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 kannst nicht mehr als 25 Themen auswählen Themen müssen entweder mit einem Buchstaben oder einer Ziffer beginnen. Sie können Bindestriche („-“) enthalten und bis zu 35 Zeichen lang sein.

test_convolve.py 10KB

vor 4 Jahren
vor 4 Jahren
vor 4 Jahren
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295
  1. import decagon_pytorch.convolve
  2. import decagon.deep.layers
  3. import torch
  4. import tensorflow as tf
  5. import numpy as np
  6. def prepare_data():
  7. np.random.seed(0)
  8. latent = np.random.random((5, 10)).astype(np.float32)
  9. latent[latent < .5] = 0
  10. latent = np.ceil(latent)
  11. adjacency_matrices = []
  12. for _ in range(5):
  13. adj_mat = np.random.random((len(latent),) * 2).astype(np.float32)
  14. adj_mat[adj_mat < .5] = 0
  15. adj_mat = np.ceil(adj_mat)
  16. adjacency_matrices.append(adj_mat)
  17. print('latent:', latent)
  18. print('adjacency_matrices[0]:', adjacency_matrices[0])
  19. return latent, adjacency_matrices
  20. def dense_to_sparse_tf(x):
  21. a, b = np.where(x)
  22. indices = np.array([a, b]).T
  23. values = x[a, b]
  24. return tf.sparse.SparseTensor(indices, values, x.shape)
  25. def dropout_sparse_tf(x, keep_prob, num_nonzero_elems):
  26. """Dropout for sparse tensors. Currently fails for very large sparse tensors (>1M elements)
  27. """
  28. noise_shape = [num_nonzero_elems]
  29. random_tensor = keep_prob
  30. random_tensor += tf.convert_to_tensor(torch.rand(noise_shape).detach().numpy())
  31. # tf.convert_to_tensor(np.random.random(noise_shape))
  32. # tf.random_uniform(noise_shape)
  33. dropout_mask = tf.cast(tf.floor(random_tensor), dtype=tf.bool)
  34. pre_out = tf.sparse_retain(x, dropout_mask)
  35. return pre_out * (1./keep_prob)
  36. def dense_graph_conv_torch():
  37. torch.random.manual_seed(0)
  38. latent, adjacency_matrices = prepare_data()
  39. latent = torch.tensor(latent)
  40. adj_mat = adjacency_matrices[0]
  41. adj_mat = torch.tensor(adj_mat)
  42. conv = decagon_pytorch.convolve.DenseGraphConv(10, 10,
  43. adj_mat)
  44. latent = conv(latent)
  45. return latent
  46. def dense_dropout_graph_conv_activation_torch(keep_prob=1.):
  47. torch.random.manual_seed(0)
  48. latent, adjacency_matrices = prepare_data()
  49. latent = torch.tensor(latent)
  50. adj_mat = adjacency_matrices[0]
  51. adj_mat = torch.tensor(adj_mat)
  52. conv = decagon_pytorch.convolve.DenseDropoutGraphConvActivation(10, 10,
  53. adj_mat, keep_prob=keep_prob)
  54. latent = conv(latent)
  55. return latent
  56. def sparse_graph_conv_torch():
  57. torch.random.manual_seed(0)
  58. latent, adjacency_matrices = prepare_data()
  59. print('latent.dtype:', latent.dtype)
  60. latent = torch.tensor(latent).to_sparse()
  61. adj_mat = adjacency_matrices[0]
  62. adj_mat = torch.tensor(adj_mat).to_sparse()
  63. print('adj_mat.dtype:', adj_mat.dtype,
  64. 'latent.dtype:', latent.dtype)
  65. conv = decagon_pytorch.convolve.SparseGraphConv(10, 10,
  66. adj_mat)
  67. latent = conv(latent)
  68. return latent
  69. def sparse_graph_conv_tf():
  70. torch.random.manual_seed(0)
  71. latent, adjacency_matrices = prepare_data()
  72. conv_torch = decagon_pytorch.convolve.SparseGraphConv(10, 10,
  73. torch.tensor(adjacency_matrices[0]).to_sparse())
  74. weight = tf.constant(conv_torch.weight.detach().numpy())
  75. latent = dense_to_sparse_tf(latent)
  76. adj_mat = dense_to_sparse_tf(adjacency_matrices[0])
  77. latent = tf.sparse_tensor_dense_matmul(latent, weight)
