tasks.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 import math, os, tqdm, warnings
   9
  10 import torch, torchvision
  11
  12 from torch import nn
  13 from torch.nn import functional as F
  14
  15 from mygpt import BracketedSequence
  16
  17 # from graph import save_attention_image
  18 save_attention_image = None
  19
  20 ######################################################################
  21
  22
  23 def masked_inplace_autoregression(
  24     model,
  25     batch_size,
  26     input,
  27     ar_mask,
  28     deterministic_synthesis,
  29     forbidden_tokens=None,
  30     logit_biases=None,
  31     progress_bar_desc="autoregression",
  32     device=torch.device("cpu"),
  33 ):
  34     assert input.size() == ar_mask.size()
  35
  36     batches = zip(input.split(batch_size), ar_mask.split(batch_size))
  37
  38     if progress_bar_desc is not None:
  39         batches = tqdm.tqdm(
  40             batches,
  41             dynamic_ncols=True,
  42             desc=progress_bar_desc,
  43             total=(input.size(0) + batch_size - 1) // batch_size,
  44         )
  45
  46     with torch.autograd.no_grad():
  47         t = model.training
  48         model.eval()
  49
  50         for input, ar_mask in batches:
  51             model.masked_inplace_autoregression(
  52                 input,
  53                 ar_mask,
  54                 deterministic_synthesis,
  55                 forbidden_tokens,
  56                 logit_biases,
  57             )
  58
  59         model.train(t)
  60
  61
  62 ######################################################################
  63
  64
  65 class Task:
  66     def batches(self, split="train", nb_to_use=-1, desc=None):
  67         pass
  68
  69     def vocabulary_size(self):
  70         pass
  71
  72     def produce_results(
  73         self, n_epoch, model, result_dir, logger, deterministic_synthesis
  74     ):
  75         pass
  76
  77
  78 class TaskFromFile(Task):
  79     def tensorize(self, pairs, shuffle):
  80         len_max = max([len(x[0]) for x in pairs])
  81
  82         input = torch.cat(
  83             [
  84                 torch.tensor(
  85                     [
  86                         [self.char2id[c] for c in s[0] + "#" * (len_max - len(s[0]))]
  87                         for s in pairs
  88                     ]
  89                 )
  90             ],
  91             0,
  92         ).to("cpu")
  93
  94         pred_mask = torch.cat(
  95             [
  96                 torch.tensor(
  97                     [
  98                         [int(c) for c in s[1] + "0" * (len_max - len(s[1]))]
  99                         for s in pairs
 100                     ]
 101                 )
 102             ],
 103             0,
 104         ).to("cpu")
 105
 106         if shuffle:
 107             i = torch.randperm(input.size(0))
 108             input = input[i].contiguous()
 109             pred_mask = pred_mask[i].contiguous()
 110
 111         return input, pred_mask
 112
 113     # trim all the tensors in the tuple z to remove as much token from
 114     # left and right in the first tensor. If z is a tuple, all its
 115     # elements are trimed according to the triming for the first
 116     def trim(self, z, token="#"):
 117         n = self.char2id[token]
 118         if type(z) == tuple:
 119             x = z[0]
 120             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 121             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 122             return tuple([t[:, a:b] for t in z])
 123         else:
 124             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 125             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 126             return z[:, a:b]
 127
 128     def __init__(
 129         self,
 130         train_filename,
 131         test_filename,
 132         nb_train_samples,
 133         nb_test_samples,
 134         batch_size,
 135         shuffle=False,
 136         device=torch.device("cpu"),
 137     ):
 138         self.batch_size = batch_size
 139         self.device = device
 140
 141         def read_file(filename, nb=-1):
 142             pairs = []
 143             with open(filename, "r") as f:
 144                 while True:
 145                     sequence = f.readline().strip()
 146                     if not sequence:
 147                         break
 148                     pred_mask = f.readline().strip()
 149                     assert len(sequence) == len(pred_mask)
 150                     assert set(pred_mask).issubset({"0", "1", "2"}), f"{set(pred_mask)}"
 151                     pairs.append((sequence, pred_mask))
 152                     if len(pairs) == nb:
 153                         break
 154
 155             if nb > 0:
 156                 pairs = pairs[:nb]
 157                 assert len(pairs) == nb
 158
 159             return pairs
 160
 161         train_pairs = read_file(train_filename, nb_train_samples)
 162         test_pairs = read_file(test_filename, nb_test_samples)
 163
 164         symbols = ["#"] + list(
 165             set("".join([x[0] for x in train_pairs + test_pairs])) - set(["#"])
 166         )
 167         self.char2id = dict([(c, n) for n, c in enumerate(symbols)])
 168         self.id2char = dict([(n, c) for c, n in self.char2id.items()])
 169
 170         self.train_input, self.train_pred_masks = self.tensorize(
 171             train_pairs, shuffle=shuffle
 172         )
 173         self.test_input, self.test_pred_masks = self.tensorize(
 174             test_pairs, shuffle=shuffle
 175         )
 176
 177     def batches(self, split="train", nb_to_use=-1, desc=None):
 178         assert split in {"train", "test"}
 179         input = self.train_input if split == "train" else self.test_input
 180         if nb_to_use > 0:
 181             input = input[:nb_to_use]
 182         if desc is None:
 183             desc = f"epoch-{split}"
 184         for batch in tqdm.tqdm(
 185             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 186         ):
 187             yield self.trim(batch).to(self.device)
 188
 189     def vocabulary_size(self):
 190         return len(self.char2id)
 191
 192     def tensor2str(self, t):
 193         return ["".join([self.id2char[x.item()] for x in s]) for s in t]
 194
 195     def produce_results(
 196         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 197     ):
 198         correct = self.trim(self.test_input[:1000]).to(self.device)
 199         result = correct.clone()
 200         pred_mask = self.test_pred_masks[:1000, : result.size(1)].to(self.device)
 201         ar_mask = (pred_mask > 0).long()
 202         result *= 1 - ar_mask  # paraaaaanoiaaaaaaa
 203
 204         logger(f"----------------------------------------------------------")
 205
 206         for e in self.tensor2str(result[:50]):
 207             logger(f"test_before {e}")
 208
 209         masked_inplace_autoregression(
 210             model,
 211             self.batch_size,
 212             result,
 213             ar_mask,
 214             deterministic_synthesis,
 215             device=self.device,
 216         )
 217
 218         logger(f"----------------------------------------------------------")
 219
 220         for e, c in zip(self.tensor2str(result[:50]), self.tensor2str(correct[:50])):
 221             logger(f"test_after  {e}")
 222             logger(f"correct     {c}")
 223
 224         logger(f"----------------------------------------------------------")
 225
 226         err_mask = (pred_mask == 2).long()
 227         nb_total = err_mask.sum().item()
 228         nb_correct = ((correct == result).long() * err_mask).sum().item()
 229
 230         logger(f"test_performance {n_epoch} {nb_total=} {nb_correct=}")
 231         logger(f"main_test_accuracy {n_epoch} {nb_correct / nb_total}")
 232
 233
 234 ####################
 235
 236 import problems
 237
 238
 239 class SandBox(Task):
 240     def __init__(
 241         self,
 242         problem,
 243         nb_train_samples,
 244         nb_test_samples,
 245         batch_size,
 246         logger=None,
 247         device=torch.device("cpu"),
 248         max_nb_codes=1024,
 249     ):
 250         super().__init__()
 251
 252         self.batch_size = batch_size
 253         self.device = device
 254         self.problem = problem
 255
 256         self.train_input, self.train_ar_mask = self.problem.generate_sequences(
 257             nb_train_samples
 258         )
 259         self.test_input, self.test_ar_mask = self.problem.generate_sequences(
 260             nb_test_samples
 261         )
 262
 263         self.train_input, self.train_ar_mask = self.train_input.to(
 264             device
 265         ), self.train_ar_mask.to(device)
 266         self.test_input, self.test_ar_mask = self.test_input.to(
 267             device
 268         ), self.test_ar_mask.to(device)
 269
 270         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 271
 272         # A bit of paranoia never hurts
 273         assert self.nb_codes <= max_nb_codes
 274         assert self.train_input.min() >= 0
 275         assert self.test_input.min() >= 0
 276         assert tuple(x.item() for x in self.train_ar_mask.unique()) in {
 277             (0,),
 278             (1,),
 279             (0, 1),
 280         }
 281         assert tuple(x.item() for x in self.test_ar_mask.unique()) in {
 282             (0,),
 283             (1,),
 284             (0, 1),
 285         }
 286
 287         if logger is not None:
 288             for s, a in zip(self.train_input[:100], self.train_ar_mask[:100]):
 289                 logger(f"train_sequences {self.problem.seq2str(s)}")
 290                 a = "".join(["01"[x.item()] for x in a])
 291                 logger(f"                {a}")
 292
 293     def batches(self, split="train", nb_to_use=-1, desc=None):
 294         assert split in {"train", "test"}
 295         input = self.train_input if split == "train" else self.test_input
 296         if nb_to_use > 0:
 297             input = input[:nb_to_use]
 298         if desc is None:
 299             desc = f"epoch-{split}"
 300         for batch in tqdm.tqdm(
 301             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 302         ):
 303             yield batch
 304
 305     def vocabulary_size(self):
 306         return self.nb_codes
 307
 308     def produce_results(
 309         self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
