pytorch 标签平滑损失 交叉熵损失 nn.logSoftmax()和nn.NLLLoss()

    import torch
    import torch.nn as nn
    import math

    entroy = nn.CrossEntropyLoss()
    input = torch.Tensor([[-0.7715, -0.6205, -0.2562]])
    target = torch.tensor([0])
    output = entroy(input, target)
    print(output)


    m = nn.LogSoftmax(dim=1)
    loss = nn.NLLLoss()
    input=m(input)
    output = loss(input, target)
    print('output:',output)

    targets = torch.zeros(input.size()).scatter_(1, target.unsqueeze(1).data.cpu(), 1)  # onehot 形式
    out = (- targets * input).sum(dim=-1).mean()  #(1,0,0)  #  这句话就是 nn.NLLLoss() 的功能
    print(out)
    print("*****************")
    print(input)
    print(targets)

    targets = (1 - 0.1) * targets + 0.1 / 3  # 这是标签平滑 先变成 onehot
    loss = (- targets * input).sum(-1).mean()
    print("标签平滑损失")
    print(loss)
    print(targets)  #tensor([[1., 0., 0.]])    to   tensor([[0.9333, 0.0333, 0.0333]])

tensor(1.3447)
output: tensor(1.3447)
tensor(1.3447)
*****************
tensor([[-1.3447, -1.1937, -0.8294]])
tensor([[1., 0., 0.]])
标签平滑损失
tensor(1.3225)
tensor([[0.9333, 0.0333, 0.0333]])

pytorch损失函数之nn.CrossEntropyLoss()、nn.NLLLoss()