二分类损失

  • NCELoss
  • NCEWithLogitsLoss
  1. import torch.nn as nn
  2. import torch
  3. import numpy as np
  5. def bce_loss(pred, y):
  6.     return -y*np.log(pred) - (1-y)*np.log(1-pred)
  8. m = nn.Sigmoid()
  9. input = torch.randn(3, requires_grad=True)
  10. input_m = m(input)
  11. target = torch.empty(3).random_(2)
  12. print("predict origin:", input.data)
  13. print("predict after sigmoid:", input_m.data)
  14. print("actual label:", target)
  16. print("BCE Loss:", nn.BCELoss(reduction='none')(input_m, target).data)
  17. print("BCEwithLogitsLoss:", nn.BCEWithLogitsLoss(reduction='none')(input, target).data)
  18. print("BCE Loss Mean:", nn.BCELoss(reduction='mean')(input_m, target).data)
  20. input_m = input_m.detach().numpy()
  21. target = target.detach().numpy()
  22. print("Loss DIY", bce_loss(input_m, target))
  24. -- output 
  25. predict origin: tensor([0.9719, 0.0456, 0.0118])
  26. predict after sigmoid: tensor([0.7255, 0.5114, 0.5029])
  27. actual label: tensor([1., 0., 0.])
  29. BCE Loss: tensor([0.3209, 0.7162, 0.6990])
  30. BCEwithLogitsLoss: tensor([0.3209, 0.7162, 0.6990])
  31. BCE Loss Mean: tensor(0.5787)
  33. Loss DIY [0.32090703 0.71618414 0.6990468 ]

根据官方文档，计算结果与手动计算结果一致，pytorch默认会是求平均。