 310     ):
 311         def compute_accuracy(input, ar_mask, logger=None):
 312             input, ar_mask = input[:nmax], ar_mask[:nmax]
 313             result = input.clone() * (1 - ar_mask)
 314
 315             masked_inplace_autoregression(
 316                 model,
 317                 self.batch_size,
 318                 result,
 319                 ar_mask,
 320                 deterministic_synthesis,
 321                 progress_bar_desc=None,
 322                 device=self.device,
 323             )
 324
 325             log_ground_truth = ar_mask.min() == 0
 326
 327             if logger is not None:
 328                 for sp, st in zip(result[:10], input[:10]):
 329                     logger(
 330                         f"test_sequences {n_epoch} prediction   {self.problem.seq2str(sp)}"
 331                     )
 332                     if log_ground_truth:
 333                         logger(
 334                             f"               {n_epoch} ground truth {self.problem.seq2str(st)}"
 335                         )
 336
 337             nb_total, nb_correct = self.problem.compute_nb_correct(
 338                 input, ar_mask, result
 339             )
 340
 341             # nb_total = ar_mask.sum().item()
 342             # nb_correct = ((result == input).long() * ar_mask).sum().item()
 343
 344             return nb_total, nb_correct
 345
 346         train_nb_total, train_nb_correct = compute_accuracy(
 347             self.train_input, self.train_ar_mask
 348         )
 349
 350         logger(
 351             f"accuracy_train {n_epoch} nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 352         )
 353
 354         test_nb_total, test_nb_correct = compute_accuracy(
 355             self.test_input, self.test_ar_mask, logger
 356         )
 357
 358         logger(
 359             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 360         )
 361
 362         logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
 363
 364         if save_attention_image is not None:
 365             for k in range(10):
 366                 ns = torch.randint(self.test_input.size(0), (1,)).item()
 367                 input = self.test_input[ns : ns + 1].clone()
 368
 369                 with torch.autograd.no_grad():
 370                     t = model.training
 371                     model.eval()
 372                     # model.record_attention(True)
 373                     model(BracketedSequence(input))
 374                     model.train(t)
 375                     # ram = model.retrieve_attention()
 376                     # model.record_attention(False)
 377
 378                 # tokens_output = [c for c in self.problem.seq2str(input[0])]
 379                 # tokens_input = ["n/a"] + tokens_output[:-1]
 380                 # for n_head in range(ram[0].size(1)):
 381                 # filename = os.path.join(
 382                 # result_dir, f"sandbox_attention_{k}_h{n_head}.pdf"
 383                 # )
 384                 # attention_matrices = [m[0, n_head] for m in ram]
 385                 # save_attention_image(
 386                 # filename,
 387                 # tokens_input,
 388                 # tokens_output,
 389                 # attention_matrices,
 390                 # k_top=10,
 391                 ##min_total_attention=0.9,
 392                 # token_gap=12,
 393                 # layer_gap=50,
 394                 # )
 395                 # logger(f"wrote {filename}")
 396
 397
 398 ######################################################################
 399
 400 import world
 401
 402
 403 class World(Task):
 404     def __init__(
 405         self,
 406         nb_train_samples,
 407         nb_test_samples,
 408         batch_size,
 409         logger=None,
 410         device=torch.device("cpu"),
 411     ):
 412         super().__init__()
 413
 414         self.batch_size = batch_size
 415         self.device = device
 416         self.height = 6
 417         self.width = 8
 418
 419         self.train_input = world.generate(
 420             nb_train_samples, height=self.height, width=self.width
 421         )
 422         self.train_ar_mask = (
 423             (torch.arange(self.train_input.size(1)) > self.train_input.size(1) // 2)
 424             .long()[None, :]
 425             .expand_as(self.train_input)
 426         )
 427
 428         self.test_input = world.generate(
 429             nb_test_samples, height=self.height, width=self.width
 430         )
 431         self.test_ar_mask = (
 432             (torch.arange(self.test_input.size(1)) > self.test_input.size(1) // 2)
 433             .long()[None, :]
 434             .expand_as(self.test_input)
 435         )
 436
 437         self.train_input, self.train_ar_mask = self.train_input.to(
 438             device
 439         ), self.train_ar_mask.to(device)
 440         self.test_input, self.test_ar_mask = self.test_input.to(
 441             device
 442         ), self.test_ar_mask.to(device)
 443
 444         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 445
 446     def batches(self, split="train", nb_to_use=-1, desc=None):
 447         assert split in {"train", "test"}
 448         input = self.train_input if split == "train" else self.test_input
 449         if nb_to_use > 0:
 450             input = input[:nb_to_use]
 451         if desc is None:
 452             desc = f"epoch-{split}"
 453         for batch in tqdm.tqdm(
 454             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 455         ):
 456             yield batch
 457
 458     def vocabulary_size(self):
 459         return self.nb_codes
 460
 461     def produce_results(
 462         self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
 463     ):
 464         def compute_accuracy(input, ar_mask, logger=None):
 465             input, ar_mask = input[:nmax], ar_mask[:nmax]
 466             result = input.clone() * (1 - ar_mask)
 467
 468             masked_inplace_autoregression(
 469                 model,
 470                 self.batch_size,
 471                 result,
 472                 ar_mask,
 473                 deterministic_synthesis,
 474                 progress_bar_desc=None,
 475                 device=self.device,
 476             )
 477
 478             nb_total, nb_correct = (
 479                 input.size(0),
 480                 (input == result).long().min(dim=1).values.sum(),
 481             )
 482
 483             return nb_total, nb_correct
 484
 485         train_nb_total, train_nb_correct = compute_accuracy(
 486             self.train_input, self.train_ar_mask
 487         )
 488
 489         logger(
 490             f"accuracy_train {n_epoch} nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 491         )
 492
 493         test_nb_total, test_nb_correct = compute_accuracy(
 494             self.test_input, self.test_ar_mask, logger
 495         )
 496
 497         logger(
 498             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 499         )
 500
 501         logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
 502
 503         ##############################
 504
 505         input, ar_mask = self.test_input[:64], self.test_ar_mask[:64]
 506         result = input.clone() * (1 - ar_mask)
 507
 508         masked_inplace_autoregression(
 509             model,
 510             self.batch_size,
 511             result,
 512             ar_mask,
 513             deterministic_synthesis,
 514             progress_bar_desc=None,
 515             device=self.device,
 516         )
 517
 518         img = world.sample2img(result.to("cpu"), self.height, self.width)
 519
 520         image_name = os.path.join(result_dir, f"world_result_{n_epoch:04d}.png")
 521         torchvision.utils.save_image(img.float() / 255.0, image_name, nrow=8, padding=2)
 522         logger(f"wrote {image_name}")
 523
 524
 525 ######################################################################
 526
 527 import picoclvr
 528
 529
 530 class PicoCLVR(Task):
 531     # Make a tensor from a list of strings
 532     def tensorize(self, descr):
 533         token_descr = [s.strip().split(" ") for s in descr]
 534         l = max([len(s) for s in token_descr])
 535         token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
 536         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
 537         return torch.tensor(id_descr, device=self.device)
 538
 539     # Make a list of strings from a tensor
 540     def detensorize(self, x):
 541         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
 542
 543     # trim all the tensors in the tuple z to remove as much token from
 544     # left and right in the first tensor. If z is a tuple, all its
 545     # elements are trimed according to the triming for the first
 546     def trim(self, z, token="<nul>"):
 547         n = self.token2id[token]
 548         if type(z) == tuple:
 549             x = z[0]
 550             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 551             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 552             return tuple([t[:, a:b] for t in z])
 553         else:
 554             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 555             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 556             return z[:, a:b]
 557
 558     ######################
 559
 560     def __init__(
 561         self,
 562         nb_train_samples,
 563         nb_test_samples,
 564         batch_size,
 565         height,
 566         width,
 567         nb_colors=5,
 568         logger=None,
 569         device=torch.device("cpu"),
 570         pruner_train=None,
 571         pruner_eval=None,
 572     ):
 573         super().__init__()
 574
 575         def generate_descr(nb, cache_suffix, pruner):
 576             return picoclvr.generate(
 577                 nb,
 578                 height=self.height,
 579                 width=self.width,
 580                 nb_colors=nb_colors,
 581                 pruner=pruner,
 582             )
 583
 584         self.height = height
 585         self.width = width
 586         self.batch_size = batch_size
 587         self.device = device
 588         self.pruner_train = pruner_train
 589         self.pruner_eval = pruner_eval
 590
 591         if logger is not None:
 592             logger(
 593                 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
 594             )
 595
 596         self.train_descr = generate_descr(
