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=25)
  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=3)
 104
 105 parser.add_argument("--snake_length", type=int, default=400)
 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         "batch_size": 25,
 135     },
 136     "mnist": {
 137         "batch_size": 10,
 138     },
 139     "maze": {
 140         "batch_size": 25,
 141     },
 142     "snake": {
 143         "batch_size": 20,
 144     },
 145 }
 146
 147 if args.task in default_args:
 148     for k, v in default_args[args.task].items():
 149         if getattr(args, k) is None:
 150             setattr(args, k, v)
 151
 152 ######################################################################
 153
 154
 155 def log_string(s):
 156     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 157
 158     if log_file is not None:
 159         log_file.write(t + s + "\n")
 160         log_file.flush()
 161
 162     print(t + s)
 163     sys.stdout.flush()
 164
 165
 166 for n in vars(args):
 167     log_string(f"args.{n} {getattr(args, n)}")
 168
 169 ######################################################################
 170
 171
 172 def masked_inplace_autoregression(
 173     model, batch_size, input, ar_mask, forbidden_tokens=None, device=torch.device("cpu")
 174 ):
 175     for input, ar_mask in tqdm.tqdm(
 176         zip(input.split(batch_size), ar_mask.split(batch_size)),
 177         dynamic_ncols=True,
 178         desc="autoregression",
 179         total=input.size(0) // batch_size,
 180     ):
 181         i = (ar_mask.sum(0) > 0).nonzero()
 182         if i.min() > 0:
 183             model(
 184                 mygpt.BracketedSequence(input, 0, i.min())
 185             )  # Needed to initialize the model's cache
 186         for s in range(i.min(), i.max() + 1):
 187             output = model(mygpt.BracketedSequence(input, s, 1)).x
 188             logits = output[:, s]
 189             if forbidden_tokens is not None:
 190                 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
 191             if args.deterministic_synthesis:
 192                 t_next = logits.argmax(1)
 193             else:
 194                 dist = torch.distributions.categorical.Categorical(logits=logits)
 195                 t_next = dist.sample()
 196             input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 197
 198
 199 ######################################################################
 200
 201
 202 class Task:
 203     def batches(self, split="train"):
 204         pass
 205
 206     def vocabulary_size(self):
 207         pass
 208
 209     def produce_results(self, n_epoch, model):
 210         pass
 211
 212
 213 ######################################################################
 214
 215 import picoclvr
 216
 217
 218 class TaskPicoCLVR(Task):
 219     # Make a tensor from a list of strings
 220     def tensorize(self, descr):
 221         token_descr = [s.strip().split(" ") for s in descr]
 222         l = max([len(s) for s in token_descr])
 223         token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
 224         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
 225         return torch.tensor(id_descr, device=self.device)
 226
 227     # Make a list of strings from a tensor
 228     def detensorize(self, x):
 229         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
 230
 231     # trim all the tensors in the tuple z to remove as much token from
 232     # left and right in the first tensor. If z is a tuple, all its
 233     # elements are trimed according to the triming for the first
 234     def trim(self, z, token="<nul>"):
 235         n = self.token2id[token]
 236         if type(z) == tuple:
 237             x = z[0]
 238             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 239             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 240             return tuple([t[:, a:b] for t in z])
 241         else:
 242             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 243             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 244             return z[:, a:b]
 245
 246     ######################
 247     # Not the cleanest part of the code
 248
 249     # Extract the last image of each sequence, from the last <img>
 250     # included, and set to <nul> all the tokens from the beginning of
 251     # that image to the end
 252     def excise_last_image(self, input):
 253         t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 254         nb_img_tokens = self.height * self.width + 1
 255
 256         input = input.clone()
 257         t = (input == t_img).long()
 258         tail_masks = (t.cumsum(dim=1) == t.sum(dim=1, keepdim=True)).long()
 259         i = (t * tail_masks).nonzero(as_tuple=True)
 260         j = (
 261             i[0][:, None],
 262             i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
 263         )
 264         images = self.trim(input[j])
 265         input[j] = t_nul
 266         loss_masks = 1 - tail_masks
 267         input, loss_masks = self.trim((input, loss_masks))
 268         return input, loss_masks, images
 269
