beaver.py

   1 #!/usr/bin/env python
   2
   3 # Any copyright is dedicated to the Public Domain.
   4 # https://creativecommons.org/publicdomain/zero/1.0/
   5
   6 # Written by Francois Fleuret <francois@fleuret.org>
   7
   8 # torch.backends.cuda.matmul.allow_tf23
   9 # torch.autocast(torch.bfloat16)
  10
  11 import math, sys, argparse, time, tqdm, itertools, os
  12
  13 import torch, torchvision
  14 from torch import nn
  15 from torch.nn import functional as F
  16
  17 import mygpt, tensorstack
  18
  19 ######################################################################
  20
  21 if torch.cuda.is_available():
  22     device = torch.device("cuda")
  23     torch.backends.cuda.matmul.allow_tf32 = True
  24 else:
  25     device = torch.device("cpu")
  26
  27 ######################################################################
  28
  29 parser = argparse.ArgumentParser(description="A maze shortest path solving with a GPT.")
  30
  31 parser.add_argument("--log_filename", type=str, default="train.log")
  32
  33 parser.add_argument("--result_dir", type=str, default="results_default")
  34
  35 parser.add_argument("--seed", type=int, default=0)
  36
  37 parser.add_argument("--nb_epochs", type=int, default=25)
  38
  39 parser.add_argument("--nb_train_samples", type=int, default=200000)
  40
  41 parser.add_argument("--nb_test_samples", type=int, default=50000)
  42
  43 parser.add_argument("--batch_size", type=int, default=25)
  44
  45 parser.add_argument("--optim", type=str, default="adam")
  46
  47 parser.add_argument("--learning_rate", type=float, default=1e-3)
  48
  49 parser.add_argument(
  50     "--learning_rate_schedule", type=str, default="10: 2e-4,20: 4e-5,30: 8e-6"
  51 )
  52
  53 parser.add_argument("--dim_model", type=int, default=512)
  54
  55 parser.add_argument("--dim_keys", type=int, default=64)
  56
  57 parser.add_argument("--dim_hidden", type=int, default=2048)
  58
  59 parser.add_argument("--nb_heads", type=int, default=8)
  60
  61 parser.add_argument("--nb_blocks", type=int, default=12)
  62
  63 parser.add_argument("--dropout", type=float, default=0.1)
  64
  65 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
  66
  67 parser.add_argument("--random_regression_order", action="store_true", default=False)
  68
  69 parser.add_argument("--noncausal_prompt", action="store_true", default=False)
  70
  71 parser.add_argument("--no_checkpoint", action="store_true", default=False)
  72
  73 parser.add_argument("--overwrite_results", action="store_true", default=False)
  74
  75 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
  76
  77 ##############################
  78 # maze options
  79
  80 parser.add_argument("--maze_height", type=int, default=13)
  81
  82 parser.add_argument("--maze_width", type=int, default=21)
  83
  84 parser.add_argument("--maze_nb_walls", type=int, default=15)
  85
  86 ##############################
  87 # one-shot prediction
  88
  89 parser.add_argument("--oneshot", action="store_true", default=False)
  90
  91 parser.add_argument("--oneshot_input", type=str, default="head")
  92
  93 parser.add_argument("--oneshot_output", type=str, default="trace")
  94
  95 ######################################################################
  96
  97 args = parser.parse_args()
  98
  99 try:
 100     os.mkdir(args.result_dir)
 101 except FileExistsError:
 102     if not args.overwrite_results:
 103         print(f"result directory {args.result_dir} already exists")
 104         exit(1)
 105
 106 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 107
 108 if args.seed >= 0:
 109     # torch.backends.cudnn.deterministic = True
 110     # torch.backends.cudnn.benchmark = False
 111     # torch.use_deterministic_algorithms(True)
 112     torch.manual_seed(args.seed)
 113     if torch.cuda.is_available():
 114         torch.cuda.manual_seed_all(args.seed)
 115
 116 ######################################################################
 117
 118
 119 def log_string(s):
 120     t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
 121
 122     if log_file is not None:
 123         log_file.write(t + s + "\n")
 124         log_file.flush()
 125
 126     print(t + s)
 127     sys.stdout.flush()
 128
 129
 130 for n in vars(args):
 131     log_string(f"args.{n} {getattr(args, n)}")
 132
 133 ######################################################################
 134
 135
 136 def generation_order(x, fixed_len=0):
 137     if args.random_regression_order:
 138         order = torch.rand(x.size(), device=x.device)
 139         order[:, :fixed_len] = torch.arange(-fixed_len, 0, device=x.device)
 140         order = order.sort(1).indices
 141     else:
 142         order = (
 143             torch.arange(x.size(1), device=x.device).unsqueeze(0).expand(x.size(0), -1)
 144         )
 145     return order
 146
 147
 148 def reorder(x, order, reverse=False):  # x is NxTxD1x...xDk, order is NxT'
 149     u = x.reshape(x.size()[:2] + (-1,))