  78. latent = tf.sparse_tensor_dense_matmul(adj_mat, latent)
  79. return latent
  80. def sparse_dropout_graph_conv_activation_torch(keep_prob=1.):
  81. torch.random.manual_seed(0)
  82. latent, adjacency_matrices = prepare_data()
  83. latent = torch.tensor(latent).to_sparse()
  84. adj_mat = adjacency_matrices[0]
  85. adj_mat = torch.tensor(adj_mat).to_sparse()
  86. conv = decagon_pytorch.convolve.SparseDropoutGraphConvActivation(10, 10,
  87. adj_mat, keep_prob=keep_prob)
  88. latent = conv(latent)
  89. return latent
  90. def sparse_dropout_graph_conv_activation_tf(keep_prob=1.):
  91. torch.random.manual_seed(0)
  92. latent, adjacency_matrices = prepare_data()
  93. conv_torch = decagon_pytorch.convolve.SparseGraphConv(10, 10,
  94. torch.tensor(adjacency_matrices[0]).to_sparse())
  95. weight = tf.constant(conv_torch.weight.detach().numpy())
  96. nonzero_feat = np.sum(latent > 0)
  97. latent = dense_to_sparse_tf(latent)
  98. latent = dropout_sparse_tf(latent, keep_prob,
  99. nonzero_feat)
  100. adj_mat = dense_to_sparse_tf(adjacency_matrices[0])
  101. latent = tf.sparse_tensor_dense_matmul(latent, weight)
  102. latent = tf.sparse_tensor_dense_matmul(adj_mat, latent)
  103. latent = tf.nn.relu(latent)
  104. return latent
  105. def test_sparse_graph_conv():
  106. latent_torch = sparse_graph_conv_torch()
  107. latent_tf = sparse_graph_conv_tf()
  108. assert np.all(latent_torch.detach().numpy() == latent_tf.eval(session = tf.Session()))
  109. def test_sparse_dropout_graph_conv_activation():
  110. for i in range(11):
  111. keep_prob = i/10. + np.finfo(np.float32).eps
  112. latent_torch = sparse_dropout_graph_conv_activation_torch(keep_prob)
  113. latent_tf = sparse_dropout_graph_conv_activation_tf(keep_prob)
  114. latent_torch = latent_torch.detach().numpy()
  115. latent_tf = latent_tf.eval(session = tf.Session())
  116. print('latent_torch:', latent_torch)
  117. print('latent_tf:', latent_tf)
  118. assert np.all(latent_torch - latent_tf < .000001)
  119. def test_sparse_multi_dgca():
  120. latent_torch = None
  121. latent_tf = []
  122. for i in range(11):
  123. keep_prob = i/10. + np.finfo(np.float32).eps
  124. latent_torch = sparse_dropout_graph_conv_activation_torch(keep_prob) \
  125. if latent_torch is None \
  126. else latent_torch + sparse_dropout_graph_conv_activation_torch(keep_prob)
  127. latent_tf.append(sparse_dropout_graph_conv_activation_tf(keep_prob))
  128. latent_torch = torch.nn.functional.normalize(latent_torch, p=2, dim=1)
  129. latent_tf = tf.add_n(latent_tf)
  130. latent_tf = tf.nn.l2_normalize(latent_tf, dim=1)
  131. latent_torch = latent_torch.detach().numpy()
  132. latent_tf = latent_tf.eval(session = tf.Session())
  133. assert np.all(latent_torch - latent_tf < .000001)
  134. def test_graph_conv():
  135. latent_dense = dense_graph_conv_torch()
  136. latent_sparse = sparse_graph_conv_torch()
  137. assert np.all(latent_dense.detach().numpy() == latent_sparse.detach().numpy())
  138. # def setup_function(fun):
  139. # if fun == test_dropout_graph_conv_activation or \
  140. # fun == test_multi_dgca:
  141. # print('Disabling dropout for testing...')
  142. # setup_function.old_dropout = decagon_pytorch.convolve.dropout, \
  143. # decagon_pytorch.convolve.dropout_sparse
  144. #
  145. # decagon_pytorch.convolve.dropout = lambda x, keep_prob: x
  146. # decagon_pytorch.convolve.dropout_sparse = lambda x, keep_prob: x
  147. #
  148. #
  149. # def teardown_function(fun):
  150. # print('Re-enabling dropout...')