 270     def add_true_image(self, input, images, loss_masks):
 271         t_nul = self.token2id["<nul>"]
 272         nb_img_tokens = self.height * self.width + 1
 273         input = F.pad(input, (0, nb_img_tokens), value=t_nul)
 274         loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
 275         t = (input == t_nul).long()
 276         i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
 277         j = (
 278             i[0][:, None],
 279             i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :],
 280         )
 281         input[j] = images
 282         loss_masks[j] = 1
 283         input, loss_masks = self.trim((input, loss_masks))
 284         return input, loss_masks
 285
 286     def add_generated_image(self, input, loss_masks, model):
 287         t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 288         nb_img_tokens = self.height * self.width + 1
 289
 290         input = F.pad(input, (0, nb_img_tokens), value=t_nul)
 291         loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0)
 292         t = (input == t_nul).long()
 293         i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True)
 294         input[i] = t_img
 295
 296         j = (
 297             i[0][:, None],
 298             i[1][:, None]
 299             + 1
 300             + torch.arange(nb_img_tokens - 1, device=input.device)[None, :],
 301         )
 302         ar_masks = input.new_zeros(input.size(), dtype=torch.int64)
 303         ar_masks[j] = 1
 304         forbidden_tokens = (
 305             torch.arange(self.vocabulary_size(), device=input.device) == t_nul
 306         )
 307         with torch.autograd.no_grad():
 308             t = model.training
 309             model.eval()
 310             masked_inplace_autoregression(
 311                 model,
 312                 self.batch_size,
 313                 input,
 314                 ar_masks,
 315                 forbidden_tokens,
 316                 device=self.device,
 317             )
 318             model.train(t)
 319
 320         input, loss_masks = self.trim((input, loss_masks))
 321
 322         return input, loss_masks
 323
 324     ######################
 325
 326     def __init__(
 327         self,
 328         nb_train_samples,
 329         nb_test_samples,
 330         batch_size,
 331         height,
 332         width,
 333         nb_colors=5,
 334         device=torch.device("cpu"),
 335         pruner_train=None,
 336         pruner_eval=None,
 337     ):
 338         def generate_descr(nb, cache_suffix, pruner):
 339             return picoclvr.generate(
 340                 nb,
 341                 height=self.height,
 342                 width=self.width,
 343                 nb_colors=nb_colors,
 344                 pruner=pruner,
 345             )
 346
 347         self.height = height
 348         self.width = width
 349         self.batch_size = batch_size
 350         self.device = device
 351         self.pruner_train = pruner_train
 352         self.pruner_eval = pruner_eval
 353
 354         param = {
 355             "nb_train_samples": nb_train_samples,
 356             "nb_test_samples": nb_test_samples,
 357             "height": height,
 358             "width": width,
 359             "nb_colors": nb_colors,
 360             "batch_size": batch_size,
 361             "rng_state": list(torch.get_rng_state()),
 362         }
 363
 364         log_string(
 365             f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
 366         )
 367         self.train_descr = generate_descr(
 368             nb_train_samples, "train", pruner=self.pruner_train
 369         )
 370         self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
 371
 372         # Build the tokenizer
 373         tokens = {"<nul>", "<img>"}
 374         for d in [self.train_descr, self.test_descr]:
 375             for s in d:
 376                 for t in s.strip().split(" "):
 377                     tokens.add(t)
 378         # make this set a sorted list to get the same tensors given
 379         # the same descr
 380         tokens = list(tokens)
 381         tokens.sort()
 382         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
 383         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
 384
 385         # Tokenize the train and test sets
 386         self.train_input = self.tensorize(self.train_descr)
 387         self.test_input = self.tensorize(self.test_descr)
 388
 389     def batches(self, split="train"):
 390         assert split in {"train", "test"}
 391         input = self.train_input if split == "train" else self.test_input
 392         for batch in tqdm.tqdm(
 393             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 394         ):
 395             yield self.trim(batch)
 396
 397     def vocabulary_size(self):
 398         return len(self.token2id)
 399
 400     def compute_missing_properties(self, n_epoch, model, pruner=None):
 401         acc_nb_requested_properties = []
 402         acc_nb_missing_properties = []
 403         acc_nb_results = 0
 404
 405         for input in tqdm.tqdm(
 406             self.test_input.split(self.batch_size),
