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