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) == {"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         print("SANITY", symbols)
 140         self.char2id = dict([(c, n) for n, c in enumerate(symbols)])
 141         self.id2char = dict([(n, c) for c, n in self.char2id.items()])
 142
 143         print(self.char2id)
 144
 145         self.train_input, self.train_pred_masks = self.tensorize(
 146             pairs[:nb_train_samples]
 147         )
 148         self.test_input, self.test_pred_masks = self.tensorize(pairs[nb_train_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         print(f"{type(t)=}")
 167         return ["".join([self.id2char[x.item()] for x in s]) for s in t]
 168
 169     def produce_results(
 170         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 171     ):
 172         correct = self.trim(self.test_input[:1000]).to(self.device)
 173         result = correct.clone()
 174         pred_mask = self.test_pred_masks[:1000, : result.size(1)].to(self.device)
 175         ar_mask = (pred_mask > 0).long()
 176         result *= 1 - ar_mask  # paraaaaanoiaaaaaaa
 177
 178         logger(f"----------------------------------------------------------")
 179
 180         for e in self.tensor2str(result[:10]):
 181             logger(f"test_before {e}")
 182
 183         masked_inplace_autoregression(
 184             model,
 185             self.batch_size,
 186             result,
 187             ar_mask,
 188             deterministic_synthesis,
 189             device=self.device,
 190         )
 191
 192         logger(f"----------------------------------------------------------")
 193
 194         for e, c in zip(self.tensor2str(result[:10]), self.tensor2str(correct[:10])):
 195             logger(f"test_after  {e}")
 196             logger(f"correct     {c}")
 197
 198         logger(f"----------------------------------------------------------")
 199
 200         err_mask = (pred_mask == 2).long()
 201         nb_total = err_mask.sum().item()
 202         nb_correct = ((correct == result).long() * err_mask).sum().item()
 203
 204         logger(f"test_performance {n_epoch} {nb_total=} {nb_correct=}")
 205         logger(f"main_test_accuracy {n_epoch} {nb_correct / nb_total}")
 206
 207
 208 ####################
 209
 210 import problems
 211
 212
 213 class SandBox(Task):
 214     def __init__(
 215         self,
 216         problem,
 217         nb_train_samples,
 218         nb_test_samples,
 219         batch_size,
 220         logger=None,
 221         device=torch.device("cpu"),
 222         max_nb_codes=1024,
 223     ):
 224         super().__init__()
 225
 226         self.batch_size = batch_size
 227         self.device = device
 228         self.problem = problem
 229
 230         self.train_input, self.train_ar_mask = self.problem.generate_sequences(
 231             nb_train_samples
 232         )
 233         self.test_input, self.test_ar_mask = self.problem.generate_sequences(
 234             nb_test_samples
 235         )
 236
 237         self.train_input, self.train_ar_mask = self.train_input.to(
 238             device
 239         ), self.train_ar_mask.to(device)
 240         self.test_input, self.test_ar_mask = self.test_input.to(
 241             device
 242         ), self.test_ar_mask.to(device)
 243
 244         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 245
 246         # A bit of paranoia never hurts
 247         assert self.nb_codes <= max_nb_codes
 248         assert self.train_input.min() >= 0
 249         assert self.test_input.min() >= 0
 250         assert tuple(x.item() for x in self.train_ar_mask.unique()) in {
 251             (0,),
 252             (1,),
 253             (0, 1),
 254         }
 255         assert tuple(x.item() for x in self.test_ar_mask.unique()) in {
 256             (0,),
 257             (1,),
 258             (0, 1),
 259         }
 260
 261         if logger is not None:
 262             for s, a in zip(self.train_input[:100], self.train_ar_mask[:100]):
 263                 logger(f"train_sequences {self.problem.seq2str(s)}")
 264                 a = "".join(["01"[x.item()] for x in a])
 265                 logger(f"                {a}")
 266
 267     def batches(self, split="train", nb_to_use=-1, desc=None):
 268         assert split in {"train", "test"}
 269         input = self.train_input if split == "train" else self.test_input
 270         if nb_to_use > 0:
 271             input = input[:nb_to_use]
 272         if desc is None:
 273             desc = f"epoch-{split}"
 274         for batch in tqdm.tqdm(
 275             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 276         ):
 277             yield batch
 278
 279     def vocabulary_size(self):
 280         return self.nb_codes
 281
 282     def produce_results(
 283         self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
 284     ):
 285         def compute_accuracy(input, ar_mask, logger=None):
 286             input, ar_mask = input[:nmax], ar_mask[:nmax]
 287             result = input.clone() * (1 - ar_mask)
 288
 289             masked_inplace_autoregression(
 290                 model,
 291                 self.batch_size,
 292                 result,
 293                 ar_mask,
 294                 deterministic_synthesis,
 295                 progress_bar_desc=None,
 296                 device=self.device,
 297             )
 298
 299             log_ground_truth = ar_mask.min() == 0
 300
 301             if logger is not None:
 302                 for sp, st in zip(result[:10], input[:10]):
 303                     logger(
 304                         f"test_sequences {n_epoch} prediction   {self.problem.seq2str(sp)}"
 305                     )
 306                     if log_ground_truth:
 307                         logger(
 308                             f"               {n_epoch} ground truth {self.problem.seq2str(st)}"
 309                         )
 310
 311             nb_total, nb_correct = self.problem.compute_nb_correct(
 312                 input, ar_mask, result
 313             )
 314
 315             # nb_total = ar_mask.sum().item()
 316             # nb_correct = ((result == input).long() * ar_mask).sum().item()
 317
 318             return nb_total, nb_correct
 319
 320         train_nb_total, train_nb_correct = compute_accuracy(
 321             self.train_input, self.train_ar_mask
 322         )
 323
 324         logger(
 325             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}%"
 326         )
 327
 328         test_nb_total, test_nb_correct = compute_accuracy(
 329             self.test_input, self.test_ar_mask, logger
 330         )
 331
 332         logger(
 333             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}%"
 334         )
 335
 336         logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
 337
 338         if save_attention_image is not None:
 339             for k in range(10):
 340                 ns = torch.randint(self.test_input.size(0), (1,)).item()
 341                 input = self.test_input[ns : ns + 1].clone()
 342
 343                 with torch.autograd.no_grad():
 344                     t = model.training
 345                     model.eval()
 346                     # model.record_attention(True)
 347                     model(BracketedSequence(input))
 348                     model.train(t)
 349                     # ram = model.retrieve_attention()
 350                     # model.record_attention(False)
 351
 352                 # tokens_output = [c for c in self.problem.seq2str(input[0])]
 353                 # tokens_input = ["n/a"] + tokens_output[:-1]
 354                 # for n_head in range(ram[0].size(1)):
 355                 # filename = os.path.join(
 356                 # result_dir, f"sandbox_attention_{k}_h{n_head}.pdf"
 357                 # )
 358                 # attention_matrices = [m[0, n_head] for m in ram]
 359                 # save_attention_image(
 360                 # filename,
 361                 # tokens_input,
 362                 # tokens_output,