 407             dynamic_ncols=True,
 408             desc=f"test-properties",
 409         ):
 410             tape, loss_masks, _ = self.excise_last_image(input)
 411             tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
 412             result_descr = self.detensorize(tape)
 413             np = picoclvr.nb_properties(
 414                 result_descr,
 415                 height=self.height,
 416                 width=self.width,
 417                 pruner=pruner,
 418             )
 419             nb_requested_properties, _, nb_missing_properties = zip(*np)
 420             acc_nb_requested_properties += nb_requested_properties
 421             acc_nb_missing_properties += nb_missing_properties
 422             acc_nb_results += len(result_descr)
 423
 424         nb_requested_properties = sum(acc_nb_requested_properties)
 425         nb_missing_properties = sum(acc_nb_missing_properties)
 426
 427         prefix = "" if pruner is None else "pruned_"
 428         log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 429         log_string(
 430             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 431         )
 432         log_string(
 433             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 434         )
 435
 436     ######################################################################
 437
 438     def produce_results(self, n_epoch, model):
 439         self.compute_missing_properties(n_epoch, model)
 440
 441         if self.pruner_eval is not None:
 442             self.compute_missing_properties(n_epoch, model, self.pruner_eval)
 443
 444         nb_tokens_to_generate = self.height * self.width + 3
 445         result_descr = []
 446         nb_per_primer = 8
 447         primer = []
 448
 449         for primer_descr in [
 450             "red above green <sep> green top <sep> blue right of red",
 451             "there is red <sep> there is yellow <sep> there is blue",
 452             "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
 453             "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
 454         ]:
 455             primer += [primer_descr] * nb_per_primer
 456
 457         tape = self.tensorize(primer)
 458         loss_masks = 1 - (tape == self.token2id["<nul>"]).long()
 459         tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
 460         result_descr = self.detensorize(tape)
 461
 462         np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
 463
 464         acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
 465         acc_nb_results = len(result_descr)
 466
 467         nb_requested_properties = sum(acc_nb_requested_properties)
 468         nb_missing_properties = sum(acc_nb_missing_properties)
 469
 470         prefix = "demo_"
 471         log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 472         log_string(
 473             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 474         )
 475         log_string(
 476             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 477         )
 478
 479         img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
 480
 481         if img.dim() == 5:
 482             if img.size(1) == 1:
 483                 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
 484             else:
 485                 img = torch.cat(
 486                     [
 487                         torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
 488                         for x in img
 489                     ],
 490                     0,
 491                 )
 492
 493         image_name = os.path.join(args.result_dir, f"picoclvr_result_{n_epoch:04d}.png")
 494         torchvision.utils.save_image(
 495             img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=1.0
 496         )
 497         log_string(f"wrote {image_name}")
 498
 499
 500 ######################################################################
 501
 502
 503 class TaskMNIST(Task):
 504     def __init__(self, batch_size, device=torch.device("cpu")):
 505         self.device = device
 506         self.batch_size = batch_size
 507
 508     def batches(self, split="train"):
 509         assert split in {"train", "test"}
 510         data_set = torchvision.datasets.MNIST(
 511             root="./data", train=(split == "train"), download=True
 512         )
 513         data_input = data_set.data.view(-1, 28 * 28).long()
 514         if args.nb_train_samples is not None:
 515             data_input = data_input[: args.nb_train_samples]
 516         for batch in tqdm.tqdm(
 517             data_input.split(self.batch_size), desc=f"epoch-{split}"
 518         ):
 519             yield batch
 520
 521     def vocabulary_size(self):
 522         return 256
 523
 524     def produce_results(self, n_epoch, model):
 525         results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
 526         ar_mask = torch.full_like(results, 1)
 527         masked_inplace_autoregression(
 528             model, self.batch_size, results, ar_mask, device=self.device
 529         )
 530         image_name = os.path.join(args.result_dir, f"mnist_result_{n_epoch:04d}.png")