 597             nb_train_samples, "train", pruner=self.pruner_train
 598         )
 599         self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
 600
 601         # Build the tokenizer
 602         tokens = {"<nul>", "<img>"}
 603         for d in [self.train_descr, self.test_descr]:
 604             for s in d:
 605                 for t in s.strip().split(" "):
 606                     tokens.add(t)
 607         # make this set a sorted list to get the same tensors given
 608         # the same descr
 609         tokens = list(tokens)
 610         tokens.sort()
 611         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
 612         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
 613         self.t_img, self.t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 614
 615         # Tokenize the train and test sets
 616         self.train_input = self.tensorize(self.train_descr)
 617         self.test_input = self.tensorize(self.test_descr)
 618
 619     def batches(self, split="train", nb_to_use=-1, desc=None):
 620         assert split in {"train", "test"}
 621         input = self.train_input if split == "train" else self.test_input
 622         for batch in tqdm.tqdm(
 623             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 624         ):
 625             yield self.trim(batch)
 626
 627     def vocabulary_size(self):
 628         return len(self.token2id)
 629
 630     def compute_missing_properties(
 631         self, n_epoch, model, logger, deterministic_synthesis, pruner=None
 632     ):
 633         acc_nb_requested_properties = []
 634         acc_nb_missing_properties = []
 635         acc_nb_results = 0
 636
 637         for input in tqdm.tqdm(
 638             self.test_input.split(self.batch_size),
 639             dynamic_ncols=True,
 640             desc=f"test-properties",
 641         ):
 642             result = input.clone()
 643             ar_mask = (result == self.t_img).long().cumsum(dim=1).clamp(max=1)
 644             result = (1 - ar_mask) * result + ar_mask * self.t_nul
 645             masked_inplace_autoregression(
 646                 model,
 647                 self.batch_size,
 648                 result,
 649                 ar_mask,
 650                 deterministic_synthesis,
 651                 progress_bar_desc=None,
 652                 device=self.device,
 653             )
 654
 655             result_descr = self.detensorize(result)
 656             np = picoclvr.nb_properties(
 657                 result_descr,
 658                 height=self.height,
 659                 width=self.width,
 660                 pruner=pruner,
 661             )
 662             nb_requested_properties, _, nb_missing_properties = zip(*np)
 663             acc_nb_requested_properties += nb_requested_properties
 664             acc_nb_missing_properties += nb_missing_properties
 665             acc_nb_results += len(result_descr)
 666
 667         nb_requested_properties = sum(acc_nb_requested_properties)
 668         nb_missing_properties = sum(acc_nb_missing_properties)
 669
 670         prefix = "" if pruner is None else "pruned_"
 671         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 672         logger(
 673             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 674         )
 675         logger(
 676             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 677         )
 678
 679         logger(
 680             f"main_test_accuracy {n_epoch} {1-nb_missing_properties/nb_requested_properties}"
 681         )
 682
 683     ######################################################################
 684
 685     def produce_results(
 686         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 687     ):
 688         self.compute_missing_properties(n_epoch, model, logger, deterministic_synthesis)
 689
 690         if self.pruner_eval is not None:
 691             self.compute_missing_properties(n_epoch, model, self.pruner_eval)
 692
 693         nb_tokens_to_generate = self.height * self.width + 3
 694         result_descr = []
 695         nb_per_primer = 8
 696         primer = []
 697
 698         for primer_descr in [
 699             "red above green <sep> green top <sep> blue right of red",
 700             "there is red <sep> there is yellow <sep> there is blue",
 701             "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
 702             "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
 703         ]:
 704             primer += [primer_descr + " <img>"] * nb_per_primer
 705
 706         result = self.tensorize(primer)
 707         fill = result.new_full(
 708             result.size()[:-1] + (self.height * self.width + 1,), self.t_nul
 709         )
 710         result = torch.cat((result, fill), 1)
 711         ar_mask = (result == self.t_nul).long()
 712         masked_inplace_autoregression(
 713             model,
 714             self.batch_size,
 715             result,
 716             ar_mask,
 717             deterministic_synthesis,
 718             device=self.device,
 719         )
 720         result_descr = self.detensorize(result)
 721
 722         np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
 723
 724         acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
 725         acc_nb_results = len(result_descr)
 726
 727         nb_requested_properties = sum(acc_nb_requested_properties)
 728         nb_missing_properties = sum(acc_nb_missing_properties)
 729
 730         prefix = "demo_"
 731         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 732         logger(
 733             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 734         )
 735         logger(
 736             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 737         )
 738
 739         img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
 740
 741         if img.dim() == 5:
 742             if img.size(1) == 1:
 743                 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
 744             else:
 745                 img = torch.cat(
 746                     [
 747                         torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
 748                         for x in img
 749                     ],
 750                     0,
 751                 )
 752
 753         image_name = os.path.join(result_dir, f"picoclvr_result_{n_epoch:04d}.png")
 754         torchvision.utils.save_image(
 755             img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0
 756         )
 757         logger(f"wrote {image_name}")
 758
 759
 760 ######################################################################
 761
 762
 763 class MNIST(Task):
 764     def __init__(
 765         self, nb_train_samples, nb_test_samples, batch_size, device=torch.device("cpu")
 766     ):
 767         super().__init__()
 768
 769         self.nb_train_samples = (nb_train_samples,)
 770         self.nb_test_samples = (nb_test_samples,)
 771         self.batch_size = batch_size
 772         self.device = device
 773         data_set = torchvision.datasets.MNIST(root="./data", train=True, download=True)
 774         self.train_input = data_set.data[:nb_train_samples].view(-1, 28 * 28).long()
 775         data_set = torchvision.datasets.MNIST(root="./data", train=False, download=True)
 776         self.test_input = data_set.data[:nb_test_samples].view(-1, 28 * 28).long()
 777
 778     def batches(self, split="train", nb_to_use=-1, desc=None):
 779         assert split in {"train", "test"}
 780         input = self.train_input if split == "train" else self.test_input
 781         if nb_to_use > 0:
 782             input = input[:nb_to_use]
 783         if desc is None:
 784             desc = f"epoch-{split}"
 785         for batch in tqdm.tqdm(
 786             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 787         ):
 788             yield batch
 789
 790     def vocabulary_size(self):
 791         return 256
 792
 793     def produce_results(
 794         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 795     ):
 796         results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
 797         ar_mask = torch.full_like(results, 1)
 798         masked_inplace_autoregression(
 799             model,
 800             self.batch_size,
 801             results,
 802             ar_mask,
 803             deterministic_synthesis,
 804             device=self.device,
 805         )
 806         image_name = os.path.join(result_dir, f"mnist_result_{n_epoch:04d}.png")
 807         torchvision.utils.save_image(
 808             1 - results.reshape(-1, 1, 28, 28) / 255.0,
 809             image_name,
 810             nrow=16,
 811             pad_value=0.8,
 812         )
 813         logger(f"wrote {image_name}")
 814
 815
 816 ######################################################################
 817
 818 import maze
 819
 820
 821 class Maze(Task):
 822     def map2seq(self, *m):
 823         return torch.cat([x.flatten(1) for x in m], 1)
 824
 825     def seq2map(self, s):
 826         s = s.reshape(s.size(0), -1, self.height, self.width)
 827         return (s[:, k] for k in range(s.size(1)))
 828
 829     def __init__(
 830         self,
 831         nb_train_samples,
 832         nb_test_samples,
 833         batch_size,
 834         height,
 835         width,
 836         nb_walls,
 837         device=torch.device("cpu"),
 838     ):
 839         super().__init__()
 840
 841         self.batch_size = batch_size
 842         self.height = height
 843         self.width = width
 844         self.device = device
 845
 846         train_mazes, train_paths, _ = maze.create_maze_data(
 847             nb_train_samples,
 848             height=height,
 849             width=width,
 850             nb_walls=nb_walls,
 851             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 852         )
 853         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 854
 855         test_mazes, test_paths, _ = maze.create_maze_data(
 856             nb_test_samples,
 857             height=height,
 858             width=width,
 859             nb_walls=nb_walls,
 860             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 861         )
 862         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 863
 864         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 865
 866     def batches(self, split="train", nb_to_use=-1, desc=None):
 867         assert split in {"train", "test"}