  151. # if fun == test_dropout_graph_conv_activation or \
  152. # fun == test_multi_dgca:
  153. # decagon_pytorch.convolve.dropout, \
  154. # decagon_pytorch.convolve.dropout_sparse = \
  155. # setup_function.old_dropout
  156. def flexible_dropout_graph_conv_activation_torch(keep_prob=1.):
  157. torch.random.manual_seed(0)
  158. latent, adjacency_matrices = prepare_data()
  159. latent = torch.tensor(latent).to_sparse()
  160. adj_mat = adjacency_matrices[0]
  161. adj_mat = torch.tensor(adj_mat).to_sparse()
  162. conv = decagon_pytorch.convolve.DropoutGraphConvActivation(10, 10,
  163. adj_mat, keep_prob=keep_prob)
  164. latent = conv(latent)
  165. return latent
  166. def _disable_dropout(monkeypatch):
  167. monkeypatch.setattr(decagon_pytorch.convolve.dense, 'dropout',
  168. lambda x, keep_prob: x)
  169. monkeypatch.setattr(decagon_pytorch.convolve.sparse, 'dropout_sparse',
  170. lambda x, keep_prob: x)
  171. monkeypatch.setattr(decagon_pytorch.convolve.universal, 'dropout',
  172. lambda x, keep_prob: x)
  173. monkeypatch.setattr(decagon_pytorch.convolve.universal, 'dropout_sparse',
  174. lambda x, keep_prob: x)
  175. def test_dropout_graph_conv_activation(monkeypatch):
  176. _disable_dropout(monkeypatch)
  177. for i in range(11):
  178. keep_prob = i/10.
  179. if keep_prob == 0:
  180. keep_prob += np.finfo(np.float32).eps
  181. print('keep_prob:', keep_prob)
  182. latent_dense = dense_dropout_graph_conv_activation_torch(keep_prob)
  183. latent_dense = latent_dense.detach().numpy()
  184. print('latent_dense:', latent_dense)
  185. latent_sparse = sparse_dropout_graph_conv_activation_torch(keep_prob)
  186. latent_sparse = latent_sparse.detach().numpy()
  187. print('latent_sparse:', latent_sparse)
  188. latent_flex = flexible_dropout_graph_conv_activation_torch(keep_prob)
  189. latent_flex = latent_flex.detach().numpy()
  190. print('latent_flex:', latent_flex)
  191. nonzero = (latent_dense != 0) & (latent_sparse != 0)
  192. assert np.all(latent_dense[nonzero] == latent_sparse[nonzero])
  193. nonzero = (latent_dense != 0) & (latent_flex != 0)
  194. assert np.all(latent_dense[nonzero] == latent_flex[nonzero])
  195. nonzero = (latent_sparse != 0) & (latent_flex != 0)
  196. assert np.all(latent_sparse[nonzero] == latent_flex[nonzero])
  197. def test_multi_dgca(monkeypatch):
  198. _disable_dropout(monkeypatch)
  199. keep_prob = .5
  200. torch.random.manual_seed(0)
  201. latent, adjacency_matrices = prepare_data()
  202. latent_sparse = torch.tensor(latent).to_sparse()
  203. latent = torch.tensor(latent)
  204. assert np.all(latent_sparse.to_dense().numpy() == latent.numpy())
  205. adjacency_matrices_sparse = [ torch.tensor(a).to_sparse() for a in adjacency_matrices ]
  206. adjacency_matrices = [ torch.tensor(a) for a in adjacency_matrices ]
  207. for i in range(len(adjacency_matrices)):
  208. assert np.all(adjacency_matrices[i].numpy() == adjacency_matrices_sparse[i].to_dense().numpy())
  209. torch.random.manual_seed(0)
  210. multi_sparse = decagon_pytorch.convolve.SparseMultiDGCA([10,] * len(adjacency_matrices), 10, adjacency_matrices_sparse, keep_prob=keep_prob)
  211. torch.random.manual_seed(0)
  212. multi = decagon_pytorch.convolve.DenseMultiDGCA([10,] * len(adjacency_matrices), 10, adjacency_matrices, keep_prob=keep_prob)
  213. print('len(adjacency_matrices):', len(adjacency_matrices))
  214. print('len(multi_sparse.sparse_dgca):', len(multi_sparse.sparse_dgca))
  215. print('len(multi.dgca):', len(multi.dgca))
  216. for i in range(len(adjacency_matrices)):
  217. assert np.all(multi_sparse.sparse_dgca[i].sparse_graph_conv.weight.detach().numpy() == multi.dgca[i].graph_conv.weight.detach().numpy())
  218. # torch.random.manual_seed(0)
  219. latent_sparse = multi_sparse([latent_sparse,] * len(adjacency_matrices))
  220. # torch.random.manual_seed(0)
  221. latent = multi([latent,] * len(adjacency_matrices))
  222. assert np.all(latent_sparse.detach().numpy() == latent.detach().numpy())