 531         torchvision.utils.save_image(
 532             1 - results.reshape(-1, 1, 28, 28) / 255.0,
 533             image_name,
 534             nrow=16,
 535             pad_value=0.8,
 536         )
 537         log_string(f"wrote {image_name}")
 538
 539
 540 ######################################################################
 541
 542 import maze
 543
 544
 545 class TaskMaze(Task):
 546     def map2seq(self, *m):
 547         return torch.cat([x.flatten(1) for x in m], 1)
 548
 549     def seq2map(self, s):
 550         s = s.reshape(s.size(0), -1, self.height, self.width)
 551         return (s[:, k] for k in range(s.size(1)))
 552
 553     def __init__(
 554         self,
 555         nb_train_samples,
 556         nb_test_samples,
 557         batch_size,
 558         height,
 559         width,
 560         nb_walls,
 561         device=torch.device("cpu"),
 562     ):
 563         self.batch_size = batch_size
 564         self.height = height
 565         self.width = width
 566         self.device = device
 567
 568         train_mazes, train_paths, _ = maze.create_maze_data(
 569             nb_train_samples,
 570             height=height,
 571             width=width,
 572             nb_walls=nb_walls,
 573             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 574         )
 575         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 576
 577         test_mazes, test_paths, _ = maze.create_maze_data(
 578             nb_test_samples,
 579             height=height,
 580             width=width,
 581             nb_walls=nb_walls,
 582             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 583         )
 584         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 585
 586         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 587
 588     def batches(self, split="train", nb_to_use=-1, desc=None):
 589         assert split in {"train", "test"}
 590         input = self.train_input if split == "train" else self.test_input
 591         if nb_to_use > 0:
 592             input = input[:nb_to_use]
 593         if desc is None:
 594             desc = f"epoch-{split}"
 595         for batch in tqdm.tqdm(
 596             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 597         ):
 598             yield batch
 599
 600     def vocabulary_size(self):
 601         return self.nb_codes
 602
 603     def compute_error(self, model, split="train", nb_to_use=-1):
 604         nb_total, nb_correct = 0, 0
 605         for input in task.batches(split, nb_to_use):
 606             result = input.clone()
 607             ar_mask = result.new_zeros(result.size())
 608             ar_mask[:, self.height * self.width :] = 1
 609             result *= 1 - ar_mask
 610             masked_inplace_autoregression(
 611                 model, self.batch_size, result, ar_mask, device=self.device
 612             )
 613             mazes, paths = self.seq2map(result)
 614             nb_correct += maze.path_correctness(mazes, paths).long().sum()
 615             nb_total += mazes.size(0)
 616
 617         return nb_total, nb_correct
 618
 619     def produce_results(self, n_epoch, model):
 620         with torch.autograd.no_grad():
 621             t = model.training
 622             model.eval()
 623
 624             train_nb_total, train_nb_correct = self.compute_error(
 625                 model, "train", nb_to_use=1000
 626             )
 627             log_string(
 628                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 629             )
 630
 631             test_nb_total, test_nb_correct = self.compute_error(
 632                 model, "test", nb_to_use=1000
 633             )
 634             log_string(
 635                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 636             )
 637
 638             input = self.test_input[:48]
 639             result = input.clone()
 640             ar_mask = result.new_zeros(result.size())
 641             ar_mask[:, self.height * self.width :] = 1
 642             result *= 1 - ar_mask
 643             masked_inplace_autoregression(
 644                 model, self.batch_size, result, ar_mask, device=self.device
 645             )
 646
 647             mazes, paths = self.seq2map(input)
 648             _, predicted_paths = self.seq2map(result)
 649
 650             filename = os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.png")
 651             maze.save_image(
 652                 filename,
 653                 mazes=mazes,
 654                 target_paths=paths,
 655                 predicted_paths=predicted_paths,
 656                 path_correct=maze.path_correctness(mazes, predicted_paths),
 657             )
 658             log_string(f"wrote {filename}")
 659
 660             model.train(t)
 661
 662
 663 ######################################################################
 664
 665
 666 def generate_snake_sequences(
 667     nb, height, width, nb_colors, length, prompt_length, device=torch.device("cpu")
 668 ):
 669     worlds = torch.randint(nb_colors, (nb, height, width), device=device)
 670     nb_prior_visits = torch.zeros(nb, height, width, device=device)
