utils.logmeanexp(outputs, dim=1)

1.utils.logmeanexp(outputs, dim)

outputs为输入张量， dim为指定维数（指定后该维消失）
特殊的，outputs一维时，dim可不加

import torch
import utils
a = torch.randn(3,2,1)
outputs = torch.tensor(a)

'''1'''
log_outputs = utils.logmeanexp(outputs, dim=2)
print('outputs:', outputs, outputs.shape, 'log_outputs:', log_outputs, log_outputs.shape)
# outputs: torch.Size([3, 2, 1]) log_outputs: torch.Size([3, 2])


'''2'''
log_outputs = utils.logmeanexp(outputs, dim=1)
print('outputs:', outputs, outputs.shape, 'log_outputs:', log_outputs, log_outputs.shape)
# outputs: torch.Size([3, 2, 1]) log_outputs: torch.Size([3, 1])

'''3'''
log_outputs = utils.logmeanexp(outputs, dim=0)
print('outputs:', outputs, outputs.shape, 'log_outputs:', log_outputs, log_outputs.shape)
# outputs: torch.Size([3, 2, 1]) log_outputs: torch.Size([3, 1])

'''4'''
outputs = torch.randn(3)
log_outputs = utils.logmeanexp(outputs)###log_outputs = utils.logmeanexp(outputs, dim=0)也可以
print('outputs:', outputs, outputs.shape, 'log_outputs:', log_outputs, log_outputs.shape)
#outputs: tensor([-0.3776, -0.5505,  1.6139]) torch.Size([3]) log_outputs: tensor(0.7395) torch.Size([])

2.代码解释

1.对于log_outputs = utils.logmeanexp(outputs, dim=2)有

outputs: tensor([[[-1.1954],
         [ 0.1321]],

        [[ 1.4576],
         [-0.0446]],

        [[-0.9327],
         [-1.3143]]]) torch.Size([3, 2, 1]) log_outputs: tensor([[-1.1954,  0.1321],
        [ 1.4576, -0.0446],
        [-0.9327, -1.3143]]) torch.Size([3, 2])

2.对于log_outputs = utils.logmeanexp(outputs, dim=1)有

outputs: tensor([[[-0.9949],
         [-0.2391]],

        [[ 0.3857],
         [-1.1477]],

        [[-0.5240],
         [-2.0448]]]) torch.Size([3, 2, 1]) log_outputs: tensor([[-0.5472],
        [-0.1120],
        [-1.0195]]) torch.Size([3, 1])

第二维消失，log_outputs为两维的结果取log平均如式（2.1）所示（注：log为e为底数，对示例中N=2）：$\log \frac{1}{N}{\sum }_{n=1}^N{e}_i^x,$（2.1）

3和4同理

3. 总结

要点：对指定的dim使用公式$\log \frac{1}{N}{\sum }_{n=1}^N{e}_i^x,$，

作用：降低维度，取log平均？？？（还不确定）