结论：Pytorch的nn.Dropout在每次被调用时dropout掉的参数都不一样，即使是同一次forward也不同。 如果模型里多次使用的dropout的dropout rate大小相同，用同一个dropout层即可。 如代码所示：
import torch import torch.nn as nn class MyModel(nn.Module): def __init__(self): super(MyModel, self).__init__() self.dropout_1 = nn.Dropout(0.5) self.dropout_2 = nn.Dropout(0.5) def forward(self, input): # print(input) drop_1 = self.dropout_1(input) print(drop_1) drop_1 = self.dropout_1(input) print(drop_1) drop_2 = self.dropout_2(input) print(drop_2) if __name__ == '__main__': i = torch.rand((5, 5)) m = MyModel() m.forward(i)结果如下：
*\python.exe */model.py tensor([[0.0000, 0.0914, 0.0000, 1.4095, 0.0000], [0.0000, 0.0000, 0.1726, 1.3800, 0.0000], [1.7651, 0.0000, 0.0000, 0.9421, 1.5603], [1.0510, 1.7290, 0.0000, 0.0000, 0.8565], [0.0000, 0.0000, 0.0000, 0.0000, 0.0000]]) tensor([[0.0000, 0.0000, 0.4722, 1.4095, 0.0000], [0.0416, 0.0000, 0.1726, 1.3800, 1.3193], [0.0000, 0.3401, 0.6550, 0.0000, 0.0000], [1.0510, 1.7290, 1.5515, 0.0000, 0.0000], [0.6388, 0.0000, 0.0000, 1.0122, 0.0000]]) tensor([[0.0000, 0.0000, 0.4722, 0.0000, 1.2689], [0.0416, 0.0000, 0.0000, 1.3800, 0.0000], [0.0000, 0.0000, 0.6550, 0.0000, 1.5603], [0.0000, 0.0000, 1.5515, 1.4596, 0.0000], [0.0000, 0.0000, 0.0000, 0.0000, 0.0000]]) Process finished with exit code 0