 671
 672     # nb x 2
 673     snake_position = torch.cat(
 674         (
 675             torch.randint(height, (nb, 1), device=device),
 676             torch.randint(width, (nb, 1), device=device),
 677         ),
 678         1,
 679     )
 680     snake_direction = torch.randint(4, (nb,), device=device)
 681     sequences = torch.empty(nb, 2 * length, device=device, dtype=torch.int64)
 682     sequences_prior_visits = torch.zeros(
 683         nb, 2 * length, device=device, dtype=torch.int64
 684     )
 685     i = torch.arange(nb, device=device)  # [:,None]
 686
 687     for l in range(length):
 688         # nb x 3
 689         snake_next_direction = torch.cat(
 690             (
 691                 (snake_direction[:, None] - 1) % 4,
 692                 snake_direction[:, None],
 693                 (snake_direction[:, None] + 1) % 4,
 694             ),
 695             1,
 696         )
 697
 698         # nb x 3
 699         vh = (snake_next_direction + 1) % 2 * (snake_next_direction - 1)
 700         vw = snake_next_direction % 2 * (snake_next_direction - 2)
 701
 702         # nb x 3 x 2
 703         snake_next_speed = torch.cat((vh[:, :, None], vw[:, :, None]), 2)
 704         snake_next_position = snake_position[:, None, :] + snake_next_speed
 705
 706         # nb x 3
 707         val = torch.logical_and(
 708             torch.logical_and(
 709                 snake_next_position[:, :, 0] >= 0, snake_next_position[:, :, 0] < height
 710             ),
 711             torch.logical_and(
 712                 snake_next_position[:, :, 1] >= 0, snake_next_position[:, :, 1] < width
 713             ),
 714         ).float()
 715         val = (
 716             # The multiplicative factors bias toward moving forward
 717             torch.rand_like(val)
 718             * val
 719             * torch.tensor([[1.0, 2.0, 1.0]], device=device)
 720         )
 721
 722         # nb
 723         j = val.argmax(1)
 724         snake_direction = snake_next_direction[i, j]
 725
 726         sequences[:, 2 * l] = worlds[i, snake_position[:, 0], snake_position[:, 1]] + 4
 727         sequences_prior_visits[:, 2 * l] = nb_prior_visits[
 728             i, snake_position[:, 0], snake_position[:, 1]
 729         ]
 730         if l < prompt_length:
 731             nb_prior_visits[i, snake_position[:, 0], snake_position[:, 1]] += 1
 732         sequences[:, 2 * l + 1] = snake_direction
 733
 734         # nb x 2
 735         snake_position = snake_next_position[i, j]
 736
 737     return sequences, sequences_prior_visits
 738
 739
 740 # generate_snake_sequences(nb=1, height=4, width=6, nb_colors=3, length=20)
 741 # exit(0)
 742
 743
 744 def snake_solver(input, ar_mask):
 745     for n in range(input.size(0)):
 746         i, j, memory = 0, 0, {}
 747         # print(input[n])
 748         # print(ar_mask[n])
 749         for l in range(input.size(1) // 2):
 750             if ar_mask[n, 2 * l] == 1:
 751                 if memory.get((i, j)) is None:
 752                     input[n, 2 * l] = -1
 753                 else:
 754                     input[n, 2 * l] = memory[(i, j)]
 755             else:
 756                 # print(f'@3 {memory=}')
 757                 if memory.get((i, j)) is None:
 758                     memory[(i, j)] = input[n, 2 * l]
 759                 else:
 760                     assert memory[(i, j)] == input[n, 2 * l], f"n={n} l={l}"
 761             # print(f'@1 {i=} {j=}')
 762             d = input[n, 2 * l + 1].item()
 763             i += (d + 1) % 2 * (d - 1)
 764             j += d % 2 * (d - 2)
 765             # print(f'@2 {i=} {j=}')
 766
 767
 768 class TaskSnake(Task):
 769     def __init__(
 770         self,
 771         nb_train_samples,
 772         nb_test_samples,
 773         batch_size,
 774         height,
 775         width,
 776         nb_colors,
 777         length,
 778         prompt_length,
 779         device=torch.device("cpu"),
 780     ):
 781         self.batch_size = batch_size
 782         self.height = height
 783         self.width = width
 784         self.device = device
 785         self.prompt_length = prompt_length
 786
 787         self.train_input, self.train_prior_visits = generate_snake_sequences(
 788             nb_train_samples,
 789             height,
 790             width,
 791             nb_colors,
 792             length,
 793             prompt_length,
 794             self.device,
 795         )
 796         self.test_input, self.test_prior_visits = generate_snake_sequences(
 797             nb_test_samples,
 798             height,
 799             width,
 800             nb_colors,
 801             length,
 802             prompt_length,
 803             self.device,
 804         )
 805
 806         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 807
 808     def batches(self, split="train", nb_to_use=-1, desc=None):
 809         assert split in {"train", "test"}
 810         input = self.train_input if split == "train" else self.test_input
 811         if nb_to_use > 0:
 812             input = input[:nb_to_use]
 813         if desc is None:
 814             desc = f"epoch-{split}"