 868         input = self.train_input if split == "train" else self.test_input
 869         if nb_to_use > 0:
 870             input = input[:nb_to_use]
 871         if desc is None:
 872             desc = f"epoch-{split}"
 873         for batch in tqdm.tqdm(
 874             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 875         ):
 876             yield batch
 877
 878     def vocabulary_size(self):
 879         return self.nb_codes
 880
 881     def compute_error(
 882         self, model, split="train", nb_to_use=-1, deterministic_synthesis=False
 883     ):
 884         model_device = next(model.parameters()).device
 885         nb_total, nb_correct = 0, 0
 886         count = torch.zeros(
 887             self.width * self.height,
 888             self.width * self.height,
 889             device=model_device,
 890             dtype=torch.int64,
 891         )
 892
 893         for input in self.batches(split, nb_to_use):
 894             input = input.to(model_device)
 895             result = input.clone()
 896             ar_mask = result.new_zeros(result.size())
 897             ar_mask[:, self.height * self.width :] = 1
 898             result *= 1 - ar_mask
 899             masked_inplace_autoregression(
 900                 model,
 901                 self.batch_size,
 902                 result,
 903                 ar_mask,
 904                 deterministic_synthesis,
 905                 progress_bar_desc=None,
 906                 device=self.device,
 907             )
 908             mazes, paths = self.seq2map(result)
 909             path_correctness = maze.path_correctness(mazes, paths)
 910             nb_correct += path_correctness.long().sum()
 911             nb_total += mazes.size(0)
 912
 913             optimal_path_lengths = (
 914                 (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
 915             )
 916             predicted_path_lengths = (
 917                 (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
 918             )
 919             optimal_path_lengths = optimal_path_lengths[path_correctness]
 920             predicted_path_lengths = predicted_path_lengths[path_correctness]
 921             count[optimal_path_lengths, predicted_path_lengths] += 1
 922
 923         if count.max() == 0:
 924             count = None
 925         else:
 926             count = count[
 927                 : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
 928             ]
 929
 930         return nb_total, nb_correct, count
 931
 932     def produce_results(
 933         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 934     ):
 935         train_nb_total, train_nb_correct, count = self.compute_error(
 936             model,
 937             "train",
 938             nb_to_use=1000,
 939             deterministic_synthesis=deterministic_synthesis,
 940         )
 941         logger(
 942             f"accuracy_train {n_epoch} nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 943         )
 944
 945         test_nb_total, test_nb_correct, count = self.compute_error(
 946             model,
 947             "test",
 948             nb_to_use=1000,
 949             deterministic_synthesis=deterministic_synthesis,
 950         )
 951         logger(
 952             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 953         )
 954
 955         logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
 956
 957         if count is not None:
 958             proportion_optimal = count.diagonal().sum().float() / count.sum()
 959             logger(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
 960             with open(
 961                 os.path.join(result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
 962             ) as f:
 963                 for i in range(count.size(0)):
 964                     for j in range(count.size(1)):
 965                         eol = " " if j < count.size(1) - 1 else "\n"
 966                         f.write(f"{count[i,j]}{eol}")
 967
 968         input = self.test_input[:48].to(next(model.parameters()).device)
 969         result = input.clone()
 970         ar_mask = result.new_zeros(result.size())
 971         ar_mask[:, self.height * self.width :] = 1
 972         result *= 1 - ar_mask
 973         masked_inplace_autoregression(
 974             model,
 975             self.batch_size,
 976             result,
 977             ar_mask,
 978             deterministic_synthesis,
 979             device=self.device,
 980         )
 981
 982         mazes, paths = self.seq2map(input)
 983         _, predicted_paths = self.seq2map(result)
 984
 985         filename = os.path.join(result_dir, f"maze_result_{n_epoch:04d}.png")
 986         maze.save_image(
 987             filename,
 988             mazes=mazes,
 989             target_paths=paths,
 990             predicted_paths=predicted_paths,
 991             path_correct=maze.path_correctness(mazes, predicted_paths),
 992             path_optimal=maze.path_optimality(paths, predicted_paths),
 993         )
 994         logger(f"wrote {filename}")
 995
 996
 997 ######################################################################
 998
 999
1000 import snake
1001
1002
1003 class Snake(Task):
1004     def __init__(
1005         self,
1006         nb_train_samples,
1007         nb_test_samples,
1008         batch_size,
1009         height,
1010         width,
1011         nb_colors,
1012         length,
1013         prompt_length,
1014         device=torch.device("cpu"),
1015     ):
1016         super().__init__()
1017
1018         self.batch_size = batch_size
1019         self.height = height
1020         self.width = width
1021         self.device = device
1022         self.prompt_length = prompt_length
1023
1024         self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
1025             nb_train_samples,
1026             height,
1027             width,
1028             nb_colors,
1029             length,
1030             prompt_length,
1031             self.device,
1032         )
1033         self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
1034             nb_test_samples,
1035             height,
1036             width,
1037             nb_colors,
1038             length,
1039             prompt_length,
1040             self.device,
1041         )
1042
1043         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1044
1045     def batches(self, split="train", nb_to_use=-1, desc=None):
1046         assert split in {"train", "test"}
1047         input = self.train_input if split == "train" else self.test_input
1048         if nb_to_use > 0:
1049             input = input[:nb_to_use]
1050         if desc is None:
1051             desc = f"epoch-{split}"
1052         for batch in tqdm.tqdm(
1053             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1054         ):
1055             yield batch
1056
1057     def vocabulary_size(self):
1058         return self.nb_codes
1059
1060     def produce_results(
1061         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1062     ):
1063         def compute_nb_correct(input, prior_visits):
1064             result = input.clone()
1065             i = torch.arange(result.size(1), device=result.device)[None, :]
1066             ar_mask = (
1067                 torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
1068                 .long()
1069                 .expand_as(result)
1070             )
1071             result *= 1 - ar_mask
1072
1073             masked_inplace_autoregression(
1074                 model,
1075                 self.batch_size,
1076                 result,
1077                 ar_mask,
1078                 deterministic_synthesis,
1079                 device=self.device,
1080             )
1081
1082             nb_total = ((prior_visits > 0) * ar_mask).sum()
1083
1084             nb_correct = ((result == input).long() * (prior_visits > 0) * ar_mask).sum()
1085
1086             return nb_total, nb_correct
1087
1088         test_nb_total, test_nb_correct = compute_nb_correct(
1089             self.test_input[:1000], self.test_prior_visits[:1000]
1090         )
1091
1092         logger(
1093             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
1094         )
1095
1096         logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
1097
1098
1099 ######################################################################
1100
1101
1102 import stack
1103
1104
1105 class Stack(Task):
1106     def __init__(
1107         self,
1108         nb_train_samples,
1109         nb_test_samples,
1110         batch_size,
1111         logger,
1112         nb_steps,
1113         nb_stacks,
1114         nb_digits,
1115         fraction_values_for_train=None,
1116         device=torch.device("cpu"),
1117     ):
1118         super().__init__()
1119
1120         self.batch_size = batch_size
1121         self.nb_steps = nb_steps
1122         self.nb_stacks = nb_stacks
1123         self.nb_digits = nb_digits
1124         self.device = device
1125
1126         if fraction_values_for_train is None:
1127             values_for_train = None
1128             values_for_test = None
1129         else:
1130             all = torch.randperm(10**nb_digits)
1131             nb_for_train = int(all.size(0) * fraction_values_for_train)
1132             values_for_train = all[:nb_for_train]
1133             values_for_test = all[nb_for_train:]
1134
1135         self.train_input, self.train_stack_counts = stack.generate_sequences(
1136             nb_train_samples,
1137             nb_steps,
1138             nb_stacks,
1139             nb_digits,
1140             values_for_train,
1141             self.device,
1142         )
1143
1144         self.test_input, self.test_stack_counts = stack.generate_sequences(
1145             nb_test_samples,
1146             nb_steps,
1147             nb_stacks,
1148             nb_digits,
1149             values_for_test,
1150             self.device,
1151         )
1152
1153         i = torch.logical_and(self.test_input % 2 == 1, self.test_input < 2 * nb_stacks)
1154         counts = self.test_stack_counts.flatten()[i.flatten()]
1155         counts = F.one_hot(counts).sum(0)
1156         logger(f"test_pop_stack_counts {counts}")
1157
1158         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1159
1160     def batches(self, split="train", nb_to_use=-1, desc=None):
1161         assert split in {"train", "test"}
1162         input = self.train_input if split == "train" else self.test_input