 363                 # attention_matrices,
 364                 # k_top=10,
 365                 ##min_total_attention=0.9,
 366                 # token_gap=12,
 367                 # layer_gap=50,
 368                 # )
 369                 # logger(f"wrote {filename}")
 370
 371
 372 ######################################################################
 373
 374 import picoclvr
 375
 376
 377 class PicoCLVR(Task):
 378     # Make a tensor from a list of strings
 379     def tensorize(self, descr):
 380         token_descr = [s.strip().split(" ") for s in descr]
 381         l = max([len(s) for s in token_descr])
 382         token_descr = [s + ["<nul>"] * (l - len(s)) for s in token_descr]
 383         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
 384         return torch.tensor(id_descr, device=self.device)
 385
 386     # Make a list of strings from a tensor
 387     def detensorize(self, x):
 388         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
 389
 390     # trim all the tensors in the tuple z to remove as much token from
 391     # left and right in the first tensor. If z is a tuple, all its
 392     # elements are trimed according to the triming for the first
 393     def trim(self, z, token="<nul>"):
 394         n = self.token2id[token]
 395         if type(z) == tuple:
 396             x = z[0]
 397             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 398             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 399             return tuple([t[:, a:b] for t in z])
 400         else:
 401             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
 402             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
 403             return z[:, a:b]
 404
 405     ######################
 406
 407     def __init__(
 408         self,
 409         nb_train_samples,
 410         nb_test_samples,
 411         batch_size,
 412         height,
 413         width,
 414         nb_colors=5,
 415         logger=None,
 416         device=torch.device("cpu"),
 417         pruner_train=None,
 418         pruner_eval=None,
 419     ):
 420         super().__init__()
 421
 422         def generate_descr(nb, cache_suffix, pruner):
 423             return picoclvr.generate(
 424                 nb,
 425                 height=self.height,
 426                 width=self.width,
 427                 nb_colors=nb_colors,
 428                 pruner=pruner,
 429             )
 430
 431         self.height = height
 432         self.width = width
 433         self.batch_size = batch_size
 434         self.device = device
 435         self.pruner_train = pruner_train
 436         self.pruner_eval = pruner_eval
 437
 438         if logger is not None:
 439             logger(
 440                 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
 441             )
 442
 443         self.train_descr = generate_descr(
 444             nb_train_samples, "train", pruner=self.pruner_train
 445         )
 446         self.test_descr = generate_descr(nb_test_samples, "test", pruner=None)
 447
 448         # Build the tokenizer
 449         tokens = {"<nul>", "<img>"}
 450         for d in [self.train_descr, self.test_descr]:
 451             for s in d:
 452                 for t in s.strip().split(" "):
 453                     tokens.add(t)
 454         # make this set a sorted list to get the same tensors given
 455         # the same descr
 456         tokens = list(tokens)
 457         tokens.sort()
 458         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
 459         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
 460         self.t_img, self.t_nul = self.token2id["<img>"], self.token2id["<nul>"]
 461
 462         # Tokenize the train and test sets
 463         self.train_input = self.tensorize(self.train_descr)
 464         self.test_input = self.tensorize(self.test_descr)
 465
 466     def batches(self, split="train"):
 467         assert split in {"train", "test"}
 468         input = self.train_input if split == "train" else self.test_input
 469         for batch in tqdm.tqdm(
 470             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
 471         ):
 472             yield self.trim(batch)
 473
 474     def vocabulary_size(self):
 475         return len(self.token2id)
 476
 477     def compute_missing_properties(
 478         self, n_epoch, model, logger, deterministic_synthesis, pruner=None
 479     ):
 480         acc_nb_requested_properties = []
 481         acc_nb_missing_properties = []
 482         acc_nb_results = 0
 483
 484         for input in tqdm.tqdm(
 485             self.test_input.split(self.batch_size),
 486             dynamic_ncols=True,
 487             desc=f"test-properties",
 488         ):
 489             result = input.clone()
 490             ar_mask = (result == self.t_img).long().cumsum(dim=1).clamp(max=1)
 491             result = (1 - ar_mask) * result + ar_mask * self.t_nul
 492             masked_inplace_autoregression(
 493                 model,
 494                 self.batch_size,
 495                 result,
 496                 ar_mask,
 497                 deterministic_synthesis,
 498                 progress_bar_desc=None,
 499                 device=self.device,
 500             )
 501
 502             result_descr = self.detensorize(result)
 503             np = picoclvr.nb_properties(
 504                 result_descr,
 505                 height=self.height,
 506                 width=self.width,
 507                 pruner=pruner,
 508             )
 509             nb_requested_properties, _, nb_missing_properties = zip(*np)
 510             acc_nb_requested_properties += nb_requested_properties
 511             acc_nb_missing_properties += nb_missing_properties
 512             acc_nb_results += len(result_descr)
 513
 514         nb_requested_properties = sum(acc_nb_requested_properties)
 515         nb_missing_properties = sum(acc_nb_missing_properties)
 516
 517         prefix = "" if pruner is None else "pruned_"
 518         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 519         logger(
 520             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 521         )
 522         logger(
 523             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 524         )
 525
 526         logger(
 527             f"main_test_accuracy {n_epoch} {1-nb_missing_properties/nb_requested_properties}"
 528         )
 529
 530     ######################################################################
 531
 532     def produce_results(
 533         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 534     ):
 535         self.compute_missing_properties(n_epoch, model, logger, deterministic_synthesis)
 536
 537         if self.pruner_eval is not None:
 538             self.compute_missing_properties(n_epoch, model, self.pruner_eval)
 539
 540         nb_tokens_to_generate = self.height * self.width + 3
 541         result_descr = []
 542         nb_per_primer = 8
 543         primer = []
 544
 545         for primer_descr in [
 546             "red above green <sep> green top <sep> blue right of red",
 547             "there is red <sep> there is yellow <sep> there is blue",
 548             "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
 549             "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
 550         ]:
 551             primer += [primer_descr + " <img>"] * nb_per_primer
 552
 553         result = self.tensorize(primer)
 554         fill = result.new_full(
 555             result.size()[:-1] + (self.height * self.width + 1,), self.t_nul
 556         )
 557         result = torch.cat((result, fill), 1)
 558         ar_mask = (result == self.t_nul).long()
 559         masked_inplace_autoregression(
 560             model,
 561             self.batch_size,
 562             result,
 563             ar_mask,
 564             deterministic_synthesis,
 565             device=self.device,
 566         )
 567         result_descr = self.detensorize(result)
 568
 569         np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
 570