 150     order = order.unsqueeze(-1).expand(-1, -1, u.size(-1))
 151     if reverse:
 152         v = u.new(u.size()).scatter_(1, order, u)
 153     else:
 154         v = u.gather(1, order)
 155     v = v.reshape(v.size()[:2] + x.size()[2:])
 156     return v
 157
 158
 159 def shuffle(x, fixed_len):
 160     order = generation_order(x, fixed_len)
 161     return reorder(x, order), order
 162
 163
 164 def eval_mygpt(model, input, mode="standard", fixed_len=0):
 165     x, order = shuffle(input, fixed_len)
 166     x = model(mygpt.BracketedSequence(x), mode=mode, order=order).x
 167     return reorder(x, order, reverse=True)
 168
 169
 170 ######################################################################
 171
 172 # ar_mask is a Boolean matrix of same shape as input, with 1s on the
 173 # tokens that should be generated
 174
 175
 176 def masked_inplace_autoregression(model, batch_size, input, ar_mask, order=None):
 177     for input, ar_mask, order in zip(
 178         input.split(batch_size), ar_mask.split(batch_size), order.split(batch_size)
 179     ):
 180         i = (ar_mask.sum(0) > 0).nonzero()
 181         if i.min() > 0:
 182             # Needed to initialize the model's cache
 183             model(mygpt.BracketedSequence(input, 0, i.min()), order=order)
 184         for s in range(i.min(), i.max() + 1):
 185             output = model(mygpt.BracketedSequence(input, s, 1), order=order).x
 186             logits = output[:, s]
 187             if args.deterministic_synthesis:
 188                 t_next = logits.argmax(1)
 189             else:
 190                 dist = torch.distributions.categorical.Categorical(logits=logits)
 191                 t_next = dist.sample()
 192             input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
 193
 194
 195 ######################################################################
 196
 197
 198 def compute_perplexity(model, task, fixed_len, split="train"):
 199     with torch.autograd.no_grad():
 200         t = model.training
 201         model.eval()
 202
 203         nb_samples, acc_loss = 0, 0.0
 204
 205         for input in task.batches(split=split):
 206             input = input.to(device)
 207             output = eval_mygpt(model, input, fixed_len=fixed_len)
 208             loss = F.cross_entropy(output.transpose(1, 2), input)
 209             acc_loss += loss.item() * input.size(0)
 210             nb_samples += input.size(0)
 211
 212         model.train(t)
 213
 214         return math.exp(min(100, acc_loss / nb_samples))
 215
 216
 217 ######################################################################
 218
 219
 220 def oneshot_policy_loss(mazes, output, policies, height, width):
 221     masks = (mazes == maze.v_empty).unsqueeze(-1)
 222     targets = policies.permute(0, 2, 1) * masks
 223     output = output * masks
 224     return -(output.log_softmax(-1) * targets).sum() / masks.sum()
 225
 226
 227 def oneshot_trace_loss(mazes, output, policies, height, width):
 228     masks = mazes == maze.v_empty
 229     targets = maze.stationary_densities(
 230         mazes.view(-1, height, width), policies.view(-1, 4, height, width)
 231     ).flatten(-2)
 232     targets = targets * masks
 233     output = output.squeeze(-1) * masks
 234     return (output - targets).abs().sum() / masks.sum()
 235
 236
 237 def oneshot(gpt, task):
 238     t = gpt.training
 239     gpt.eval()
 240
 241     if args.oneshot_input == "head":
 242         dim_in = args.dim_model
 243     elif args.oneshot_input == "deep":
 244         dim_in = args.dim_model * args.nb_blocks * 2
 245     else:
 246         raise ValueError(f"{args.oneshot_input=}")
 247
 248     if args.oneshot_output == "policy":
 249         dim_out = 4
 250         compute_loss = oneshot_policy_loss
 251     elif args.oneshot_output == "trace":
 252         dim_out = 1
 253         compute_loss = oneshot_trace_loss
 254     else:
 255         raise ValueError(f"{args.oneshot_output=}")
 256
 257     model = nn.Sequential(
 258         nn.Linear(dim_in, args.dim_model),
 259         nn.ReLU(),
 260         nn.Linear(args.dim_model, args.dim_model),
 261         nn.ReLU(),
 262         nn.Linear(args.dim_model, dim_out),
 263     ).to(device)
 264
 265     for n_epoch in range(args.nb_epochs):
 266         learning_rate = learning_rate_schedule[n_epoch]
 267         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 268
 269         acc_train_loss, nb_train_samples = 0, 0
 270         for mazes, policies in task.policy_batches(split="train"):
 271             output_gpt = eval_mygpt(
 272                 gpt, mazes, mode=args.oneshot_input, fixed_len=task.height * task.width
 273             )
 274             output = model(output_gpt)
 275
 276             loss = compute_loss(mazes, output, policies, task.height, task.width)
 277             acc_train_loss += loss.item() * mazes.size(0)
 278             nb_train_samples += mazes.size(0)
 279
 280             optimizer.zero_grad()
