IF YOU WOULD LIKE TO GET AN ACCOUNT, please write an email to s dot adaszewski at gmail dot com. User accounts are meant only to report issues and/or generate pull requests. This is a purpose-specific Git hosting for ADARED projects. Thank you for your understanding!
Você não pode selecionar mais de 25 tópicos Os tópicos devem começar com uma letra ou um número, podem incluir traços ('-') e podem ter até 35 caracteres.

101 linhas
3.5KB

  1. from .model import Model
  2. import torch
  3. from .batch import PredictionsBatch, \
  4. flatten_predictions, \
  5. gather_batch_indices
  6. from typing import Callable
  7. from types import FunctionType
  8. import time
  9. class TrainLoop(object):
  10. def __init__(
  11. self,
  12. model: Model,
  13. lr: float = 0.001,
  14. loss: Callable[[torch.Tensor, torch.Tensor], torch.Tensor] = \
  15. torch.nn.functional.binary_cross_entropy_with_logits,
  16. batch_size: int = 100,
  17. shuffle: bool = False,
  18. generator: torch.Generator = None) -> None:
  19. if not isinstance(model, Model):
  20. raise TypeError('model must be an instance of Model')
  21. lr = float(lr)
  22. if not isinstance(loss, FunctionType):
  23. raise TypeError('loss must be a function')
  24. batch_size = int(batch_size)
  25. if generator is not None and not isinstance(generator, torch.Generator):
  26. raise TypeError('generator must be an instance of torch.Generator')
  27. self.model = model
  28. self.lr = lr
  29. self.loss = loss
  30. self.batch_size = batch_size
  31. self.shuffle = shuffle
  32. self.generator = generator or torch.default_generator
  33. self.opt = None
  34. self.build()
  35. def build(self) -> None:
  36. opt = torch.optim.Adam(self.model.parameters(), lr=self.lr)
  37. self.opt = opt
  38. def run_epoch(self):
  39. batch = PredictionsBatch(self.model.prep_d, batch_size=self.batch_size,
  40. shuffle = self.shuffle, generator=self.generator)
  41. # pred = self.model(None)
  42. # n = len(list(iter(batch)))
  43. loss_sum = 0
  44. for i, indices in enumerate(batch):
  45. print('%.2f%% (%d/%d)' % (i * batch.batch_size * 100 / batch.total_edge_count, i * batch.batch_size, batch.total_edge_count))
  46. t = time.time()
  47. self.opt.zero_grad()
  48. print('zero_grad() took:', time.time() - t)
  49. t = time.time()
  50. pred = self.model(None)
  51. print('model() took:', time.time() - t)
  52. t = time.time()
  53. pred = flatten_predictions(pred)
  54. print('flatten_predictions() took:', time.time() - t)
  55. # batch = PredictionsBatch(pred, batch_size=self.batch_size, shuffle=True)
  56. # seed = torch.rand(1).item()
  57. # rng_state = torch.get_rng_state()
  58. # torch.manual_seed(seed)
  59. #it = iter(batch)
  60. #torch.set_rng_state(rng_state)
  61. #for k in range(i):
  62. #_ = next(it)
  63. #(input, target) = next(it)
  64. t = time.time()
  65. (input, target) = gather_batch_indices(pred, indices)
  66. print('gather_batch_indices() took:', time.time() - t)
  67. t = time.time()
  68. loss = self.loss(input, target)
  69. print('loss() took:', time.time() - t)
  70. t = time.time()
  71. loss.backward()
  72. print('backward() took:', time.time() - t)
  73. t = time.time()
  74. self.opt.step()
  75. print('step() took:', time.time() - t)
  76. loss_sum += loss.detach().cpu().item()
  77. return loss_sum
  78. def train(self, max_epochs):
  79. best_loss = None
  80. best_epoch = None
  81. for i in range(max_epochs):
  82. loss = self.run_epoch()
  83. if best_loss is None or loss < best_loss:
  84. best_loss = loss
  85. best_epoch = i
  86. return loss, best_loss, best_epoch