 571         acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np)
 572         acc_nb_results = len(result_descr)
 573
 574         nb_requested_properties = sum(acc_nb_requested_properties)
 575         nb_missing_properties = sum(acc_nb_missing_properties)
 576
 577         prefix = "demo_"
 578         logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}")
 579         logger(
 580             f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}"
 581         )
 582         logger(
 583             f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%"
 584         )
 585
 586         img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
 587
 588         if img.dim() == 5:
 589             if img.size(1) == 1:
 590                 img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64)
 591             else:
 592                 img = torch.cat(
 593                     [
 594                         torchvision.utils.make_grid(x, padding=1, pad_value=64)[None]
 595                         for x in img
 596                     ],
 597                     0,
 598                 )
 599
 600         image_name = os.path.join(result_dir, f"picoclvr_result_{n_epoch:04d}.png")
 601         torchvision.utils.save_image(
 602             img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0
 603         )
 604         logger(f"wrote {image_name}")
 605
 606
 607 ######################################################################
 608
 609
 610 class MNIST(Task):
 611     def __init__(
 612         self, nb_train_samples, nb_test_samples, batch_size, device=torch.device("cpu")
 613     ):
 614         super().__init__()
 615
 616         self.nb_train_samples = (nb_train_samples,)
 617         self.nb_test_samples = (nb_test_samples,)
 618         self.batch_size = batch_size
 619         self.device = device
 620         data_set = torchvision.datasets.MNIST(root="./data", train=True, download=True)
 621         self.train_input = data_set.data[:nb_train_samples].view(-1, 28 * 28).long()
 622         data_set = torchvision.datasets.MNIST(root="./data", train=False, download=True)
 623         self.test_input = data_set.data[:nb_test_samples].view(-1, 28 * 28).long()
 624
 625     def batches(self, split="train", nb_to_use=-1, desc=None):
 626         assert split in {"train", "test"}
 627         input = self.train_input if split == "train" else self.test_input
 628         if nb_to_use > 0:
 629             input = input[:nb_to_use]
 630         if desc is None:
 631             desc = f"epoch-{split}"
 632         for batch in tqdm.tqdm(
 633             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 634         ):
 635             yield batch
 636
 637     def vocabulary_size(self):
 638         return 256
 639
 640     def produce_results(
 641         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 642     ):
 643         results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
 644         ar_mask = torch.full_like(results, 1)
 645         masked_inplace_autoregression(
 646             model,
 647             self.batch_size,
 648             results,
 649             ar_mask,
 650             deterministic_synthesis,
 651             device=self.device,
 652         )
 653         image_name = os.path.join(result_dir, f"mnist_result_{n_epoch:04d}.png")
 654         torchvision.utils.save_image(
 655             1 - results.reshape(-1, 1, 28, 28) / 255.0,
 656             image_name,
 657             nrow=16,
 658             pad_value=0.8,
 659         )
 660         logger(f"wrote {image_name}")
 661
 662
 663 ######################################################################
 664
 665 import maze
 666
 667
 668 class Maze(Task):
 669     def map2seq(self, *m):
 670         return torch.cat([x.flatten(1) for x in m], 1)
 671
 672     def seq2map(self, s):
 673         s = s.reshape(s.size(0), -1, self.height, self.width)
 674         return (s[:, k] for k in range(s.size(1)))
 675
 676     def __init__(
 677         self,
 678         nb_train_samples,
 679         nb_test_samples,
 680         batch_size,
 681         height,
 682         width,
 683         nb_walls,
 684         device=torch.device("cpu"),
 685     ):
 686         super().__init__()
 687
 688         self.batch_size = batch_size
 689         self.height = height
 690         self.width = width
 691         self.device = device
 692
 693         train_mazes, train_paths, _ = maze.create_maze_data(
 694             nb_train_samples,
 695             height=height,
 696             width=width,
 697             nb_walls=nb_walls,
 698             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 699         )
 700         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 701
 702         test_mazes, test_paths, _ = maze.create_maze_data(
 703             nb_test_samples,
 704             height=height,
 705             width=width,
 706             nb_walls=nb_walls,
 707             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 708         )
 709         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 710
 711         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 712
 713     def batches(self, split="train", nb_to_use=-1, desc=None):
 714         assert split in {"train", "test"}
 715         input = self.train_input if split == "train" else self.test_input
 716         if nb_to_use > 0:
 717             input = input[:nb_to_use]
 718         if desc is None:
 719             desc = f"epoch-{split}"
 720         for batch in tqdm.tqdm(
 721             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 722         ):
 723             yield batch
 724
 725     def vocabulary_size(self):
 726         return self.nb_codes
 727
 728     def compute_error(
 729         self, model, split="train", nb_to_use=-1, deterministic_synthesis=False
 730     ):
 731         nb_total, nb_correct = 0, 0
 732         count = torch.zeros(
 733             self.width * self.height,
 734             self.width * self.height,
 735             device=self.device,
 736             dtype=torch.int64,
 737         )
 738
 739         for input in self.batches(split, nb_to_use):
 740             result = input.clone()
 741             ar_mask = result.new_zeros(result.size())
 742             ar_mask[:, self.height * self.width :] = 1
 743             result *= 1 - ar_mask
 744             masked_inplace_autoregression(
 745                 model,
 746                 self.batch_size,
 747                 result,
 748                 ar_mask,
 749                 deterministic_synthesis,
 750                 progress_bar_desc=None,
 751                 device=self.device,
 752             )
 753             mazes, paths = self.seq2map(result)
 754             path_correctness = maze.path_correctness(mazes, paths)
 755             nb_correct += path_correctness.long().sum()
 756             nb_total += mazes.size(0)
 757
 758             optimal_path_lengths = (
 759                 (input[:, self.height * self.width :] == maze.v_path).long().sum(1)
 760             )
 761             predicted_path_lengths = (
 762                 (result[:, self.height * self.width :] == maze.v_path).long().sum(1)
 763             )
 764             optimal_path_lengths = optimal_path_lengths[path_correctness]
 765             predicted_path_lengths = predicted_path_lengths[path_correctness]
 766             count[optimal_path_lengths, predicted_path_lengths] += 1
 767
 768         if count.max() == 0:
 769             count = None
 770         else:
 771             count = count[
 772                 : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1
 773             ]
 774
 775         return nb_total, nb_correct, count
 776
 777     def produce_results(
 778         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 779     ):
 780         train_nb_total, train_nb_correct, count = self.compute_error(
 781             model,
 782             "train",
 783             nb_to_use=1000,
 784             deterministic_synthesis=deterministic_synthesis,
 785         )
