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.

235 linhas
9.3KB

  1. #
  2. # Copyright (C) Stanislaw Adaszewski, 2020
  3. # License: GPLv3
  4. #
  5. from icosagon.input import OneHotInputLayer
  6. from icosagon.convolve import DropoutGraphConvActivation
  7. from icosagon.convlayer import DecagonLayer
  8. from icosagon.declayer import DecodeLayer, \
  9. Predictions, \
  10. RelationFamilyPredictions, \
  11. RelationPredictions
  12. from icosagon.decode import DEDICOMDecoder, \
  13. InnerProductDecoder
  14. from icosagon.data import Data
  15. from icosagon.trainprep import prepare_training, \
  16. TrainValTest
  17. import torch
  18. def test_decode_layer_01():
  19. d = Data()
  20. d.add_node_type('Dummy', 100)
  21. fam = d.add_relation_family('Dummy-Dummy', 0, 0, False)
  22. fam.add_relation_type('Dummy Relation 1',
  23. torch.rand((100, 100), dtype=torch.float32).round().to_sparse())
  24. prep_d = prepare_training(d, TrainValTest(.8, .1, .1))
  25. in_layer = OneHotInputLayer(d)
  26. d_layer = DecagonLayer(in_layer.output_dim, 32, d)
  27. seq = torch.nn.Sequential(in_layer, d_layer)
  28. last_layer_repr = seq(None)
  29. dec = DecodeLayer(input_dim=d_layer.output_dim, data=prep_d, keep_prob=1.,
  30. activation=lambda x: x)
  31. pred = dec(last_layer_repr)
  32. assert isinstance(pred, Predictions)
  33. assert isinstance(pred.relation_families, list)
  34. assert len(pred.relation_families) == 1
  35. assert isinstance(pred.relation_families[0], RelationFamilyPredictions)
  36. assert isinstance(pred.relation_families[0].relation_types, list)
  37. assert len(pred.relation_families[0].relation_types) == 1
  38. assert isinstance(pred.relation_families[0].relation_types[0], RelationPredictions)
  39. tmp = pred.relation_families[0].relation_types[0]
  40. assert isinstance(tmp.edges_pos, TrainValTest)
  41. assert isinstance(tmp.edges_neg, TrainValTest)
  42. assert isinstance(tmp.edges_back_pos, TrainValTest)
  43. assert isinstance(tmp.edges_back_neg, TrainValTest)
  44. def test_decode_layer_02():
  45. d = Data()
  46. d.add_node_type('Dummy', 100)
  47. fam = d.add_relation_family('Dummy-Dummy', 0, 0, False)
  48. fam.add_relation_type('Dummy Relation 1',
  49. torch.rand((100, 100), dtype=torch.float32).round().to_sparse())
  50. prep_d = prepare_training(d, TrainValTest(.8, .1, .1))
  51. in_layer = OneHotInputLayer(d)
  52. d_layer = DecagonLayer(in_layer.output_dim, 32, d)
  53. dec_layer = DecodeLayer(input_dim=d_layer.output_dim, data=prep_d,
  54. keep_prob=1., activation=lambda x: x)
  55. seq = torch.nn.Sequential(in_layer, d_layer, dec_layer)
  56. pred = seq(None)
  57. assert isinstance(pred, Predictions)
  58. assert len(pred.relation_families) == 1
  59. assert isinstance(pred.relation_families[0], RelationFamilyPredictions)
  60. assert isinstance(pred.relation_families[0].relation_types, list)
  61. assert len(pred.relation_families[0].relation_types) == 1
  62. def test_decode_layer_03():
  63. d = Data()
  64. d.add_node_type('Dummy 1', 100)
  65. d.add_node_type('Dummy 2', 100)
  66. fam = d.add_relation_family('Dummy 1-Dummy 2', 0, 1, True)
  67. fam.add_relation_type('Dummy Relation 1',
  68. torch.rand((100, 100), dtype=torch.float32).round().to_sparse())