1163         if nb_to_use > 0:
1164             input = input[:nb_to_use]
1165         if desc is None:
1166             desc = f"epoch-{split}"
1167         for batch in tqdm.tqdm(
1168             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1169         ):
1170             yield batch
1171
1172     def vocabulary_size(self):
1173         return self.nb_codes
1174
1175     def produce_results(
1176         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1177     ):
1178         def compute_nb_correct(input):
1179             result = input.clone()
1180             stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
1181             ar_mask = (result != input).long()
1182             masked_inplace_autoregression(
1183                 model,
1184                 self.batch_size,
1185                 result,
1186                 ar_mask,
1187                 deterministic_synthesis,
1188                 device=self.device,
1189             )
1190
1191             errors = ((result != input).long() * ar_mask).reshape(
1192                 -1, 1 + self.nb_digits
1193             )
1194             ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits)
1195
1196             nb_total = ar_mask.max(1).values.sum()
1197             nb_correct = nb_total - errors.max(1).values.sum()
1198
1199             return nb_total, nb_correct
1200
1201         test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
1202
1203         logger(
1204             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
1205         )
1206
1207         logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
1208
1209         ##############################################################
1210         # Log a few generated sequences
1211         input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
1212         result = input.clone()
1213         stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
1214         ar_mask = (result != input).long()
1215
1216         # for n in range(result.size(0)):
1217         # logger(
1218         # f"test_before {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
1219         # )
1220
1221         masked_inplace_autoregression(
1222             model,
1223             self.batch_size,
1224             result,
1225             ar_mask,
1226             deterministic_synthesis,
1227             device=self.device,
1228         )
1229
1230         #!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
1231         for label, input in [
1232             ("train", self.train_input[:32]),
1233             ("test", self.test_input[:32]),
1234         ]:
1235             output = model(BracketedSequence(input)).x
1236             output = output.log_softmax(dim=-1)
1237             filename = os.path.join(
1238                 result_dir, f"stack_with_crossentropy_{n_epoch:04d}_{label}.txt"
1239             )
1240             with open(filename, "w") as f:
1241                 for n in range(input.size(0)):
1242                     s = stack.seq_to_str(
1243                         input[n], nb_stacks=self.nb_stacks, nb_digits=self.nb_digits
1244                     )
1245                     for t, k, w in zip(range(input[n].size(0)), input[n], s.split(" ")):
1246                         u = (
1247                             " " * (10 - len(w))
1248                             + w
1249                             + " "
1250                             + str(output[n][t][k].exp().item())
1251                             + "\n"
1252                         )
1253                         f.write(u)
1254                     f.write("\n")
1255             logger(f"wrote {filename}")
1256         #!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
1257
1258         for n in range(result.size(0)):
1259             logger(
1260                 f"test_after  {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
1261             )
1262         ##############################################################
1263
1264
1265 ######################################################################
1266
1267 import rpl
1268
1269
1270 class RPL(Task):
1271     def tensorize(self, sequences):
1272         len_max = max([len(x) for x in sequences])
1273         return torch.cat(
1274             [
1275                 torch.tensor(
1276                     [
1277                         [
1278                             self.token2id[str(c)]
1279                             for c in s + ["<nul>"] * (len_max - len(s))
1280                         ]
1281                         for s in sequences
1282                     ]
1283                 )
1284             ],
1285             0,
1286         )
1287
1288     def seq2str(self, seq):
1289         return " ".join([self.id2token[i] for i in seq])
1290
1291     def __init__(
1292         self,
1293         nb_train_samples,
1294         nb_test_samples,
1295         batch_size,
1296         nb_starting_values=3,
1297         max_input=9,
1298         prog_len=6,
1299         nb_runs=5,
1300         no_prog=False,
1301         logger=None,
1302         device=torch.device("cpu"),
1303     ):
1304         super().__init__()
1305
1306         self.batch_size = batch_size
1307         self.device = device
1308         self.no_prog = no_prog
1309
1310         train_sequences = [
1311             rpl.generate(
1312                 nb_starting_values=nb_starting_values,
1313                 nb_result_values_max=4 * nb_starting_values,
1314                 max_input=max_input,
1315                 prog_len=prog_len,
1316                 nb_runs=nb_runs,
1317             )
1318             for _ in tqdm.tqdm(range(nb_train_samples), desc="train-data")
1319         ]
1320
1321         test_sequences = [
1322             rpl.generate(
1323                 nb_starting_values=nb_starting_values,
1324                 nb_result_values_max=4 * nb_starting_values,
1325                 max_input=max_input,
1326                 prog_len=prog_len,
1327                 nb_runs=nb_runs,
1328             )
1329             for _ in tqdm.tqdm(range(nb_test_samples), desc="test-data")
1330         ]
1331
1332         symbols = list(
1333             set(["<nul>"] + [x for l in train_sequences + test_sequences for x in l])
1334         )
1335         val_max = max([x if type(x) is int else 0 for x in symbols])
1336         symbols = list(filter(lambda x: type(x) is str, symbols))
1337         symbols.sort()
1338         symbols += [str(n) for n in range(val_max + 1)]
1339         self.token2id = dict([(c, n) for n, c in enumerate(symbols)])
1340         self.id2token = dict([(n, c) for c, n in self.token2id.items()])
1341
1342         self.t_nul = self.token2id["<nul>"]
1343         self.t_input = self.token2id["<in>"]
1344         self.t_output = self.token2id["<out>"]
1345         self.t_prog = self.token2id["<prg>"]
1346         self.t_end = self.token2id["<end>"]
1347
1348         self.train_input = self.tensorize(train_sequences)
1349         self.test_input = self.tensorize(test_sequences)
1350
1351         if no_prog:
1352             # Excise the program from every train and test example
1353             k = torch.arange(self.train_input.size(1), device=self.train_input.device)[
1354                 None, :
1355             ]
1356             p = (
1357                 ((self.train_input == self.t_prog).long() * k)
1358                 .max(1, keepdim=True)
1359                 .values
1360             )
1361             self.train_input = (
1362                 self.train_input * (k <= p).long()
1363                 + self.t_end * (k == p + 1).long()
1364                 + self.t_nul * (k > p + 1).long()
1365             )
1366             k = torch.arange(self.test_input.size(1), device=self.test_input.device)[
1367                 None, :
1368             ]
1369             p = (
1370                 ((self.test_input == self.t_prog).long() * k)
1371                 .max(1, keepdim=True)
1372                 .values
1373             )
1374             self.test_input = (
1375                 self.test_input * (k <= p).long()
1376                 + self.t_end * (k == p + 1).long()
1377                 + self.t_nul * (k > p + 1).long()
1378             )
1379
1380         if logger is not None:
1381             logger(f"value_max {val_max}")
1382             for x in self.train_input[:25]:
1383                 end = (x != self.t_nul).nonzero().max().item() + 1
1384                 seq = [self.id2token[i.item()] for i in x[:end]]
1385                 s = " ".join(seq)
1386                 logger(f"example_seq {s}")
1387
1388         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1389
1390     def batches(self, split="train", nb_to_use=-1, desc=None):
1391         assert split in {"train", "test"}
1392         input = self.train_input if split == "train" else self.test_input
1393         if nb_to_use > 0:
1394             input = input[:nb_to_use]
1395         if desc is None:
1396             desc = f"epoch-{split}"
1397         for batch in tqdm.tqdm(
1398             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1399         ):
1400             last = (batch != self.t_nul).max(0).values.nonzero().max() + 3
1401             batch = batch[:, :last].to(self.device)
1402             yield batch
1403
1404     def vocabulary_size(self):
1405         return self.nb_codes
1406
1407     def produce_results(
1408         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1409     ):
1410         # --------------------------------------------------------------------
1411         def compute_nb_errors_prog(input, nb_to_log=0):
1412             result = input.clone()
1413             s = (result == self.t_prog).long()
1414             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1415             result = (1 - ar_mask) * result + ar_mask * self.t_nul
1416
1417             masked_inplace_autoregression(
1418                 model,
1419                 self.batch_size,
1420                 result,
1421                 ar_mask,
1422                 deterministic_synthesis,
1423                 device=self.device,
1424             )
1425
1426             sum_nb_total, sum_nb_errors = 0, 0
1427             for one_input, one_result in zip(input, result):
1428                 seq = [self.id2token[i.item()] for i in one_result]
1429                 nb_total, nb_errors, prog, stacks = rpl.compute_nb_errors(seq)
1430                 sum_nb_total += 1
1431                 sum_nb_errors += 0 if nb_errors == 0 else 1