 786         logger(
 787             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}%"
 788         )
 789
 790         test_nb_total, test_nb_correct, count = self.compute_error(
 791             model,
 792             "test",
 793             nb_to_use=1000,
 794             deterministic_synthesis=deterministic_synthesis,
 795         )
 796         logger(
 797             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}%"
 798         )
 799
 800         logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
 801
 802         if count is not None:
 803             proportion_optimal = count.diagonal().sum().float() / count.sum()
 804             logger(f"proportion_optimal_test {proportion_optimal*100:.02f}%")
 805             with open(
 806                 os.path.join(result_dir, f"maze_result_{n_epoch:04d}.txt"), "w"
 807             ) as f:
 808                 for i in range(count.size(0)):
 809                     for j in range(count.size(1)):
 810                         eol = " " if j < count.size(1) - 1 else "\n"
 811                         f.write(f"{count[i,j]}{eol}")
 812
 813         input = self.test_input[:48]
 814         result = input.clone()
 815         ar_mask = result.new_zeros(result.size())
 816         ar_mask[:, self.height * self.width :] = 1
 817         result *= 1 - ar_mask
 818         masked_inplace_autoregression(
 819             model,
 820             self.batch_size,
 821             result,
 822             ar_mask,
 823             deterministic_synthesis,
 824             device=self.device,
 825         )
 826
 827         mazes, paths = self.seq2map(input)
 828         _, predicted_paths = self.seq2map(result)
 829
 830         filename = os.path.join(result_dir, f"maze_result_{n_epoch:04d}.png")
 831         maze.save_image(
 832             filename,
 833             mazes=mazes,
 834             target_paths=paths,
 835             predicted_paths=predicted_paths,
 836             path_correct=maze.path_correctness(mazes, predicted_paths),
 837             path_optimal=maze.path_optimality(paths, predicted_paths),
 838         )
 839         logger(f"wrote {filename}")
 840
 841
 842 ######################################################################
 843
 844
 845 import snake
 846
 847
 848 class Snake(Task):
 849     def __init__(
 850         self,
 851         nb_train_samples,
 852         nb_test_samples,
 853         batch_size,
 854         height,
 855         width,
 856         nb_colors,
 857         length,
 858         prompt_length,
 859         device=torch.device("cpu"),
 860     ):
 861         super().__init__()
 862
 863         self.batch_size = batch_size
 864         self.height = height
 865         self.width = width
 866         self.device = device
 867         self.prompt_length = prompt_length
 868
 869         self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences(
 870             nb_train_samples,
 871             height,
 872             width,
 873             nb_colors,
 874             length,
 875             prompt_length,
 876             self.device,
 877         )
 878         self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences(
 879             nb_test_samples,
 880             height,
 881             width,
 882             nb_colors,
 883             length,
 884             prompt_length,
 885             self.device,
 886         )
 887
 888         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
 889
 890     def batches(self, split="train", nb_to_use=-1, desc=None):
 891         assert split in {"train", "test"}
 892         input = self.train_input if split == "train" else self.test_input
 893         if nb_to_use > 0:
 894             input = input[:nb_to_use]
 895         if desc is None:
 896             desc = f"epoch-{split}"
 897         for batch in tqdm.tqdm(
 898             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 899         ):
 900             yield batch
 901
 902     def vocabulary_size(self):
 903         return self.nb_codes
 904
 905     def produce_results(
 906         self, n_epoch, model, result_dir, logger, deterministic_synthesis
 907     ):
 908         def compute_nb_correct(input, prior_visits):
 909             result = input.clone()
 910             i = torch.arange(result.size(1), device=result.device)[None, :]
 911             ar_mask = (
 912                 torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0)
 913                 .long()
 914                 .expand_as(result)
 915             )
 916             result *= 1 - ar_mask
 917
 918             masked_inplace_autoregression(
 919                 model,
 920                 self.batch_size,
 921                 result,
 922                 ar_mask,
 923                 deterministic_synthesis,
 924                 device=self.device,
 925             )
 926
 927             nb_total = ((prior_visits > 0) * ar_mask).sum()
 928
 929             nb_correct = ((result == input).long() * (prior_visits > 0) * ar_mask).sum()
 930
 931             return nb_total, nb_correct
 932
 933         test_nb_total, test_nb_correct = compute_nb_correct(
 934             self.test_input[:1000], self.test_prior_visits[:1000]
 935         )
 936
 937         logger(
 938             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}%"
 939         )
 940
 941         logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
 942
 943
 944 ######################################################################
 945
 946
 947 import stack
 948
 949
 950 class Stack(Task):
 951     def __init__(
 952         self,
 953         nb_train_samples,
 954         nb_test_samples,
 955         batch_size,
 956         logger,
 957         nb_steps,
 958         nb_stacks,
 959         nb_digits,
 960         fraction_values_for_train=None,
 961         device=torch.device("cpu"),
 962     ):
 963         super().__init__()
 964
 965         self.batch_size = batch_size
 966         self.nb_steps = nb_steps
 967         self.nb_stacks = nb_stacks
 968         self.nb_digits = nb_digits
 969         self.device = device
 970
 971         if fraction_values_for_train is None:
 972             values_for_train = None
 973             values_for_test = None
 974         else:
 975             all = torch.randperm(10**nb_digits)
 976             nb_for_train = int(all.size(0) * fraction_values_for_train)
 977             values_for_train = all[:nb_for_train]
 978             values_for_test = all[nb_for_train:]
 979
 980         self.train_input, self.train_stack_counts = stack.generate_sequences(
 981             nb_train_samples,
 982             nb_steps,
 983             nb_stacks,
 984             nb_digits,
 985             values_for_train,
 986             self.device,
 987         )
 988
 989         self.test_input, self.test_stack_counts = stack.generate_sequences(
 990             nb_test_samples,
 991             nb_steps,
 992             nb_stacks,
 993             nb_digits,
 994             values_for_test,
 995             self.device,
 996         )
 997
 998         i = torch.logical_and(self.test_input % 2 == 1, self.test_input < 2 * nb_stacks)
 999         counts = self.test_stack_counts.flatten()[i.flatten()]
1000         counts = F.one_hot(counts).sum(0)
1001         logger(f"test_pop_stack_counts {counts}")
1002
1003         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1004
1005     def batches(self, split="train", nb_to_use=-1, desc=None):
1006         assert split in {"train", "test"}
1007         input = self.train_input if split == "train" else self.test_input
1008         if nb_to_use > 0:
1009             input = input[:nb_to_use]
1010         if desc is None:
1011             desc = f"epoch-{split}"
1012         for batch in tqdm.tqdm(
1013             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1014         ):
1015             yield batch
1016
1017     def vocabulary_size(self):
1018         return self.nb_codes
1019
1020     def produce_results(
1021         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1022     ):