 281             loss.backward()
 282             optimizer.step()
 283
 284         acc_test_loss, nb_test_samples = 0, 0
 285         for mazes, policies in task.policy_batches(split="test"):
 286             output_gpt = eval_mygpt(
 287                 gpt, mazes, mode=args.oneshot_input, fixed_len=task.height * task.width
 288             )
 289             output = model(output_gpt)
 290             loss = compute_loss(mazes, output, policies, task.height, task.width)
 291             acc_test_loss += loss.item() * mazes.size(0)
 292             nb_test_samples += mazes.size(0)
 293
 294         log_string(
 295             f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
 296         )
 297
 298         # -------------------
 299         mazes = task.test_input[:32, : task.height * task.width]
 300         policies = task.test_policies[:32]
 301         output_gpt = eval_mygpt(
 302             gpt, mazes, mode=args.oneshot_input, fixed_len=task.height * task.width
 303         )
 304         output = model(output_gpt)
 305         if args.oneshot_output == "policy":
 306             targets = policies.permute(0, 2, 1)
 307             scores = (
 308                 (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
 309             ).float()
 310         elif args.oneshot_output == "trace":
 311             targets = maze.stationary_densities(
 312                 mazes.view(-1, task.height, task.width),
 313                 policies.view(-1, 4, task.height, task.width),
 314             ).flatten(-2)
 315             scores = output
 316         else:
 317             raise ValueError(f"{args.oneshot_output=}")
 318
 319         scores = scores.reshape(-1, task.height, task.width)
 320         mazes = mazes.reshape(-1, task.height, task.width)
 321         targets = targets.reshape(-1, task.height, task.width)
 322         filename = (
 323             f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png"
 324         )
 325         maze.save_image(
 326             os.path.join(args.result_dir, filename),
 327             mazes=mazes,
 328             score_paths=scores,
 329             score_truth=targets,
 330         )
 331         log_string(f"wrote {filename}")
 332
 333         # -------------------
 334
 335     gpt.train(t)
 336
 337
 338 ######################################################################
 339
 340
 341 class Task:
 342     def batches(self, split="train", nb_to_use=-1, desc=None):
 343         pass
 344
 345     def vocabulary_size(self):
 346         pass
 347
 348     def produce_results(self, n_epoch, model):
 349         pass
 350
 351
 352 ######################################################################
 353
 354 import maze
 355
 356
 357 class TaskMaze(Task):
 358     def map2seq(self, *m):
 359         return torch.cat([x.flatten(1) for x in m], 1)
 360
 361     def seq2map(self, s):
 362         s = s.reshape(s.size(0), -1, self.height, self.width)
 363         return (s[:, k] for k in range(s.size(1)))
 364
 365     def __init__(
 366         self,
 367         nb_train_samples,
 368         nb_test_samples,
 369         batch_size,
 370         height,
 371         width,
 372         nb_walls,
 373         device=torch.device("cpu"),
 374     ):
 375         self.batch_size = batch_size
 376         self.height = height
 377         self.width = width
 378         self.device = device
 379
 380         train_mazes, train_paths, train_policies = maze.create_maze_data(
 381             nb_train_samples,
 382             height=height,
 383             width=width,
 384             nb_walls=nb_walls,
 385             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
 386         )
 387         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
 388         self.train_policies = train_policies.flatten(-2).to(device)
 389
 390         test_mazes, test_paths, test_policies = maze.create_maze_data(
 391             nb_test_samples,
 392             height=height,
 393             width=width,
 394             nb_walls=nb_walls,
 395             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
 396         )
 397         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
 398         self.test_policies = test_policies.flatten(-2).to(device)
 399
 400         self.nb_codes = self.train_input.max() + 1
 401
 402     def batches(self, split="train", nb_to_use=-1, desc=None):
 403         assert split in {"train", "test"}
 404         input = self.train_input if split == "train" else self.test_input
 405         if nb_to_use > 0:
 406             input = input[:nb_to_use]
 407         if desc is None:
 408             desc = f"epoch-{split}"
 409         for batch in tqdm.tqdm(
 410             input.split(self.batch_size), dynamic_ncols=True, desc=desc
 411         ):
 412             yield batch
 413
 414     def policy_batches(self, split="train", nb_to_use=-1, desc=None):
 415         assert split in {"train", "test"}