  69. prep_d = prepare_training(d, TrainValTest(.8, .1, .1))
  70. in_layer = OneHotInputLayer(d)
  71. d_layer = DecagonLayer(in_layer.output_dim, 32, d)
  72. dec_layer = DecodeLayer(input_dim=d_layer.output_dim, data=prep_d,
  73. keep_prob=1., activation=lambda x: x)
  74. seq = torch.nn.Sequential(in_layer, d_layer, dec_layer)
  75. pred = seq(None)
  76. assert isinstance(pred, Predictions)
  77. assert len(pred.relation_families) == 1
  78. assert isinstance(pred.relation_families[0], RelationFamilyPredictions)
  79. assert isinstance(pred.relation_families[0].relation_types, list)
  80. assert len(pred.relation_families[0].relation_types) == 1
  81. assert isinstance(pred.relation_families[0].relation_types[0], RelationPredictions)
  82. def test_decode_layer_04():
  83. d = Data()
  84. d.add_node_type('Dummy', 100)
  85. assert len(d.relation_families) == 0
  86. prep_d = prepare_training(d, TrainValTest(.8, .1, .1))
  87. in_layer = OneHotInputLayer(d)
  88. d_layer = DecagonLayer(in_layer.output_dim, 32, d)
  89. dec_layer = DecodeLayer(input_dim=d_layer.output_dim, data=prep_d,
  90. keep_prob=1., activation=lambda x: x)
  91. seq = torch.nn.Sequential(in_layer, d_layer, dec_layer)
  92. pred = seq(None)
  93. assert isinstance(pred, Predictions)
  94. assert len(pred.relation_families) == 0
  95. def test_decode_layer_05():
  96. d = Data()
  97. d.add_node_type('Dummy', 10)
  98. mat = torch.rand((10, 10))
  99. mat = (mat + mat.transpose(0, 1)) / 2
  100. mat = mat.round()
  101. fam = d.add_relation_family('Dummy-Dummy', 0, 0, True,
  102. decoder_class=InnerProductDecoder)
  103. fam.add_relation_type('Dummy Rel', mat.to_sparse())
  104. prep_d = prepare_training(d, TrainValTest(1., 0., 0.))
  105. in_layer = OneHotInputLayer(d)
  106. conv_layer = DecagonLayer(in_layer.output_dim, 32, prep_d,
  107. rel_activation=lambda x: x, layer_activation=lambda x: x)
  108. dec_layer = DecodeLayer(conv_layer.output_dim, prep_d,
  109. keep_prob=1., activation=lambda x: x)
  110. seq = torch.nn.Sequential(in_layer, conv_layer, dec_layer)
  111. pred = seq(None)
  112. rel_pred = pred.relation_families[0].relation_types[0]
  113. for edge_type in ['edges_pos', 'edges_neg', 'edges_back_pos', 'edges_back_neg']:
  114. edge_pred = getattr(rel_pred, edge_type)
  115. assert isinstance(edge_pred, TrainValTest)
  116. for part_type in ['train', 'val', 'test']:
  117. part_pred = getattr(edge_pred, part_type)
  118. assert isinstance(part_pred, torch.Tensor)
  119. assert len(part_pred.shape) == 1
  120. print(edge_type, part_type, part_pred.shape)
  121. if (edge_type, part_type) not in [('edges_pos', 'train'), ('edges_neg', 'train')]:
  122. assert part_pred.shape[0] == 0
  123. else:
  124. assert part_pred.shape[0] > 0
  125. prep_rel = prep_d.relation_families[0].relation_types[0]
  126. assert len(rel_pred.edges_pos.train) == len(prep_rel.edges_pos.train)
  127. assert len(rel_pred.edges_neg.train) == len(prep_rel.edges_neg.train)
  128. assert len(prep_rel.edges_pos.train) == torch.sum(mat)
  129. # print('Predictions for positive edges:')
  130. # print(rel_pred.edges_pos.train)
  131. # print('Predictions for negative edges:')