1023         def compute_nb_correct(input):
1024             result = input.clone()
1025             stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
1026             ar_mask = (result != input).long()
1027             masked_inplace_autoregression(
1028                 model,
1029                 self.batch_size,
1030                 result,
1031                 ar_mask,
1032                 deterministic_synthesis,
1033                 device=self.device,
1034             )
1035
1036             errors = ((result != input).long() * ar_mask).reshape(
1037                 -1, 1 + self.nb_digits
1038             )
1039             ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits)
1040
1041             nb_total = ar_mask.max(1).values.sum()
1042             nb_correct = nb_total - errors.max(1).values.sum()
1043
1044             return nb_total, nb_correct
1045
1046         test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000])
1047
1048         logger(
1049             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}%"
1050         )
1051
1052         logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
1053
1054         ##############################################################
1055         # Log a few generated sequences
1056         input = self.test_input[:10, : 12 * (1 + self.nb_digits)]
1057         result = input.clone()
1058         stack.remove_popped_values(result, self.nb_stacks, self.nb_digits)
1059         ar_mask = (result != input).long()
1060
1061         # for n in range(result.size(0)):
1062         # logger(
1063         # f"test_before {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
1064         # )
1065
1066         masked_inplace_autoregression(
1067             model,
1068             self.batch_size,
1069             result,
1070             ar_mask,
1071             deterministic_synthesis,
1072             device=self.device,
1073         )
1074
1075         for n in range(result.size(0)):
1076             logger(
1077                 f"test_after  {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}"
1078             )
1079         ##############################################################
1080
1081
1082 ######################################################################
1083
1084 import rpl
1085
1086
1087 class RPL(Task):
1088     def tensorize(self, sequences):
1089         len_max = max([len(x) for x in sequences])
1090         return torch.cat(
1091             [
1092                 torch.tensor(
1093                     [
1094                         [
1095                             self.token2id[str(c)]
1096                             for c in s + ["<nul>"] * (len_max - len(s))
1097                         ]
1098                         for s in sequences
1099                     ]
1100                 )
1101             ],
1102             0,
1103         )
1104
1105     def seq2str(self, seq):
1106         return " ".join([self.id2token[i] for i in seq])
1107
1108     def __init__(
1109         self,
1110         nb_train_samples,
1111         nb_test_samples,
1112         batch_size,
1113         nb_starting_values=3,
1114         max_input=9,
1115         prog_len=6,
1116         nb_runs=5,
1117         no_prog=False,
1118         logger=None,
1119         device=torch.device("cpu"),
1120     ):
1121         super().__init__()
1122
1123         self.batch_size = batch_size
1124         self.device = device
1125         self.no_prog = no_prog
1126
1127         train_sequences = [
1128             rpl.generate(
1129                 nb_starting_values=nb_starting_values,
1130                 nb_result_values_max=4 * nb_starting_values,
1131                 max_input=max_input,
1132                 prog_len=prog_len,
1133                 nb_runs=nb_runs,
1134             )
1135             for _ in tqdm.tqdm(range(nb_train_samples), desc="train-data")
1136         ]
1137
1138         test_sequences = [
1139             rpl.generate(
1140                 nb_starting_values=nb_starting_values,
1141                 nb_result_values_max=4 * nb_starting_values,
1142                 max_input=max_input,
1143                 prog_len=prog_len,
1144                 nb_runs=nb_runs,
1145             )
1146             for _ in tqdm.tqdm(range(nb_test_samples), desc="test-data")
1147         ]
1148
1149         symbols = list(
1150             set(["<nul>"] + [x for l in train_sequences + test_sequences for x in l])
1151         )
1152         val_max = max([x if type(x) is int else 0 for x in symbols])
1153         symbols = list(filter(lambda x: type(x) is str, symbols))
1154         symbols.sort()
1155         symbols += [str(n) for n in range(val_max + 1)]
1156         self.token2id = dict([(c, n) for n, c in enumerate(symbols)])
1157         self.id2token = dict([(n, c) for c, n in self.token2id.items()])
1158
1159         self.t_nul = self.token2id["<nul>"]
1160         self.t_input = self.token2id["<in>"]
1161         self.t_output = self.token2id["<out>"]
1162         self.t_prog = self.token2id["<prg>"]
1163         self.t_end = self.token2id["<end>"]
1164
1165         self.train_input = self.tensorize(train_sequences)
1166         self.test_input = self.tensorize(test_sequences)
1167
1168         if no_prog:
1169             # Excise the program from every train and test example
1170             k = torch.arange(self.train_input.size(1), device=self.train_input.device)[
1171                 None, :
1172             ]
1173             p = (
1174                 ((self.train_input == self.t_prog).long() * k)
1175                 .max(1, keepdim=True)
1176                 .values
1177             )
1178             self.train_input = (
1179                 self.train_input * (k <= p).long()
1180                 + self.t_end * (k == p + 1).long()
1181                 + self.t_nul * (k > p + 1).long()
1182             )
1183             k = torch.arange(self.test_input.size(1), device=self.test_input.device)[
1184                 None, :
1185             ]
1186             p = (
1187                 ((self.test_input == self.t_prog).long() * k)
1188                 .max(1, keepdim=True)
1189                 .values
1190             )
1191             self.test_input = (
1192                 self.test_input * (k <= p).long()
1193                 + self.t_end * (k == p + 1).long()
1194                 + self.t_nul * (k > p + 1).long()
1195             )
1196
1197         if logger is not None:
1198             logger(f"value_max {val_max}")
1199             for x in self.train_input[:25]:
1200                 end = (x != self.t_nul).nonzero().max().item() + 1
1201                 seq = [self.id2token[i.item()] for i in x[:end]]
1202                 s = " ".join(seq)
1203                 logger(f"example_seq {s}")
1204
1205         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1206
1207     def batches(self, split="train", nb_to_use=-1, desc=None):
1208         assert split in {"train", "test"}
1209         input = self.train_input if split == "train" else self.test_input
1210         if nb_to_use > 0:
1211             input = input[:nb_to_use]
1212         if desc is None:
1213             desc = f"epoch-{split}"
1214         for batch in tqdm.tqdm(
1215             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1216         ):
1217             last = (batch != self.t_nul).max(0).values.nonzero().max() + 3
1218             batch = batch[:, :last].to(self.device)
1219             yield batch
1220
1221     def vocabulary_size(self):
1222         return self.nb_codes
1223
1224     def produce_results(
1225         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1226     ):
1227         # --------------------------------------------------------------------
1228         def compute_nb_errors_prog(input, nb_to_log=0):
1229             result = input.clone()
1230             s = (result == self.t_prog).long()
1231             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1232             result = (1 - ar_mask) * result + ar_mask * self.t_nul
1233
1234             masked_inplace_autoregression(