 416         input = self.train_input if split == "train" else self.test_input
 417         policies = self.train_policies if split == "train" else self.test_policies
 418         input = input[:, : self.height * self.width]
 419         policies = policies * (input != maze.v_wall)[:, None]
 420
 421         if nb_to_use > 0:
 422             input = input[:nb_to_use]
 423             policies = policies[:nb_to_use]
 424
 425         if desc is None:
 426             desc = f"epoch-{split}"
 427         for batch in tqdm.tqdm(
 428             zip(input.split(self.batch_size), policies.split(self.batch_size)),
 429             dynamic_ncols=True,
 430             desc=desc,
 431         ):
 432             yield batch
 433
 434     def vocabulary_size(self):
 435         return self.nb_codes
 436
 437     def compute_error(self, model, split="train", nb_to_use=-1):
 438         nb_total, nb_correct = 0, 0
 439         for input in task.batches(split, nb_to_use):
 440             result = input.clone()
 441             ar_mask = result.new_zeros(result.size())
 442             ar_mask[:, self.height * self.width :] = 1
 443             result *= 1 - ar_mask
 444             x, order = shuffle(result, self.height * self.width)
 445             masked_inplace_autoregression(
 446                 model, self.batch_size, x, ar_mask, order=order
 447             )
 448             result = reorder(x, order, reverse=True)
 449             mazes, paths = self.seq2map(result)
 450             nb_correct += maze.path_correctness(mazes, paths).long().sum()
 451             nb_total += mazes.size(0)
 452
 453         return nb_total, nb_correct
 454
 455     def produce_results(self, n_epoch, model):
 456         with torch.autograd.no_grad():
 457             t = model.training
 458             model.eval()
 459
 460             train_nb_total, train_nb_correct = self.compute_error(
 461                 model, "train", nb_to_use=1000
 462             )
 463             log_string(
 464                 f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
 465             )
 466
 467             test_nb_total, test_nb_correct = self.compute_error(
 468                 model, "test", nb_to_use=1000
 469             )
 470             log_string(
 471                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
 472             )
 473
 474             input = self.test_input[:32]
 475             result = input.clone()
 476             ar_mask = result.new_zeros(result.size())
 477             ar_mask[:, self.height * self.width :] = 1
 478             result *= 1 - ar_mask
 479             x, order = shuffle(result, self.height * self.width)
 480             masked_inplace_autoregression(
 481                 model, self.batch_size, x, ar_mask, order=order
 482             )
 483             result = reorder(x, order, reverse=True)
 484
 485             mazes, paths = self.seq2map(input)
 486             _, predicted_paths = self.seq2map(result)
 487             filename = f"result_{n_epoch:04d}.png"
 488             maze.save_image(
 489                 os.path.join(args.result_dir, filename),
 490                 mazes=mazes,
 491                 target_paths=paths,
 492                 predicted_paths=predicted_paths,
 493                 path_correct=maze.path_correctness(mazes, predicted_paths),
 494             )
 495             log_string(f"wrote {filename}")
 496
 497             model.train(t)
 498
 499
 500 ######################################################################
 501
 502 log_string(f"device {device}")
 503
 504
 505 task = TaskMaze(
 506     nb_train_samples=args.nb_train_samples,
 507     nb_test_samples=args.nb_test_samples,
 508     batch_size=args.batch_size,
 509     height=args.maze_height,
 510     width=args.maze_width,
 511     nb_walls=args.maze_nb_walls,
 512     device=device,
 513 )
 514
 515
 516 vocabulary_size = task.vocabulary_size()
 517
 518 log_string(f"vocabulary_size {vocabulary_size}")
 519
 520 ##############################
 521
 522 amm_generator = None
 523
 524 if args.noncausal_prompt:
 525     amm_generator = lambda d: torch.logical_and(
 526         torch.arange(d)[None, None, :, None] < torch.arange(d)[None, None, None, :],
 527         torch.arange(d)[None, None, :, None] >= d // 2,
 528     )
 529
 530 model = mygpt.MyGPT(
 531     vocabulary_size=vocabulary_size,
 532     dim_model=args.dim_model,
 533     dim_keys=args.dim_keys,
 534     dim_hidden=args.dim_hidden,
 535     nb_heads=args.nb_heads,
 536     nb_blocks=args.nb_blocks,
 537     causal=True,
 538     dropout=args.dropout,
 539     amm_generator=amm_generator,
 540 )
 541
 542 model.to(device)
 543
 544 nb_parameters = sum(p.numel() for p in model.parameters())
 545 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 546
 547 ######################################################################
 548
 549 nb_epochs_finished = 0
 550
 551 if args.no_checkpoint:
 552     log_string(f"not trying to load checkpoint.")
