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