1235                 model,
1236                 self.batch_size,
1237                 result,
1238                 ar_mask,
1239                 deterministic_synthesis,
1240                 device=self.device,
1241             )
1242
1243             sum_nb_total, sum_nb_errors = 0, 0
1244             for one_input, one_result in zip(input, result):
1245                 seq = [self.id2token[i.item()] for i in one_result]
1246                 nb_total, nb_errors, prog, stacks = rpl.compute_nb_errors(seq)
1247                 sum_nb_total += 1
1248                 sum_nb_errors += 0 if nb_errors == 0 else 1
1249                 if nb_to_log > 0:
1250                     gt_seq = [self.id2token[i.item()] for i in one_input]
1251                     _, _, gt_prog, _ = rpl.compute_nb_errors(gt_seq)
1252                     gt_prog = " ".join([str(x) for x in gt_prog])
1253                     prog = " ".join([str(x) for x in prog])
1254                     comment = "*" if nb_errors == 0 else "-"
1255                     logger(f"{comment} PROG [{gt_prog}] PREDICTED [{prog}]")
1256                     for start_stack, target_stack, result_stack, correct in stacks:
1257                         comment = "*" if correct else "-"
1258                         start_stack = " ".join([str(x) for x in start_stack])
1259                         target_stack = " ".join([str(x) for x in target_stack])
1260                         result_stack = " ".join([str(x) for x in result_stack])
1261                         logger(
1262                             f"  {comment} [{start_stack}] -> [{target_stack}] PREDICTED [{result_stack}]"
1263                         )
1264                     nb_to_log -= 1
1265
1266             return sum_nb_total, sum_nb_errors
1267
1268         # --------------------------------------------------------------------
1269         def compute_nb_errors_output(input, nb_to_log=0):
1270             result = input.clone()
1271             k = torch.arange(result.size(1), device=result.device)[None, :]
1272             last_output_idx = (
1273                 ((result == self.t_output) * k).max(dim=1, keepdim=True).values
1274             )
1275             first_prog_idx = (
1276                 ((result == self.t_prog) * k).max(dim=1, keepdim=True).values
1277             )
1278             ar_mask = (k > last_output_idx).long() * (k < first_prog_idx).long()
1279             result = (1 - ar_mask) * result + ar_mask * self.t_nul
1280
1281             masked_inplace_autoregression(
1282                 model,
1283                 self.batch_size,
1284                 result,
1285                 ar_mask,
1286                 deterministic_synthesis,
1287                 device=self.device,
1288             )
1289
1290             sum_nb_total, sum_nb_errors = 0, 0
1291             for one_input, one_result, i, j in zip(
1292                 input, result, last_output_idx, first_prog_idx
1293             ):
1294                 seq = [self.id2token[i.item()] for i in one_result]
1295                 sum_nb_total += 1
1296                 correct = (one_input - one_result).abs().max() == 0
1297                 sum_nb_errors += 0 if correct else 1
1298                 if nb_to_log > 0:
1299                     result_stack = [
1300                         self.id2token[i.item()] for i in one_result[i : j + 1]
1301                     ]
1302                     target_stack = [
1303                         self.id2token[i.item()] for i in one_input[i : j + 1]
1304                     ]
1305                     comment = "*" if correct else "-"
1306                     result_stack = " ".join([str(x) for x in result_stack])
1307                     target_stack = " ".join([str(x) for x in target_stack])
1308                     logger(
1309                         f"output_test {comment} [{target_stack}] PREDICTED [{result_stack}]"
1310                     )
1311                     nb_to_log -= 1
1312
1313             return sum_nb_total, sum_nb_errors
1314
1315         # --------------------------------------------------------------------
1316
1317         if not self.no_prog:
1318             test_nb_total, test_nb_errors = compute_nb_errors_prog(
1319                 self.test_input[:1000].to(self.device), nb_to_log=10
1320             )
1321
1322             logger(
1323                 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}%"
1324             )
1325
1326             logger(f"main_test_accuracy {n_epoch} {1-test_nb_errors/test_nb_total}")
1327
1328         test_nb_total, test_nb_errors = compute_nb_errors_output(
1329             self.test_input[:1000].to(self.device), nb_to_log=10
1330         )
1331
1332         logger(
1333             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}%"
1334         )
1335
1336         if save_attention_image is None:
1337             logger("no save_attention_image (is pycairo installed?)")
1338         else:
1339             ns = torch.randint(self.test_input.size(0), (1,)).item()
1340             input = self.test_input[ns : ns + 1].clone()
1341             last = (input != self.t_nul).max(0).values.nonzero().max() + 3
1342             input = input[:, :last].to(self.device)
1343
1344             with torch.autograd.no_grad():
1345                 t = model.training
1346                 model.eval()
1347                 model.record_attention(True)
1348                 model(BracketedSequence(input))
1349                 model.train(t)
1350                 ram = model.retrieve_attention()
1351                 model.record_attention(False)
1352
1353             tokens_output = [self.id2token[i.item()] for i in input[0]]
1354             tokens_input = ["n/a"] + tokens_output[:-1]
1355             for n_head in range(ram[0].size(1)):
1356                 filename = os.path.join(
1357                     result_dir, f"rpl_attention_{n_epoch}_h{n_head}.pdf"
1358                 )
1359                 attention_matrices = [m[0, n_head] for m in ram]
1360                 save_attention_image(
1361                     filename,
1362                     tokens_input,
1363                     tokens_output,
1364                     attention_matrices,
1365                     k_top=10,
1366                     # min_total_attention=0.9,
1367                     token_gap=12,
1368                     layer_gap=50,
1369                 )
1370                 logger(f"wrote {filename}")
1371
1372
1373 ######################################################################
1374
1375
1376 import expr
1377
1378
1379 class Expr(Task):
1380     def tensorize(self, sequences):
1381         len_max = max([len(x) for x in sequences])
1382         return torch.cat(
1383             [
1384                 torch.tensor(
1385                     [
1386                         [self.char2id[c] for c in s + "#" * (len_max - len(s))]
1387                         for s in sequences
1388                     ]
1389                 )
1390             ],
1391             0,
1392         ).to(self.device)
1393
1394     def __init__(
1395         self,
1396         nb_train_samples,
1397         nb_test_samples,
1398         nb_variables,
1399         sequence_length,
1400         operand_max,
1401         result_max,
1402         batch_size,
1403         device=torch.device("cpu"),
1404     ):
1405         super().__init__()
1406
1407         self.batch_size = batch_size
1408         self.device = device
1409
1410         train_sequences = expr.generate_sequences(
1411             nb_train_samples,
1412             nb_variables=nb_variables,
1413             length=sequence_length,
1414             operand_max=operand_max,
1415             result_max=result_max,
1416         )
1417
1418         test_sequences = expr.generate_sequences(
1419             nb_test_samples,
1420             nb_variables=nb_variables,
1421             length=sequence_length,
1422             operand_max=operand_max,
1423             result_max=result_max,
1424         )
1425
1426         symbols = list(set("#" + "".join(train_sequences + test_sequences)))