 553
 554 else:
 555     try:
 556         checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 557         checkpoint = torch.load(checkpoint_name)
 558         nb_epochs_finished = checkpoint["nb_epochs_finished"]
 559         model.load_state_dict(checkpoint["model_state"])
 560         torch.set_rng_state(checkpoint["rng_state"])
 561         if torch.cuda.is_available():
 562             torch.cuda.set_rng_state(checkpoint["cuda_rng_state"])
 563
 564         log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
 565
 566     except FileNotFoundError:
 567         log_string("starting from scratch.")
 568
 569     except:
 570         log_string("error when loading the checkpoint.")
 571         exit(1)
 572
 573 ######################################################################
 574
 575 token_count = 0
 576 for input in task.batches(split="train"):
 577     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
 578 token_probas = token_count / token_count.sum()
 579 entropy = -torch.xlogy(token_probas, token_probas).sum()
 580 train_set_perplexity = math.exp(entropy)
 581
 582 ##############################
 583
 584 if args.learning_rate_schedule == "cos":
 585     learning_rate_schedule = {}
 586     for n_epoch in range(args.nb_epochs):
 587         u = n_epoch / args.nb_epochs * math.pi
 588         learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
 589 else:
 590     u = {
 591         int(k): float(v)
 592         for k, v in [
 593             tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
 594         ]
 595     }
 596
 597     learning_rate_schedule = {}
 598     learning_rate = args.learning_rate
 599     for n_epoch in range(args.nb_epochs):
 600         if n_epoch in u:
 601             learning_rate = u[n_epoch]
 602         learning_rate_schedule[n_epoch] = learning_rate
 603
 604 log_string(f"learning_rate_schedule {learning_rate_schedule}")
 605
 606 ##############################
 607
 608 if nb_epochs_finished >= args.nb_epochs:
 609     n_epoch = nb_epochs_finished
 610     train_perplexity = compute_perplexity(
 611         model, task, fixed_len=task.height * task.width, split="train"
 612     )
 613     test_perplexity = compute_perplexity(
 614         model, task, fixed_len=task.height * task.width, split="test"
 615     )
 616
 617     log_string(
 618         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 619     )
 620
 621     task.produce_results(n_epoch, model)
 622
 623 ##############################
 624
 625 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
 626     learning_rate = learning_rate_schedule[n_epoch]
 627
 628     log_string(f"learning_rate {learning_rate}")
 629
 630     if args.optim == "sgd":
 631         optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
 632     elif args.optim == "adam":
 633         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 634     elif args.optim == "adamw":
 635         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
 636     else:
 637         raise ValueError(f"{args.optim=}")
 638
 639     model.train()
 640
 641     nb_train_samples, acc_train_loss = 0, 0.0
 642
 643     for input in task.batches(split="train"):
 644         input = input.to(device)
 645         output = eval_mygpt(
 646             model, input, mode=args.oneshot_input, fixed_len=task.height * task.width
 647         )
 648         loss = F.cross_entropy(output.transpose(1, 2), input)
 649         acc_train_loss += loss.item() * input.size(0)
 650         nb_train_samples += input.size(0)
 651
 652         optimizer.zero_grad()
 653         loss.backward()
 654         optimizer.step()
 655
 656     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 657     test_perplexity = compute_perplexity(
 658         model, task, fixed_len=task.height * task.width, split="test"
 659     )
 660
 661     log_string(
 662         f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
 663     )
 664
 665     task.produce_results(n_epoch, model)
 666
 667     checkpoint = {
 668         "nb_epochs_finished": n_epoch + 1,
 669         "model_state": model.state_dict(),
 670         "rng_state": torch.get_rng_state(),
 671     }
 672
 673     if torch.cuda.is_available():
 674         checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
 675
 676     checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
 677     torch.save(checkpoint, checkpoint_name)
 678     log_string(f"saved checkpoint {checkpoint_name}")
 679
 680 ######################################################################
 681
 682 if args.oneshot:
 683     oneshot(model, task)
 684
 685 ######################################################################