1432                 if nb_to_log > 0:
1433                     gt_seq = [self.id2token[i.item()] for i in one_input]
1434                     _, _, gt_prog, _ = rpl.compute_nb_errors(gt_seq)
1435                     gt_prog = " ".join([str(x) for x in gt_prog])
1436                     prog = " ".join([str(x) for x in prog])
1437                     comment = "*" if nb_errors == 0 else "-"
1438                     logger(f"{comment} PROG [{gt_prog}] PREDICTED [{prog}]")
1439                     for start_stack, target_stack, result_stack, correct in stacks:
1440                         comment = "*" if correct else "-"
1441                         start_stack = " ".join([str(x) for x in start_stack])
1442                         target_stack = " ".join([str(x) for x in target_stack])
1443                         result_stack = " ".join([str(x) for x in result_stack])
1444                         logger(
1445                             f"  {comment} [{start_stack}] -> [{target_stack}] PREDICTED [{result_stack}]"
1446                         )
1447                     nb_to_log -= 1
1448
1449             return sum_nb_total, sum_nb_errors
1450
1451         # --------------------------------------------------------------------
1452         def compute_nb_errors_output(input, nb_to_log=0):
1453             result = input.clone()
1454             k = torch.arange(result.size(1), device=result.device)[None, :]
1455             last_output_idx = (
1456                 ((result == self.t_output) * k).max(dim=1, keepdim=True).values
1457             )
1458             first_prog_idx = (
1459                 ((result == self.t_prog) * k).max(dim=1, keepdim=True).values
1460             )
1461             ar_mask = (k > last_output_idx).long() * (k < first_prog_idx).long()
1462             result = (1 - ar_mask) * result + ar_mask * self.t_nul
1463
1464             masked_inplace_autoregression(
1465                 model,
1466                 self.batch_size,
1467                 result,
1468                 ar_mask,
1469                 deterministic_synthesis,
1470                 device=self.device,
1471             )
1472
1473             sum_nb_total, sum_nb_errors = 0, 0
1474             for one_input, one_result, i, j in zip(
1475                 input, result, last_output_idx, first_prog_idx
1476             ):
1477                 seq = [self.id2token[i.item()] for i in one_result]
1478                 sum_nb_total += 1
1479                 correct = (one_input - one_result).abs().max() == 0
1480                 sum_nb_errors += 0 if correct else 1
1481                 if nb_to_log > 0:
1482                     result_stack = [
1483                         self.id2token[i.item()] for i in one_result[i : j + 1]
1484                     ]
1485                     target_stack = [
1486                         self.id2token[i.item()] for i in one_input[i : j + 1]
1487                     ]
1488                     comment = "*" if correct else "-"
1489                     result_stack = " ".join([str(x) for x in result_stack])
1490                     target_stack = " ".join([str(x) for x in target_stack])
1491                     logger(
1492                         f"output_test {comment} [{target_stack}] PREDICTED [{result_stack}]"
1493                     )
1494                     nb_to_log -= 1
1495
1496             return sum_nb_total, sum_nb_errors
1497
1498         # --------------------------------------------------------------------
1499
1500         if not self.no_prog:
1501             test_nb_total, test_nb_errors = compute_nb_errors_prog(
1502                 self.test_input[:1000].to(self.device), nb_to_log=10
1503             )
1504
1505             logger(
1506                 f"accuracy_prog_test {n_epoch} nb_total {test_nb_total} nb_errors {test_nb_errors} accuracy {100.0*(1-test_nb_errors/test_nb_total):.02f}%"
1507             )
1508
1509             logger(f"main_test_accuracy {n_epoch} {1-test_nb_errors/test_nb_total}")
1510
1511         test_nb_total, test_nb_errors = compute_nb_errors_output(
1512             self.test_input[:1000].to(self.device), nb_to_log=10
1513         )
1514
1515         logger(
1516             f"accuracy_output_test {n_epoch} nb_total {test_nb_total} nb_errors {test_nb_errors} accuracy {100.0*(1-test_nb_errors/test_nb_total):.02f}%"
1517         )
1518
1519         if save_attention_image is None:
1520             logger("no save_attention_image (is pycairo installed?)")
1521         else:
1522             ns = torch.randint(self.test_input.size(0), (1,)).item()
1523             input = self.test_input[ns : ns + 1].clone()
1524             last = (input != self.t_nul).max(0).values.nonzero().max() + 3
1525             input = input[:, :last].to(self.device)
1526
1527             with torch.autograd.no_grad():
1528                 t = model.training
1529                 model.eval()
1530                 model.record_attention(True)
1531                 model(BracketedSequence(input))
1532                 model.train(t)
1533                 ram = model.retrieve_attention()
1534                 model.record_attention(False)
1535
1536             tokens_output = [self.id2token[i.item()] for i in input[0]]
1537             tokens_input = ["n/a"] + tokens_output[:-1]
1538             for n_head in range(ram[0].size(1)):
1539                 filename = os.path.join(
1540                     result_dir, f"rpl_attention_{n_epoch}_h{n_head}.pdf"
1541                 )
1542                 attention_matrices = [m[0, n_head] for m in ram]
1543                 save_attention_image(
1544                     filename,
1545                     tokens_input,
1546                     tokens_output,
1547                     attention_matrices,
1548                     k_top=10,
1549                     # min_total_attention=0.9,
1550                     token_gap=12,
1551                     layer_gap=50,
1552                 )
1553                 logger(f"wrote {filename}")
1554
1555
1556 ######################################################################
1557
1558
1559 import expr
1560
1561
1562 class Expr(Task):
1563     def tensorize(self, sequences):
1564         len_max = max([len(x) for x in sequences])
1565         return torch.cat(
1566             [
1567                 torch.tensor(
1568                     [
1569                         [self.char2id[c] for c in s + "#" * (len_max - len(s))]
1570                         for s in sequences
1571                     ]
1572                 )
1573             ],
1574             0,
1575         ).to(self.device)
1576
1577     def __init__(
1578         self,
1579         nb_train_samples,
1580         nb_test_samples,
1581         nb_variables,
1582         sequence_length,
1583         operand_max,
1584         result_max,
1585         batch_size,
1586         device=torch.device("cpu"),
1587     ):
1588         super().__init__()
1589
1590         self.batch_size = batch_size
1591         self.device = device
1592
1593         train_sequences = expr.generate_sequences(
1594             nb_train_samples,
1595             nb_variables=nb_variables,
1596             length=sequence_length,
1597             operand_max=operand_max,
1598             result_max=result_max,
1599         )
1600
1601         test_sequences = expr.generate_sequences(
1602             nb_test_samples,
1603             nb_variables=nb_variables,
1604             length=sequence_length,
1605             operand_max=operand_max,
1606             result_max=result_max,
1607         )
1608
1609         symbols = list(set("#" + "".join(train_sequences + test_sequences)))
1610         symbols.sort()
1611
1612         self.char2id = dict([(c, n) for n, c in enumerate(symbols)])
1613         self.id2char = dict([(n, c) for c, n in self.char2id.items()])
1614
1615         self.filler, self.space = self.char2id["#"], self.char2id[" "]
1616
1617         self.train_input = self.tensorize(train_sequences)
1618         self.test_input = self.tensorize(test_sequences)
1619
1620         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1621
1622     def batches(self, split="train", nb_to_use=-1, desc=None):
1623         assert split in {"train", "test"}
1624         input = self.train_input if split == "train" else self.test_input
1625         if nb_to_use > 0:
1626             input = input[:nb_to_use]
1627         if desc is None:
1628             desc = f"epoch-{split}"
1629         for batch in tqdm.tqdm(
1630             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1631         ):
1632             last = (batch != self.filler).max(0).values.nonzero().max() + 3
1633             batch = batch[:, :last]
1634             yield batch
1635
1636     def vocabulary_size(self):
1637         return self.nb_codes
1638
1639     def seq2str(self, s):
1640         return "".join([self.id2char[k.item()] for k in s])
1641
1642     def produce_results(
1643         self,
1644         n_epoch,
1645         model,
1646         result_dir,
1647         logger,
1648         deterministic_synthesis,
1649         input_file=None,
1650     ):
1651         def compute_nb_correct(input):
1652             result = input.clone()
1653             s = (result == self.space).long()
1654             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1655             result = (1 - ar_mask) * result + ar_mask * self.filler
1656             masked_inplace_autoregression(
1657                 model,
1658                 self.batch_size,
1659                 result,
1660                 ar_mask,
1661                 deterministic_synthesis,
1662                 device=self.device,
1663             )
1664
1665             nb_total = input.size(0)
1666             nb_correct = (input == result).long().min(1).values.sum()
1667
1668             #######################################################################
1669             # Comput predicted vs. true variable values
1670
1671             nb_delta = torch.zeros(5, dtype=torch.int64)
1672             nb_missed = 0
1673
1674             values_input = expr.extract_results([self.seq2str(s) for s in input])
1675             values_result = expr.extract_results([self.seq2str(s) for s in result])
1676
1677             filename = os.path.join(result_dir, f"expr_result_{n_epoch:04d}.txt")
1678
1679             with open(filename, "w") as f:
1680                 for i, r in zip(values_input, values_result):
1681                     for n, vi in i.items():
