batch_size, result, ar_mask, device=self.device
 662             )
 663
 664             mazes, paths = self.seq2map(input)
 665             _, predicted_paths = self.seq2map(result)
 666
 667             filename = os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.png")
 668             maze.save_image(
 669                 filename,
 670                 mazes=mazes,
 671                 target_paths=paths,
 672                 predicted_paths=predicted_paths,
 673                 path_correct=maze.path_correctness(mazes, predicted_paths),
 674             )
 675             log_string(f"wrote {filename}")
 676
 677             model.train(t)
 678
 679
 680 ######################################################################
 681
 682
 683 import snake
 684
 685
 686 class TaskSnake(Task):
 687     def __init__(
 688         self,
 689         nb_train_samples,
 690         nb_test_samples,
 691         batch_size,
 692         height,
 693         width,
 694         nb_colors,
 695         length,
 696         prompt_length,
 697         device=torch.device("cpu"),
 698     ):
 699         self.batch_size = batch_size
 700         self.height = height
 701         self.width = width
 702         self.device = device
 703         self.prompt_length = prompt_length
 704
 705         self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
 706             nb_train_samples,
 707             height,
 708             width,
 709             nb_colors,
 710             length,
 711             prompt_length,
 712             self.device,
 713         )
 714         self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
 715             nb_test_samples,
 716             height,
 717             width,
 718             nb_colors,
 719             length,
 720             prompt_length,
 721             self.device,
 722         )
 723
 724         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 725
 726     def batches(self, split="train", nb_to_use=-1, desc=None):
 727         assert split in {"train", "test"}
 728         input = self.train_input if split == "train" else self.test_input
 729         if nb_to_use > 0:
 730             input = input[:nb_to_use]
 731         if desc is None:
 732             desc = f"epoch-{split}"
 733         for batch in tqdm.tqdm(
 734             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 735         ):
 736             yield batch
 737
 738     def vocabulary_size(self):
 739         return self.nb_codes
 740
 741     def produce_results(self, n_epoch, model):
 742         with torch.autograd.no_grad():
 743             t = model.training
 744             model.eval()
 745
 746             def compute_nb_correct(input, prior_visits):
 747                 result = input.clone()
 748                 i = torch.arange(result.size(1), device=result.device)[None, :]
 749                 ar_mask = (
 750                     torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
 751                     .long()
 752                     .expand_as(result)
 753                 )
 754                 result *= 1 - ar_mask
 755
 756                 # snake.solver(result,ar_mask)
 757
 758                 masked_inplace_autoregression(
 759                     model, self.batch_size, result, ar_mask, device=self.device
 760                 )
 761
 762                 nb_total = ((prior_visits > 0) * ar_mask).sum()
 763
 764                 nb_correct = (
 765                     (result == input).long() * (prior_visits > 0) * ar_mask
 766                 ).sum()
 767
 768                 # nb_total = result.size(0)
 769                 # nb_correct = ((result - input).abs().sum(1) == 0).sum()
 770
 771                 return nb_total, nb_correct
 772
 773             # train_nb_total, train_nb_correct = compute_nb_correct(
 774             # self.train_input, self.train_prior_visits
 775             # )
 776
 777             # log_string(
 778             # f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 779             # )
 780
 781             test_nb_total, test_nb_correct = compute_nb_correct(
 782                 self.test_input[:1000], self.test_prior_visits[:1000]
 783             )
 784
 785             log_string(
 786                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 787             )
 788
 789             model.train(t)
 790
 791
 792 ######################################################################
 793
 794
 795 def picoclvr_pruner_horizontal_green(p):
 796     return not ("green" in p and ("left" in p or "right" in p))
 797
 798
 799 picoclvr_pruner_train = (
 800     picoclvr_pruner_horizontal_green
 801     if args.picocvlr_prune_properties in {"train+eval"}
 802     else None
 803 )
 804
 805 picoclvr_pruner_eval = (
 806     (lambda p: not picoclvr_pruner_horizontal_green(p))
 807     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 808     else None
 809 )
 810
 811 ######################################################################
 812
 813 if args.task == "picoclvr":
 814     task = TaskPicoCLVR(
 815         nb_train_samples=args.nb_train_samples,
 816         nb_test_samples=args.nb_test_samples,
 817         batch_size=args.batch_size,
 818         height=args.picoclvr_height,
 819         width=args.picoclvr_width,
 820         nb_colors=args.picoclvr_nb_colors,
 821         device=device,
 822         pruner_train=picoclvr_pruner_train,
 823         pruner_eval=picoclvr_pruner_eval,
 824     )
 825
 826 elif args.task == "mnist":
 827     task = TaskMNIST(
 828         batch_size=args.batch_size,
 829         device=device,
 830     )
 831
 832 elif args.task == "maze":
 833     task = TaskMaze(
 834         nb_train_samples=args.nb_train_samples,
 835         nb_test_samples=args.nb_test_samples,
 836         batch_size=args.batch_size,
 837         height=args.maze_height,
 838         width=args.maze_width,
 839         nb_walls=args.maze_nb_walls,
 840         device=device,
 841     )
 842
 843 elif args.task == "snake":
 844     task = TaskSnake(
 845         nb_train_samples=args.nb_train_samples,
 846         nb_test_samples=args.nb_test_samples,
 847         batch_size=args.batch_size,
 848         height=args.snake_height,
 849         width=args.snake_width,
 850         nb_colors=args.snake_nb_colors,
 851         length=args.snake_length,
 852         prompt_length=args.snake_length // 2,
 853         device=device,
 854     )
 855
 856 else:
 857     raise ValueError(f"Unknown task {args.task}")
 858
 859 ######################################################################
 860
 861 log_string(f"device {device}")
 862
 863 vocabulary_size = task.vocabulary_size()
 864
 865 log_string(f"vocabulary_size {vocabulary_size}")
 866
 867 ##############################
 868
 869 model = mygpt.MyGPT(
 870     vocabulary_size=vocabulary_size,
 871     dim_model=args.dim_model,
 872     dim_keys=args.dim_keys,
 873     dim_hidden=args.dim_hidden,
 874     nb_heads=args.nb_heads,
 875     nb_blocks=args.nb_blocks,
 876     causal=True,
 877     dropout=args.dropout,
 878 )
 879
 880 model.to(device)
 881
 882 nb_parameters = sum(p.numel() for p in model.parameters())
 883 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 884
 885 ######################################################################
 886
 887 nb_epochs_finished = 0
 888
 889 if args.no_checkpoint:
 890     log_string(f"not trying to load checkpoint.")