1427         symbols.sort()
1428
1429         self.char2id = dict([(c, n) for n, c in enumerate(symbols)])
1430         self.id2char = dict([(n, c) for c, n in self.char2id.items()])
1431
1432         self.filler, self.space = self.char2id["#"], self.char2id[" "]
1433
1434         self.train_input = self.tensorize(train_sequences)
1435         self.test_input = self.tensorize(test_sequences)
1436
1437         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1438
1439     def batches(self, split="train", nb_to_use=-1, desc=None):
1440         assert split in {"train", "test"}
1441         input = self.train_input if split == "train" else self.test_input
1442         if nb_to_use > 0:
1443             input = input[:nb_to_use]
1444         if desc is None:
1445             desc = f"epoch-{split}"
1446         for batch in tqdm.tqdm(
1447             input.split(self.batch_size), dynamic_ncols=True, desc=desc
1448         ):
1449             last = (batch != self.filler).max(0).values.nonzero().max() + 3
1450             batch = batch[:, :last]
1451             yield batch
1452
1453     def vocabulary_size(self):
1454         return self.nb_codes
1455
1456     def seq2str(self, s):
1457         return "".join([self.id2char[k.item()] for k in s])
1458
1459     def produce_results(
1460         self,
1461         n_epoch,
1462         model,
1463         result_dir,
1464         logger,
1465         deterministic_synthesis,
1466         input_file=None,
1467     ):
1468         def compute_nb_correct(input):
1469             result = input.clone()
1470             s = (result == self.space).long()
1471             ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1472             result = (1 - ar_mask) * result + ar_mask * self.filler
1473             masked_inplace_autoregression(
1474                 model,
1475                 self.batch_size,
1476                 result,
1477                 ar_mask,
1478                 deterministic_synthesis,
1479                 device=self.device,
1480             )
1481
1482             nb_total = input.size(0)
1483             nb_correct = (input == result).long().min(1).values.sum()
1484
1485             #######################################################################
1486             # Comput predicted vs. true variable values
1487
1488             nb_delta = torch.zeros(5, dtype=torch.int64)
1489             nb_missed = 0
1490
1491             values_input = expr.extract_results([self.seq2str(s) for s in input])
1492             values_result = expr.extract_results([self.seq2str(s) for s in result])
1493
1494             filename = os.path.join(result_dir, f"expr_result_{n_epoch:04d}.txt")
1495
1496             with open(filename, "w") as f:
1497                 for i, r in zip(values_input, values_result):
1498                     for n, vi in i.items():
1499                         vr = r.get(n)
1500                         f.write(f"{vi} {-1 if vr is None else vr}\n")
1501
1502                         if vr is None or vr < 0:
1503                             nb_missed += 1
1504                         else:
1505                             d = abs(vr - vi)
1506                             if d >= nb_delta.size(0):
1507                                 nb_missed += 1
1508                             else:
1509                                 nb_delta[d] += 1
1510
1511             ######################################################################
1512
1513             return nb_total, nb_correct, nb_delta, nb_missed
1514
1515         (
1516             test_nb_total,
1517             test_nb_correct,
1518             test_nb_delta,
1519             test_nb_missed,
1520         ) = compute_nb_correct(self.test_input[:10000])
1521
1522         logger(
1523             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}%"
1524         )
1525
1526         logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
1527
1528         nb_total = test_nb_delta.sum() + test_nb_missed
1529         for d in range(test_nb_delta.size(0)):
1530             logger(
1531                 f"error_value {n_epoch} delta {d} {test_nb_delta[d]} {test_nb_delta[d]*100/nb_total:.02f}%"
1532             )
1533         logger(
1534             f"error_value {n_epoch} missed {test_nb_missed} {test_nb_missed*100/nb_total:.02f}%"
1535         )
1536
1537         ##############################################################
1538         # Log a few generated sequences
1539         if input_file is None:
1540             input = self.test_input[:10]
1541         else:
1542             with open(input_file, "r") as f:
1543                 sequences = [e.strip() for e in f.readlines()]
1544                 sequences = [s + " " + "#" * 50 for s in sequences]
1545                 input = self.tensorize(sequences)
1546
1547         result = input.clone()
1548         s = (result == self.space).long()
1549         ar_mask = (s.cumsum(dim=1) - s).clamp(min=0, max=1)
1550         result = (1 - ar_mask) * result + ar_mask * self.filler
1551
1552         for n in range(result.size(0)):
1553             logger(f"test_before {self.seq2str(result[n])}")
1554
1555         masked_inplace_autoregression(
1556             model,
1557             self.batch_size,
1558             result,
1559             ar_mask,
1560             deterministic_synthesis,
1561             device=self.device,
1562         )
1563
1564         correct = (1 - ar_mask) * self.space + ar_mask * input
1565         for n in range(result.size(0)):
1566             comment = "GOOD" if (result[n] - input[n]).abs().max() == 0 else ""
1567             logger(f"test_after  {self.seq2str(result[n])} {comment}")
1568             logger(f"truth       {self.seq2str(correct[n])}")
1569         ##############################################################
1570
1571
1572 ######################################################################
1573
1574 import grid
1575
1576
1577 class Grid(Task):
1578     # Make a tensor from a list of strings
1579     def str2tensor(self, descr):
1580         token_descr = [s.strip().split(" ") for s in descr]
1581         l = max([len(s) for s in token_descr])
1582         token_descr = [s + ["#"] * (l - len(s)) for s in token_descr]
1583         id_descr = [[self.token2id[u] for u in s] for s in token_descr]
1584         return torch.tensor(id_descr, device=self.device)
1585
1586     # Make a list of strings from a tensor
1587     def tensor2str(self, x):
1588         return [" ".join([self.id2token[t.item()] for t in r]) for r in x]
1589
1590     # trim all the tensors in the tuple z to remove as much token from
1591     # left and right in the first tensor. If z is a tuple, all its
1592     # elements are trimed according to the triming for the first
1593     def trim(self, z, token="#"):
1594         n = self.token2id[token]
1595         if type(z) == tuple:
1596             x = z[0]
1597             i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
1598             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
1599             return tuple([t[:, a:b] for t in z])
1600         else:
1601             i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0)
1602             a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
1603             return z[:, a:b]
1604
1605     ######################
1606
1607     def __init__(
1608         self,
1609         nb_train_samples,
1610         nb_test_samples,
1611         batch_size,
1612         size,
1613         fraction_play=0.0,
1614         logger=None,
1615         device=torch.device("cpu"),
1616     ):
1617         super().__init__()
1618
1619         self.device = device
1620         self.batch_size = batch_size
1621         self.grid_factory = grid.GridFactory(size=size)
1622         self.fraction_play = fraction_play
1623
1624         if logger is not None:
1625             logger(
1626                 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
1627             )
1628
1629         self.train_descr = self.grid_factory.generate_samples(
1630             nb=nb_train_samples,
