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.

185 Zeilen
5.0KB

  1. from icosagon.data import Data, \
  2. _equal
  3. from icosagon.trainprep import prepare_training, \
  4. TrainValTest
  5. from icosagon.model import Model
  6. from icosagon.trainloop import TrainLoop
  7. import torch
  8. import pytest
  9. import pdb
  10. import time
  11. def test_train_loop_01():
  12. d = Data()
  13. d.add_node_type('Dummy', 10)
  14. fam = d.add_relation_family('Dummy-Dummy', 0, 0, False)
  15. fam.add_relation_type('Dummy Rel', torch.rand(10, 10).round())
  16. prep_d = prepare_training(d, TrainValTest(.8, .1, .1))
  17. m = Model(prep_d)
  18. loop = TrainLoop(m)
  19. assert loop.model == m
  20. assert loop.lr == 0.001
  21. assert loop.loss == torch.nn.functional.binary_cross_entropy_with_logits
  22. assert loop.batch_size == 100
  23. def test_train_loop_02():
  24. d = Data()
  25. d.add_node_type('Dummy', 10)
  26. fam = d.add_relation_family('Dummy-Dummy', 0, 0, False)
  27. fam.add_relation_type('Dummy Rel', torch.rand(10, 10).round())
  28. prep_d = prepare_training(d, TrainValTest(.8, .1, .1))
  29. m = Model(prep_d)
  30. for prm in m.parameters():
  31. print(prm.shape, prm.is_leaf, prm.requires_grad)
  32. loop = TrainLoop(m)
  33. loop.run_epoch()
  34. for prm in m.parameters():
  35. print(prm.shape, prm.is_leaf, prm.requires_grad)
  36. def test_train_loop_03():
  37. # pdb.set_trace()
  38. if torch.cuda.device_count() == 0:
  39. pytest.skip('CUDA required for this test')
  40. adj_mat = torch.rand(10, 10).round()
  41. dev = torch.device('cuda:0')
  42. adj_mat = adj_mat.to(dev)
  43. d = Data()
  44. d.add_node_type('Dummy', 10)
  45. fam = d.add_relation_family('Dummy-Dummy', 0, 0, False)
  46. fam.add_relation_type('Dummy Rel', adj_mat)
  47. prep_d = prepare_training(d, TrainValTest(.8, .1, .1))
  48. # pdb.set_trace()
  49. m = Model(prep_d)
  50. m = m.to(dev)
  51. print(list(m.parameters()))
  52. for prm in m.parameters():
  53. assert prm.device == dev
  54. loop = TrainLoop(m)
  55. loop.run_epoch()
  56. def test_train_loop_04():
  57. adj_mat = torch.rand(10, 10).round()
  58. d = Data()
  59. d.add_node_type('Dummy', 10)
  60. fam = d.add_relation_family('Dummy-Dummy', 0, 0, False)
  61. fam.add_relation_type('Dummy Rel', adj_mat)
  62. prep_d = prepare_training(d, TrainValTest(.8, .1, .1))
  63. m = Model(prep_d)
  64. old_values = []
  65. for prm in m.parameters():
  66. old_values.append(prm.clone().detach())
  67. loop = TrainLoop(m)
  68. loop.run_epoch()
  69. for i, prm in enumerate(m.parameters()):
  70. assert not prm.requires_grad or \
  71. not torch.all(_equal(prm, old_values[i]))
  72. def test_train_loop_05():
  73. adj_mat = torch.rand(10, 10).round().to_sparse()
  74. d = Data()
  75. d.add_node_type('Dummy', 10)
  76. fam = d.add_relation_family('Dummy-Dummy', 0, 0, False)
  77. fam.add_relation_type('Dummy Rel', adj_mat)
  78. prep_d = prepare_training(d, TrainValTest(.8, .1, .1))
  79. m = Model(prep_d)
  80. old_values = []
  81. for prm in m.parameters():
  82. old_values.append(prm.clone().detach())
  83. loop = TrainLoop(m)
  84. loop.run_epoch()
  85. for i, prm in enumerate(m.parameters()):
  86. assert not prm.requires_grad or \
  87. not torch.all(_equal(prm, old_values[i]))
  88. def test_timing_01():
  89. adj_mat = (torch.rand(2000, 2000) < .001).to(torch.float32).to_sparse()
  90. rep = torch.eye(2000).requires_grad_(True)
  91. t = time.time()
  92. for _ in range(1300):
  93. _ = torch.sparse.mm(adj_mat, rep)
  94. print('Elapsed:', time.time() - t)
  95. def test_timing_02():
  96. adj_mat = (torch.rand(2000, 2000) < .001).to(torch.float32)
  97. adj_mat_batch = [adj_mat.view(1, 2000, 2000)] * 1300
  98. adj_mat_batch = torch.cat(adj_mat_batch)
  99. rep = torch.eye(2000).requires_grad_(True)
  100. t = time.time()
  101. res = torch.matmul(adj_mat_batch, rep)
  102. print('Elapsed:', time.time() - t)
  103. print('res.shape:', res.shape)
  104. def test_timing_03():
  105. adj_mat = (torch.rand(2000, 2000) < .001).to(torch.float32)
  106. adj_mat_batch = [adj_mat.view(1, 2000, 2000).to_sparse()] * 1300
  107. adj_mat_batch = torch.cat(adj_mat_batch)
  108. rep = torch.eye(2000).requires_grad_(True)
  109. rep_batch = [rep.view(1, 2000, 2000)] * 1300
  110. rep_batch = torch.cat(rep_batch)
  111. t = time.time()
  112. with pytest.raises(RuntimeError):
  113. _ = torch.bmm(adj_mat_batch, rep)
  114. print('Elapsed:', time.time() - t)
  115. def test_timing_04():
  116. adj_mat = (torch.rand(2000, 2000) < .0001).to(torch.float32).to_sparse()
  117. rep = torch.eye(2000).requires_grad_(True)
  118. t = time.time()
  119. for _ in range(1300):
  120. _ = torch.sparse.mm(adj_mat, rep)
  121. print('Elapsed:', time.time() - t)
  122. def test_timing_05():
  123. if torch.cuda.device_count() == 0:
  124. pytest.skip('Test requires CUDA')
  125. dev = torch.device('cuda:0')
  126. adj_mat = (torch.rand(2000, 2000) < .001).to(torch.float32).to_sparse().to(dev)
  127. rep = torch.eye(2000).requires_grad_(True).to(dev)
  128. t = time.time()
  129. for _ in range(1300):
  130. _ = torch.sparse.mm(adj_mat, rep)
  131. torch.cuda.synchronize()
  132. print('Elapsed:', time.time() - t)