 815         for batch in tqdm.tqdm(
 816             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 817         ):
 818             yield batch
 819
 820     def vocabulary_size(self):
 821         return self.nb_codes
 822
 823     def produce_results(self, n_epoch, model):
 824         with torch.autograd.no_grad():
 825             t = model.training
 826             model.eval()
 827
 828             def compute_nb_correct(input, prior_visits):
 829                 result = input.clone()
 830                 i = torch.arange(result.size(1), device=result.device)[None, :]
 831                 ar_mask = (
 832                     torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
 833                     .long()
 834                     .expand_as(result)
 835                 )
 836                 result *= 1 - ar_mask
 837
 838                 # snake_solver(result,ar_mask)
 839
 840                 masked_inplace_autoregression(
 841                     model, self.batch_size, result, ar_mask, device=self.device
 842                 )
 843
 844                 nb_total = ((prior_visits > 0) * ar_mask).sum()
 845
 846                 nb_correct = (
 847                     (result == input).long() * (prior_visits > 0) * ar_mask
 848                 ).sum()
 849
 850                 # nb_total = result.size(0)
 851                 # nb_correct = ((result - input).abs().sum(1) == 0).sum()
 852
 853                 return nb_total, nb_correct
 854
 855             # train_nb_total, train_nb_correct = compute_nb_correct(
 856             # self.train_input, self.train_prior_visits
 857             # )
 858
 859             # log_string(
 860             # f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 861             # )
 862
 863             test_nb_total, test_nb_correct = compute_nb_correct(
 864                 self.test_input[:1000], self.test_prior_visits[:1000]
 865             )
 866
 867             log_string(
 868                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 869             )
 870
 871             model.train(t)
 872
 873
 874 ######################################################################
 875
 876
 877 def picoclvr_pruner_horizontal_green(p):
 878     return not ("green" in p and ("left" in p or "right" in p))
 879
 880
 881 picoclvr_pruner_train = (
 882     picoclvr_pruner_horizontal_green
 883     if args.picocvlr_prune_properties in {"train+eval"}
 884     else None
 885 )
 886
 887 picoclvr_pruner_eval = (
 888     (lambda p: not picoclvr_pruner_horizontal_green(p))
 889     if args.picocvlr_prune_properties in {"train+eval", "eval"}
 890     else None
 891 )
 892
 893 ######################################################################
 894
 895 if args.task == "picoclvr":
 896     task = TaskPicoCLVR(
 897         nb_train_samples=args.nb_train_samples,
 898         nb_test_samples=args.nb_test_samples,
 899         batch_size=args.batch_size,
 900         height=args.picoclvr_height,
 901         width=args.picoclvr_width,
 902         nb_colors=args.picoclvr_nb_colors,
 903         device=device,
 904         pruner_train=picoclvr_pruner_train,
 905         pruner_eval=picoclvr_pruner_eval,
 906     )
 907
 908 elif args.task == "mnist":
 909     task = TaskMNIST(
 910         batch_size=args.batch_size,
 911         device=device,
 912     )
 913
 914 elif args.task == "maze":
 915     task = TaskMaze(
 916         nb_train_samples=args.nb_train_samples,
 917         nb_test_samples=args.nb_test_samples,
 918         batch_size=args.batch_size,
 919         height=args.maze_height,
 920         width=args.maze_width,
 921         nb_walls=args.maze_nb_walls,
 922         device=device,
 923     )
 924
 925 elif args.task == "snake":
 926     task = TaskSnake(
 927         nb_train_samples=args.nb_train_samples,
 928         nb_test_samples=args.nb_test_samples,
 929         batch_size=args.batch_size,
 930         height=args.snake_height,
 931         width=args.snake_width,
 932         nb_colors=args.snake_nb_colors,
 933         length=args.snake_length,
 934         prompt_length=args.snake_length // 2,
 935         device=device,
 936     )
 937
 938 else:
 939     raise ValueError(f"Unknown task {args.task}")
 940
 941 ######################################################################
 942
 943 log_string(f"device {device}")
 944
 945 vocabulary_size = task.vocabulary_size()
 946
 947 log_string(f"vocabulary_size {vocabulary_size}")
 948
 949 ##############################
 950
 951 model = mygpt.MyGPT(
 952     vocabulary_size=vocabulary_size,
 953     dim_model=args.dim_model,
 954     dim_keys=args.dim_keys,
 955     dim_hidden=args.dim_hidden,
 956     nb_heads=args.nb_heads,
 957     nb_blocks=args.nb_blocks,
 958     causal=True,
 959     dropout=args.dropout,
 960 )
 961
 962 model.to(device)
 963
 964 nb_parameters = sum(p.numel() for p in model.parameters())
 965 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 966
 967 ######################################################################
 968
 969 nb_epochs_finished = 0
 970
 971 if args.no_checkpoint:
 972     log_string(f"not trying to load checkpoint.")