1682                         vr = r.get(n)
1683                         f.write(f"{vi} {-1 if vr is None else vr}\n")
1684
1685                         if vr is None or vr < 0:
1686                             nb_missed += 1
1687                         else:
1688                             d = abs(vr - vi)
1689                             if d >= nb_delta.size(0):
1690                                 nb_missed += 1
1691                             else:
1692                                 nb_delta[d] += 1
1693
1694             ######################################################################
1695
1696             return nb_total, nb_correct, nb_delta, nb_missed
1697
1698         (
1699             test_nb_total,
1700             test_nb_correct,
1701             test_nb_delta,
1702             test_nb_missed,
1703         ) = compute_nb_correct(self.test_input[:10000])
1704
1705         logger(
1706             f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
1707         )
1708
1709         logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
1710
1711         nb_total = test_nb_delta.sum() + test_nb_missed
1712         for d in range(test_nb_delta.size(0)):
1713             logger(
1714                 f"error_value {n_epoch} delta {d} {test_nb_delta[d]} {test_nb_delta[d]*100/nb_total:.02f}%"
1715             )
1716         logger(
1717             f"error_value {n_epoch} missed {test_nb_missed} {test_nb_missed*100/nb_total:.02f}%"
1718         )
1719
1720         ##############################################################
1721         # Log a few generated sequences
1722         if input_file is None:
1723             input = self.test_input[:10]
1724         else:
1725             with open(input_file, "r") as f:
1726                 sequences = [e.strip() for e in f.readlines()]
1727                 sequences = [s + " " + "#" * 50 for s in sequences]
1728                 input = self.tensorize(sequences)
1729
1730         result = input.clone()
1731         s = (result == self.space).long()
1732         ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1733         result = (1 - ar_mask) * result + ar_mask * self.filler
1734
1735         for n in range(result.size(0)):
1736             logger(f"test_before {self.seq2str(result[n])}")
1737
1738         masked_inplace_autoregression(
1739             model,
1740             self.batch_size,
1741             result,
1742             ar_mask,
1743             deterministic_synthesis,
1744             device=self.device,
1745         )
1746
1747         correct = (1 - ar_mask) * self.space + ar_mask * input
1748         for n in range(result.size(0)):
1749             comment = "GOOD" if (result[n] - input[n]).abs().max() == 0 else ""
1750             logger(f"test_after  {self.seq2str(result[n])} {comment}")
1751             logger(f"truth       {self.seq2str(correct[n])}")
1752         ##############################################################
1753
1754
1755 ######################################################################
1756
1757 import grid
1758
1759
1760 class Grid(Task):
1761     # Make a tensor from a list of strings
1762     def str2tensor(self, descr):
1763         token_descr = [s.strip().split(" ") for s in descr]
1764         l = max([len(s) for s in token_descr])
1765         token_descr = [s + ["#"] * (l - len(s)) for s in token_descr]
1766         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
1767         return torch.tensor(id_descr, device=self.device)
1768
1769     # Make a list of strings from a tensor
1770     def tensor2str(self, x):
1771         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
1772
1773     # trim all the tensors in the tuple z to remove as much token from
1774     # left and right in the first tensor. If z is a tuple, all its
1775     # elements are trimed according to the triming for the first
1776     def trim(self, z, token="#"):
1777         n = self.token2id[token]
1778         if type(z) == tuple:
1779             x = z[0]
1780             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
1781             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
1782             return tuple([t[:, a:b] for t in z])
1783         else:
1784             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
1785             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
1786             return z[:, a:b]
1787
1788     ######################
1789
1790     def __init__(
1791         self,
1792         nb_train_samples,
1793         nb_test_samples,
1794         batch_size,
1795         size,
1796         fraction_play=0.0,
1797         logger=None,
1798         device=torch.device("cpu"),
1799     ):
1800         super().__init__()
1801
1802         self.device = device
1803         self.batch_size = batch_size
1804         self.grid_factory = grid.GridFactory(size=size)
1805         self.fraction_play = fraction_play
1806
1807         if logger is not None:
1808             logger(
1809                 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
1810             )
1811
1812         self.train_descr = self.grid_factory.generate_samples(
1813             nb=nb_train_samples,
1814             fraction_play=fraction_play,
1815             progress_bar=lambda r: tqdm.tqdm(r),
1816         )
1817
1818         self.test_descr = self.grid_factory.generate_samples(
1819             nb=nb_test_samples, fraction_play=0.0, progress_bar=lambda r: tqdm.tqdm(r)
1820         )
1821
1822         if fraction_play > 0:
1823             self.play_descr = self.grid_factory.generate_samples(
1824                 nb=25, fraction_play=1.0, progress_bar=lambda r: tqdm.tqdm(r)
1825             )
1826         else:
1827             self.play_descr = []
1828
1829         # Build the tokenizer
1830         tokens = set()
1831         for d in [self.train_descr, self.test_descr, self.play_descr]:
1832             for s in d:
1833                 for t in s.strip().split(" "):
1834                     tokens.add(t)
1835         # make this set a sorted list to get the same tensors given
1836         # the same descr
1837         tokens = list(tokens)
1838         tokens.sort()
1839         tokens = ["#"] + tokens
1840         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
1841         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
1842         self.t_nul = self.token2id["#"]
1843         self.t_true = self.token2id["true"]
1844         self.t_false = self.token2id["false"]
1845         # self.t_pipe = self.token2id["|"]
1846
1847         # Tokenize the train and test sets
1848         self.train_input = self.str2tensor(self.train_descr)
1849         self.test_input = self.str2tensor(self.test_descr)
1850         self.play_input = (
1851             None if len(self.play_descr) == 0 else self.str2tensor(self.play_descr)
1852         )
1853
1854     def batches(self, split="train", nb_to_use=-1, desc=None):
1855         assert split in {"train", "test"}
1856         input = self.train_input if split == "train" else self.test_input
1857         for batch in tqdm.tqdm(
1858             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
1859         ):
1860             yield self.trim(batch)
1861
1862     def vocabulary_size(self):
1863         return len(self.token2id)
1864
1865     def produce_results(
1866         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1867     ):
1868         correct = self.test_input[:1000]
1869         result = correct.clone()
1870         ar_mask = torch.logical_or(result == self.t_true, result == self.t_false).long()
1871         result *= 1 - ar_mask  # paraaaaanoiaaaaaaa
1872
1873         logger(f"----------------------------------------------------------")
1874
1875         for e in self.tensor2str(result[:10]):
1876             logger(f"test_before {e}")
1877
1878         masked_inplace_autoregression(
1879             model,
1880             self.batch_size,
1881             result,
1882             ar_mask,
1883             deterministic_synthesis,
1884             device=self.device,
1885         )
1886
1887         logger(f"----------------------------------------------------------")
1888
1889         for e in self.tensor2str(result[:10]):
1890             logger(f"test_after  {e}")
1891
1892         logger(f"----------------------------------------------------------")
1893
1894         nb_total = ar_mask.sum().item()
1895         nb_correct = ((correct == result).long() * ar_mask).sum().item()
1896
1897         logger(f"test_performance {n_epoch} {nb_total=} {nb_correct=}")
1898         logger(f"main_test_accuracy {n_epoch} {nb_correct / nb_total}")
1899
1900         if self.play_input is not None:
1901             result = self.play_input.clone()
1902             ar_mask = (result == self.t_pipe).long().cumsum(dim=1).clamp(max=1)
1903             result *= 1 - ar_mask  # paraaaaanoiaaaaaaa
1904
1905             logger(f"----------------------------------------------------------")
1906
1907             for e in self.tensor2str(result[:10]):
1908                 logger(f"play_before {e}")
1909
1910             masked_inplace_autoregression(
1911                 model,
1912                 self.batch_size,
1913                 result,
1914                 ar_mask,
1915                 deterministic_synthesis,
1916                 device=self.device,
1917             )
1918
1919             logger(f"----------------------------------------------------------")
1920
1921             for e in self.tensor2str(result[:10]):
1922                 logger(f"play_after  {e}")
1923
1924             logger(f"----------------------------------------------------------")
1925
1926
1927 ######################################################################
1928
1929 import qmlp
1930
1931
1932 class QMLP(Task):
1933     ######################
1934
1935     def __init__(
1936         self,
1937         nb_train_samples,
1938         nb_test_samples,
1939         batch_size,
1940         result_dir,
1941         logger=None,
1942         device=torch.device("cpu"),
1943     ):
1944         super().__init__()
1945
1946         self.device = device
1947         self.batch_size = batch_size
1948         self.nb_samples_per_mlp = 256
1949
1950         if logger is not None:
1951             logger(
1952                 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
1953             )
1954
1955         seq, q_test_set, test_error = qmlp.generate_sequence_and_test_set(