 891
 892 else:
 893     try:
 894         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 895         checkpoint = torch.load(checkpoint_name)
 896         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 897         model.load_state_dict(checkpoint["model_state"])
 898         torch.set_rng_state(checkpoint["rng_state"])
 899         if torch.cuda.is_available():
 900             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 901
 902         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 903
 904     except FileNotFoundError:
 905         log_string("starting from scratch.")
 906
 907     except:
 908         log_string("error when loading the checkpoint.")
 909         exit(1)
 910
 911 ######################################################################
 912
 913 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 914
 915 token_count = 0
 916 for input in task.batches(split="train"):
 917     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 918 token_probas = token_count / token_count.sum()
 919 entropy = -torch.xlogy(token_probas, token_probas).sum()
 920 train_set_perplexity = math.exp(entropy)
 921
 922 ##############################
 923
 924 if args.learning_rate_schedule == "cos":
 925     learning_rate_schedule = {}
 926     for n_epoch in range(args.nb_epochs):
 927         u = n_epoch / args.nb_epochs * math.pi
 928         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 929 else:
 930     u = {
 931         int(k): float(v)
 932         for k, v in [
 933             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 934         ]
 935     }
 936
 937     learning_rate_schedule = {}
 938     learning_rate = args.learning_rate
 939     for n_epoch in range(args.nb_epochs):
 940         if n_epoch in u:
 941             learning_rate = u[n_epoch]
 942         learning_rate_schedule[n_epoch] = learning_rate
 943
 944 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 945
 946 ##############################
 947
 948 nb_samples_seen = 0
 949
 950 if nb_epochs_finished >= nb_epochs:
 951     task.produce_results(nb_epochs_finished, model)
 952
 953 for n_epoch in range(nb_epochs_finished, nb_epochs):
 954     learning_rate = learning_rate_schedule[n_epoch]
 955
 956     log_string(f"learning_rate {learning_rate}")
 957
 958     if args.optim == "sgd":
 959         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 960     elif args.optim == "adam":
 961         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 962     elif args.optim == "adamw":
 963         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 964     else:
 965         raise ValueError(f"Unknown optimizer {args.optim}.")
 966
 967     model.train()
 968
 969     nb_train_samples, acc_train_loss = 0, 0.0
 970
 971     for input in task.batches(split="train"):
 972         input = input.to(device)
 973         output = model(mygpt.BracketedSequence(input)).x
 974         loss = F.cross_entropy(output.transpose(1, 2), input)
 975         acc_train_loss += loss.item() * input.size(0)
 976         nb_train_samples += input.size(0)
 977         nb_samples_seen += input.size(0)
 978
 979         optimizer.zero_grad()
 980         loss.backward()
 981         optimizer.step()
 982
 983     with torch.autograd.no_grad():
 984         model.eval()
 985
 986         nb_test_samples, acc_test_loss = 0, 0.0
 987
 988         for input in task.batches(split="test"):
 989             input = input.to(device)
 990
 991             output = model(mygpt.BracketedSequence(input)).x
 992             loss = F.cross_entropy(output.transpose(1, 2), input)
 993             acc_test_loss += loss.item() * input.size(0)
 994             nb_test_samples += input.size(0)
 995
 996         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 997         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
 998
 999         log_string(
1000             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
1001         )
1002
1003         task.produce_results(n_epoch, model)
1004
1005     checkpoint = {
1006         "nb_epochs_finished": n_epoch + 1,
1007         "model_state": model.state_dict(),
1008         "rng_state": torch.get_rng_state(),
1009     }
1010
1011     if torch.cuda.is_available():
1012         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
1013
1014     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
1015     torch.save(checkpoint, checkpoint_name)
1016     log_string(f"saved checkpoint {checkpoint_name}")
1017
1018 ######################################################################