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