1956             nb_mlps=nb_train_samples + nb_test_samples,
1957             nb_samples=self.nb_samples_per_mlp,
1958             device=self.device,
1959             batch_size=64,
1960             nb_epochs=250,
1961             nb_mlps_per_batch=1024,
1962         )
1963
1964         self.train_input = seq[:nb_train_samples]
1965         self.train_q_test_set = q_test_set[:nb_train_samples]
1966         self.train_ref_test_errors = test_error[:nb_train_samples]
1967         self.test_input = seq[nb_train_samples:]
1968         self.test_q_test_set = q_test_set[nb_train_samples:]
1969         self.test_ref_test_errors = test_error[nb_train_samples:]
1970
1971         filename = os.path.join(result_dir, f"train_errors_ref.dat")
1972         with open(filename, "w") as f:
1973             for e in self.train_ref_test_errors:
1974                 f.write(f"{e}\n")
1975
1976         filename = os.path.join(result_dir, f"test_errors_ref.dat")
1977         with open(filename, "w") as f:
1978             for e in self.test_ref_test_errors:
1979                 f.write(f"{e}\n")
1980
1981         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1982
1983     def batches(self, split="train", nb_to_use=-1, desc=None):
1984         assert split in {"train", "test"}
1985         input = self.train_input if split == "train" else self.test_input
1986         for batch in tqdm.tqdm(
1987             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
1988         ):
1989             yield batch
1990
1991     def vocabulary_size(self):
1992         return self.nb_codes
1993
1994     def produce_results(
1995         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1996     ):
1997         correct = self.test_input[:1000]
1998         result = correct.clone()
1999         ar_mask = (
2000             torch.arange(result.size(1), device=result.device)
2001             > self.nb_samples_per_mlp * 3 + 1
2002         ).long()[None, :]
2003         ar_mask = ar_mask.expand_as(result)
2004         result *= 1 - ar_mask  # paraaaaanoiaaaaaaa
2005
2006         masked_inplace_autoregression(
2007             model,
2008             self.batch_size,
2009             result,
2010             ar_mask,
2011             deterministic_synthesis,
2012             device=self.device,
2013         )
2014
2015         q_train_set = result[:, : self.nb_samples_per_mlp * 3]
2016         q_params = result[:, self.nb_samples_per_mlp * 3 + 1 :]
2017         error_test = qmlp.evaluate_q_params(q_params, self.test_q_test_set)
2018
2019         filename = os.path.join(result_dir, f"test_errors_{n_epoch:04d}.dat")
2020         with open(filename, "w") as f:
2021             for e in error_test:
2022                 f.write(f"{e}\n")
2023
2024
2025 ######################################################################
2026
2027 import greed
2028
2029
2030 class Greed(Task):
2031     def __init__(
2032         self,
2033         nb_train_samples,
2034         nb_test_samples,
2035         batch_size,
2036         height,
2037         width,
2038         T,
2039         nb_walls,
2040         nb_coins,
2041         logger=None,
2042         device=torch.device("cpu"),
2043     ):
2044         super().__init__()
2045
2046         self.batch_size = batch_size
2047         self.device = device
2048
2049         self.world = greed.GreedWorld(height, width, T, nb_walls, nb_coins)
2050
2051         states, actions, rewards = self.world.generate_episodes(
2052             nb_train_samples + nb_test_samples
2053         )
2054         seq = self.world.episodes2seq(states, actions, rewards)
2055         self.train_input = seq[:nb_train_samples].to(self.device)
2056         self.test_input = seq[nb_train_samples:].to(self.device)
2057
2058     def wipe_lookahead_rewards(self, batch):
2059         t = torch.arange(batch.size(1), device=batch.device)[None, :]
2060         u = torch.randint(batch.size(1), (batch.size(0), 1), device=batch.device)
2061         lr_mask = (t <= u).long() * (
2062             t % self.world.it_len == self.world.index_lookahead_reward
2063         ).long()
2064
2065         return (
2066             lr_mask * self.world.lookahead_reward2code(greed.REWARD_UNKNOWN)
2067             + (1 - lr_mask) * batch
2068         )
2069
2070     def batches(self, split="train", nb_to_use=-1, desc=None):
2071         assert split in {"train", "test"}
2072         input = self.train_input if split == "train" else self.test_input
2073         if nb_to_use > 0:
2074             input = input[:nb_to_use]
2075         if desc is None:
2076             desc = f"epoch-{split}"
2077         for batch in tqdm.tqdm(
2078             input.split(self.batch_size), dynamic_ncols=True, desc=desc
2079         ):
2080             yield self.wipe_lookahead_rewards(batch)
2081
2082     def vocabulary_size(self):
2083         return self.world.nb_codes
2084
2085     def thinking_autoregression(
2086         self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
2087     ):
2088         snapshots = []
2089
2090         def ar(result, ar_mask, logit_biases=None):
2091             ar_mask = ar_mask.expand_as(result)
2092             result *= 1 - ar_mask
2093             masked_inplace_autoregression(
2094                 model,
2095                 self.batch_size,
2096                 result,
2097                 ar_mask,
2098                 deterministic_synthesis=deterministic_synthesis,
2099                 logit_biases=logit_biases,
2100                 device=self.device,
2101                 progress_bar_desc=None,
2102             )
2103             warnings.warn("keeping thinking snapshots", RuntimeWarning)
2104             snapshots.append(result[:100].detach().clone())
2105
2106         # Generate iteration after iteration
2107
2108         result = self.test_input[:250].clone()
2109         # Erase all the content but that of the first iteration
2110         result[:, self.world.it_len :] = -1
2111         # Set the lookahead_reward of the firs to UNKNOWN
2112         result[:, self.world.index_lookahead_reward] = self.world.lookahead_reward2code(
2113             greed.REWARD_UNKNOWN
2114         )
2115
2116         t = torch.arange(result.size(1), device=result.device)[None, :]
2117
2118         for u in tqdm.tqdm(
2119             range(0, result.size(1), self.world.it_len),
2120             desc="thinking",
2121         ):
2122             # Generate the next state but keep the initial one, the
2123             # lookahead_reward of previous iterations are set to
2124             # UNKNOWN
2125             if u > 0:
2126                 result[
2127                     :, u + self.world.index_lookahead_reward
2128                 ] = self.world.lookahead_reward2code(greed.REWARD_UNKNOWN)
2129                 ar_mask = (t >= u + self.world.index_states).long() * (
2130                     t < u + self.world.index_states + self.world.state_len
2131                 ).long()
2132                 ar(result, ar_mask)
2133
2134             # Generate the action and reward with lookahead_reward to +1
2135             result[
2136                 :, u + self.world.index_lookahead_reward
2137             ] = self.world.lookahead_reward2code(greed.REWARD_PLUS)
2138             ar_mask = (t >= u + self.world.index_reward).long() * (
2139                 t <= u + self.world.index_action
2140             ).long()
2141             ar(result, ar_mask)
2142
2143             # Set the lookahead_reward to UNKNOWN for the next iterations
2144             result[
2145                 :, u + self.world.index_lookahead_reward
2146             ] = self.world.lookahead_reward2code(greed.REWARD_UNKNOWN)
2147
2148         filename = os.path.join(result_dir, f"test_thinking_compute_{n_epoch:04d}.txt")
2149         with open(filename, "w") as f:
2150             for n in range(snapshots[0].size(0)):
2151                 for s in snapshots:
2152                     lr, s, a, r = self.world.seq2episodes(
2153                         s[n : n + 1],
2154                     )
2155                     str = self.world.episodes2str(
2156                         lr, s, a, r, unicode=True, ansi_colors=True
2157                     )
2158                     f.write(str)
2159                 f.write("\n\n")
2160
2161         # Saving the generated sequences
2162
2163         lr, s, a, r = self.world.seq2episodes(result)
2164         str = self.world.episodes2str(lr, s, a, r, unicode=True, ansi_colors=True)
2165
2166         filename = os.path.join(result_dir, f"test_thinking_seq_{n_epoch:04d}.txt")
2167         with open(filename, "w") as f:
2168             f.write(str)
2169             logger(f"wrote {filename}")
2170
2171     def produce_results(
2172         self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
2173     ):
2174         result = self.wipe_lookahead_rewards(self.test_input[:250].clone())
2175
2176         # Saving the ground truth
2177
2178         lr, s, a, r = self.world.seq2episodes(
2179             result,
2180         )
2181         str = self.world.episodes2str(lr, s, a, r, unicode=True, ansi_colors=True)
2182
2183         filename = os.path.join(result_dir, f"test_true_seq_{n_epoch:04d}.txt")
2184         with open(filename, "w") as f:
2185             f.write(str)
2186             logger(f"wrote {filename}")
2187
2188         # Re-generating from the first frame
2189
2190         ar_mask = (
2191             torch.arange(result.size(1), device=result.device) >= self.world.it_len
2192         ).long()[None, :]
2193         ar_mask = ar_mask.expand_as(result)
2194         result *= 1 - ar_mask  # paraaaaanoiaaaaaaa
2195
2196         masked_inplace_autoregression(
2197             model,
2198             self.batch_size,
2199             result,
2200             ar_mask,
2201             deterministic_synthesis,
2202             device=self.device,
2203         )
2204
2205         # Saving the generated sequences
2206
2207         lr, s, a, r = self.world.seq2episodes(
2208             result,
2209         )
2210         str = self.world.episodes2str(lr, s, a, r, unicode=True, ansi_colors=True)
2211
2212         filename = os.path.join(result_dir, f"test_seq_{n_epoch:04d}.txt")
2213         with open(filename, "w") as f:
2214             f.write(str)
2215             logger(f"wrote {filename}")
2216
2217         self.thinking_autoregression(
2218             n_epoch, model, result_dir, logger, deterministic_synthesis, nmax
2219         )
2220
2221
2222 ######################################################################