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!
Du kan inte välja fler än 25 ämnen Ämnen måste starta med en bokstav eller siffra, kan innehålla bindestreck ('-') och vara max 35 tecken långa.

59 lines
2.0KB

  1. #
  2. # Copyright (C) Stanislaw Adaszewski, 2020
  3. # License: GPLv3
  4. #
  5. import torch
  6. from .dropout import dropout
  7. from .weights import init_glorot
  8. from typing import List, Callable
  9. import pdb
  10. class GraphConv(torch.nn.Module):
  11. def __init__(self, in_channels: int, out_channels: int,
  12. adjacency_matrix: torch.Tensor, **kwargs) -> None:
  13. super().__init__(**kwargs)
  14. self.in_channels = in_channels
  15. self.out_channels = out_channels
  16. self.weight = torch.nn.Parameter(init_glorot(in_channels, out_channels))
  17. self.adjacency_matrix = adjacency_matrix
  18. def forward(self, x: torch.Tensor) -> torch.Tensor:
  19. x = torch.sparse.mm(x, self.weight) \
  20. if x.is_sparse \
  21. else torch.mm(x, self.weight)
  22. x = torch.sparse.mm(self.adjacency_matrix, x) \
  23. if self.adjacency_matrix.is_sparse \
  24. else torch.mm(self.adjacency_matrix, x)
  25. return x
  26. class DropoutGraphConvActivation(torch.nn.Module):
  27. def __init__(self, input_dim: int, output_dim: int,
  28. adjacency_matrix: torch.Tensor, keep_prob: float=1.,
  29. activation: Callable[[torch.Tensor], torch.Tensor]=torch.nn.functional.relu,
  30. **kwargs) -> None:
  31. super().__init__(**kwargs)
  32. self.input_dim = input_dim
  33. self.output_dim = output_dim
  34. self.adjacency_matrix = adjacency_matrix
  35. self.keep_prob = keep_prob
  36. self.activation = activation
  37. self.graph_conv = GraphConv(input_dim, output_dim, adjacency_matrix)
  38. def forward(self, x: torch.Tensor) -> torch.Tensor:
  39. # pdb.set_trace()
  40. x = dropout(x, self.keep_prob)
  41. x = self.graph_conv(x)
  42. x = self.activation(x)
  43. return x
  44. def clone(self, adjacency_matrix) -> 'DropoutGraphConvActivation':
  45. res = DropoutGraphConvActivation(self.input_dim,
  46. self.output_dim, adjacency_matrix, self.keep_prob,
  47. self.activation)
  48. res.graph_conv.weight = self.graph_conv.weight
  49. return res