1631             fraction_play=fraction_play,
1632             progress_bar=lambda r: tqdm.tqdm(r),
1633         )
1634
1635         self.test_descr = self.grid_factory.generate_samples(
1636             nb=nb_test_samples, fraction_play=0.0, progress_bar=lambda r: tqdm.tqdm(r)
1637         )
1638
1639         if fraction_play > 0:
1640             self.play_descr = self.grid_factory.generate_samples(
1641                 nb=25, fraction_play=1.0, progress_bar=lambda r: tqdm.tqdm(r)
1642             )
1643         else:
1644             self.play_descr = []
1645
1646         # Build the tokenizer
1647         tokens = set()
1648         for d in [self.train_descr, self.test_descr, self.play_descr]:
1649             for s in d:
1650                 for t in s.strip().split(" "):
1651                     tokens.add(t)
1652         # make this set a sorted list to get the same tensors given
1653         # the same descr
1654         tokens = list(tokens)
1655         tokens.sort()
1656         tokens = ["#"] + tokens
1657         self.token2id = dict([(t, n) for n, t in enumerate(tokens)])
1658         self.id2token = dict([(n, t) for n, t in enumerate(tokens)])
1659         self.t_nul = self.token2id["#"]
1660         self.t_true = self.token2id["true"]
1661         self.t_false = self.token2id["false"]
1662         self.t_pipe = self.token2id["|"]
1663
1664         # Tokenize the train and test sets
1665         self.train_input = self.str2tensor(self.train_descr)
1666         self.test_input = self.str2tensor(self.test_descr)
1667         self.play_input = (
1668             None if len(self.play_descr) == 0 else self.str2tensor(self.play_descr)
1669         )
1670
1671     def batches(self, split="train"):
1672         assert split in {"train", "test"}
1673         input = self.train_input if split == "train" else self.test_input
1674         for batch in tqdm.tqdm(
1675             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
1676         ):
1677             yield self.trim(batch)
1678
1679     def vocabulary_size(self):
1680         return len(self.token2id)
1681
1682     def produce_results(
1683         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1684     ):
1685         correct = self.test_input[:1000]
1686         result = correct.clone()
1687         ar_mask = torch.logical_or(result == self.t_true, result == self.t_false).long()
1688         result *= 1 - ar_mask  # paraaaaanoiaaaaaaa
1689
1690         logger(f"----------------------------------------------------------")
1691
1692         for e in self.tensor2str(result[:10]):
1693             logger(f"test_before {e}")
1694
1695         masked_inplace_autoregression(
1696             model,
1697             self.batch_size,
1698             result,
1699             ar_mask,
1700             deterministic_synthesis,
1701             device=self.device,
1702         )
1703
1704         logger(f"----------------------------------------------------------")
1705
1706         for e in self.tensor2str(result[:10]):
1707             logger(f"test_after  {e}")
1708
1709         logger(f"----------------------------------------------------------")
1710
1711         nb_total = ar_mask.sum().item()
1712         nb_correct = ((correct == result).long() * ar_mask).sum().item()
1713
1714         logger(f"test_performance {n_epoch} {nb_total=} {nb_correct=}")
1715         logger(f"main_test_accuracy {n_epoch} {nb_correct / nb_total}")
1716
1717         if self.play_input is not None:
1718             result = self.play_input.clone()
1719             ar_mask = (result == self.t_pipe).long().cumsum(dim=1).clamp(max=1)
1720             result *= 1 - ar_mask  # paraaaaanoiaaaaaaa
1721
1722             logger(f"----------------------------------------------------------")
1723
1724             for e in self.tensor2str(result[:10]):
1725                 logger(f"play_before {e}")
1726
1727             masked_inplace_autoregression(
1728                 model,
1729                 self.batch_size,
1730                 result,
1731                 ar_mask,
1732                 deterministic_synthesis,
1733                 device=self.device,
1734             )
1735
1736             logger(f"----------------------------------------------------------")
1737
1738             for e in self.tensor2str(result[:10]):
1739                 logger(f"play_after  {e}")
1740
1741             logger(f"----------------------------------------------------------")
1742
1743
1744 ######################################################################
1745
1746 import qmlp
1747
1748
1749 class QMLP(Task):
1750     ######################
1751
1752     def __init__(
1753         self,
1754         nb_train_samples,
1755         nb_test_samples,
1756         batch_size,
1757         result_dir,
1758         logger=None,
1759         device=torch.device("cpu"),
1760     ):
1761         super().__init__()
1762
1763         self.device = device
1764         self.batch_size = batch_size
1765         self.nb_samples_per_mlp = 256
1766
1767         if logger is not None:
1768             logger(
1769                 f"generating {nb_train_samples+nb_test_samples} samples (can take some time)"
1770             )
1771
1772         seq, q_test_set, test_error = qmlp.generate_sequence_and_test_set(
1773             nb_mlps=nb_train_samples + nb_test_samples,
1774             nb_samples=self.nb_samples_per_mlp,
1775             device=self.device,
1776             batch_size=64,
1777             nb_epochs=250,
1778             nb_mlps_per_batch=1024,
1779         )
1780
1781         self.train_input = seq[:nb_train_samples]
1782         self.train_q_test_set = q_test_set[:nb_train_samples]
1783         self.train_ref_test_errors = test_error[:nb_train_samples]
1784         self.test_input = seq[nb_train_samples:]
1785         self.test_q_test_set = q_test_set[nb_train_samples:]
1786         self.test_ref_test_errors = test_error[nb_train_samples:]
1787
1788         filename = os.path.join(result_dir, f"train_errors_ref.dat")
1789         with open(filename, "w") as f:
1790             for e in self.train_ref_test_errors:
1791                 f.write(f"{e}\n")
1792
1793         filename = os.path.join(result_dir, f"test_errors_ref.dat")
1794         with open(filename, "w") as f:
1795             for e in self.test_ref_test_errors:
1796                 f.write(f"{e}\n")
1797
1798         self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
1799
1800     def batches(self, split="train"):
1801         assert split in {"train", "test"}
1802         input = self.train_input if split == "train" else self.test_input
1803         for batch in tqdm.tqdm(
1804             input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}"
1805         ):
1806             yield batch
1807
1808     def vocabulary_size(self):
1809         return self.nb_codes
1810
1811     def produce_results(
1812         self, n_epoch, model, result_dir, logger, deterministic_synthesis
1813     ):
1814         correct = self.test_input[:1000]
1815         result = correct.clone()
1816         ar_mask = (
1817             torch.arange(result.size(1), device=result.device)
1818             > self.nb_samples_per_mlp * 3 + 1
1819         ).long()[None, :]
1820         ar_mask = ar_mask.expand_as(result)
1821         result *= 1 - ar_mask  # paraaaaanoiaaaaaaa
1822
1823         masked_inplace_autoregression(
1824             model,
1825             self.batch_size,
1826             result,
1827             ar_mask,
1828             deterministic_synthesis,
1829             device=self.device,
1830         )
1831
1832         q_train_set = result[:, : self.nb_samples_per_mlp * 3]
1833         q_params = result[:, self.nb_samples_per_mlp * 3 + 1 :]
1834         error_test = qmlp.evaluate_q_params(q_params, self.test_q_test_set)
1835
1836         filename = os.path.join(result_dir, f"test_errors_{n_epoch:04d}.dat")
1837         with open(filename, "w") as f:
1838             for e in error_test:
1839                 f.write(f"{e}\n")
1840
1841
1842 ######################################################################