 973
 974 else:
 975     try:
 976         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 977         checkpoint = torch.load(checkpoint_name)
 978         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 979         model.load_state_dict(checkpoint["model_state"])
 980         torch.set_rng_state(checkpoint["rng_state"])
 981         if torch.cuda.is_available():
 982             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 983
 984         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 985
 986     except FileNotFoundError:
 987         log_string("starting from scratch.")
 988
 989     except:
 990         log_string("error when loading the checkpoint.")
 991         exit(1)
 992
 993 ######################################################################
 994
 995 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 996
 997 token_count = 0
 998 for input in task.batches(split="train"):
 999     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
1000 token_probas = token_count / token_count.sum()
1001 entropy = -torch.xlogy(token_probas, token_probas).sum()
1002 train_set_perplexity = math.exp(entropy)
1003
1004 ##############################
1005
1006 if args.learning_rate_schedule == "cos":
1007     learning_rate_schedule = {}
1008     for n_epoch in range(args.nb_epochs):
1009         u = n_epoch / args.nb_epochs * math.pi
1010         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
1011 else:
1012     u = {
1013         int(k): float(v)
1014         for k, v in [
1015             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
1016         ]
1017     }
1018
1019     learning_rate_schedule = {}
1020     learning_rate = args.learning_rate
1021     for n_epoch in range(args.nb_epochs):
1022         if n_epoch in u:
1023             learning_rate = u[n_epoch]
1024         learning_rate_schedule[n_epoch] = learning_rate
1025
1026 log_string(f"learning_rate_schedule {learning_rate_schedule}")
1027
1028 ##############################
1029
1030 nb_samples_seen = 0
1031
1032 if nb_epochs_finished >= nb_epochs:
1033     task.produce_results(nb_epochs_finished, model)
1034
1035 for n_epoch in range(nb_epochs_finished, nb_epochs):
1036     learning_rate = learning_rate_schedule[n_epoch]
1037
1038     log_string(f"learning_rate {learning_rate}")
1039
1040     if args.optim == "sgd":
1041         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
1042     elif args.optim == "adam":
1043         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
1044     elif args.optim == "adamw":
1045         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
1046     else:
1047         raise ValueError(f"Unknown optimizer {args.optim}.")
1048
1049     model.train()
1050
1051     nb_train_samples, acc_train_loss = 0, 0.0
1052
1053     for input in task.batches(split="train"):
1054         input = input.to(device)
1055         output = model(mygpt.BracketedSequence(input)).x
1056         loss = F.cross_entropy(output.transpose(1, 2), input)
1057         acc_train_loss += loss.item() * input.size(0)
1058         nb_train_samples += input.size(0)
1059         nb_samples_seen += input.size(0)
1060
1061         optimizer.zero_grad()
1062         loss.backward()
1063         optimizer.step()
1064
1065     with torch.autograd.no_grad():
1066         model.eval()
1067
1068         nb_test_samples, acc_test_loss = 0, 0.0
1069
1070         for input in task.batches(split="test"):
1071             input = input.to(device)
1072
1073             # input, loss_masks, true_images = task.excise_last_image(input)
1074             # input, loss_masks = task.add_true_image(input, true_images, loss_masks)
1075
1076             output = model(mygpt.BracketedSequence(input)).x
1077             loss = F.cross_entropy(output.transpose(1, 2), input)
1078             acc_test_loss += loss.item() * input.size(0)
1079             nb_test_samples += input.size(0)
1080
1081         train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
1082         test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
1083
1084         log_string(
1085             f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
1086         )
1087
1088         task.produce_results(n_epoch, model)
1089
1090     checkpoint = {
1091         "nb_epochs_finished": n_epoch + 1,
1092         "model_state": model.state_dict(),
1093         "rng_state": torch.get_rng_state(),
1094     }
1095
1096     if torch.cuda.is_available():
1097         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
1098
1099     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
1100     torch.save(checkpoint, checkpoint_name)
1101     log_string(f"saved checkpoint {checkpoint_name}")
1102
1103 ######################################################################