  132. # print(rel_pred.edges_neg.train)
  133. repr_in = in_layer(None)
  134. assert isinstance(repr_in, list)
  135. assert len(repr_in) == 1
  136. assert isinstance(repr_in[0], torch.Tensor)
  137. assert torch.all(repr_in[0].to_dense() == torch.eye(10))
  138. assert len(conv_layer.next_layer_repr[0]) == 1
  139. assert len(conv_layer.next_layer_repr[0][0].convolutions) == 1
  140. assert conv_layer.rel_activation(0) == 0
  141. assert conv_layer.rel_activation(1) == 1
  142. assert conv_layer.rel_activation(-1) == -1
  143. assert conv_layer.layer_activation(0) == 0
  144. assert conv_layer.layer_activation(1) == 1
  145. assert conv_layer.layer_activation(-1) == -1
  146. graph_conv = conv_layer.next_layer_repr[0][0].convolutions[0]
  147. assert isinstance(graph_conv, DropoutGraphConvActivation)
  148. assert graph_conv.activation(0) == 0
  149. assert graph_conv.activation(1) == 1
  150. assert graph_conv.activation(-1) == -1
  151. weight = graph_conv.graph_conv.weight
  152. adj_mat = prep_d.relation_families[0].relation_types[0].adjacency_matrix
  153. repr_conv = torch.sparse.mm(repr_in[0], weight)
  154. repr_conv = torch.mm(adj_mat, repr_conv)
  155. repr_conv = torch.nn.functional.normalize(repr_conv, p=2, dim=1)
  156. repr_conv_expect = conv_layer(repr_in)[0]
  157. print('repr_conv:\n', repr_conv)
  158. # print(repr_conv_expect)
  159. assert torch.all(repr_conv == repr_conv_expect)
  160. assert repr_conv.shape[1] == 32
  161. dec = InnerProductDecoder(32, 1, keep_prob=1., activation=lambda x: x)
  162. x, y = torch.meshgrid(torch.arange(0, 10), torch.arange(0, 10))
  163. x = x.flatten()
  164. y = y.flatten()
  165. repr_dec_expect = dec(repr_conv[x], repr_conv[y], 0)
  166. repr_dec_expect = repr_dec_expect.view(10, 10)
  167. repr_dec = torch.mm(repr_conv, torch.transpose(repr_conv, 0, 1))
  168. # repr_dec = torch.flatten(repr_dec)
  169. # repr_dec -= torch.eye(10)
  170. assert torch.all(torch.abs(repr_dec - repr_dec_expect) < 0.000001)
  171. repr_dec_expect = torch.zeros((10, 10))
  172. x = prep_d.relation_families[0].relation_types[0].edges_pos.train
  173. repr_dec_expect[x[:, 0], x[:, 1]] = pred.relation_families[0].relation_types[0].edges_pos.train
  174. x = prep_d.relation_families[0].relation_types[0].edges_neg.train
  175. repr_dec_expect[x[:, 0], x[:, 1]] = pred.relation_families[0].relation_types[0].edges_neg.train
  176. print(repr_dec)
  177. print(repr_dec_expect)
  178. repr_dec = torch.zeros((10, 10))
  179. x = prep_d.relation_families[0].relation_types[0].edges_pos.train
  180. repr_dec[x[:, 0], x[:, 1]] = dec(repr_conv[x[:, 0]], repr_conv[x[:, 1]], 0)
  181. x = prep_d.relation_families[0].relation_types[0].edges_neg.train
  182. repr_dec[x[:, 0], x[:, 1]] = dec(repr_conv[x[:, 0]], repr_conv[x[:, 1]], 0)
  183. assert torch.all(torch.abs(repr_dec - repr_dec_expect) < 0.000001)
  184. #print(prep_rel.edges_pos.train)
  185. #print(prep_rel.edges_neg.train)
  186. # assert isinstance(edge_pred.train)
  187. # assert isinstance(rel_pred.edges_pos, TrainValTest)
  188. # assert isinstance(rel_pred.edges_neg, TrainValTest)
  189. # assert isinstance(rel_pred.edges_back_pos, TrainValTest)
  190. # assert isinstance(rel_pred.edges_back_neg, TrainValTest)