LabelSmoothing.py

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable

# Wangleiofficial
# https://github.com/pytorch/pytorch/issues/7455#issuecomment-720100742
class LabelSmoothingLoss(torch.nn.Module):
    def __init__(self, smoothing: float = 0.1,
                 reduction
 
="mean", weight=None):
        super(LabelSmoothingLoss, self).__init__()
        self.smoothing   = smoothing
        self.reduction = reduction
        self.weight    = weight

    def reduce_loss(self, loss):
        return loss.mean() if self.reduction == 'mean' else loss.sum() \
         if self.reduction == '
 
sum' else loss

    def linear_combination(self, x, y):
        return self.smoothing * x + (1 - self.smoothing) * y

    def forward(self, preds, target):
        assert 0 <= self.smoothing < 1

        if self.weight is not None:
            self.weight = self.weight.to(preds.device)

        n 
 
= preds.size(-1)
        log_preds = F.log_softmax(preds, dim=-1)
        loss = self.reduce_loss(-log_preds.sum(dim=-1))
        nll = F.nll_loss(
            log_preds, target, reduction=self.reduction, weight=self.weight
        )
        return self.linear_combination(loss / n, nll)


# NVIDIA
# https://github.com/NVIDIA/DeepLearningExamples/blob/8d8b21a933fff3defb692e0527fca15532da5dc6/PyTorch/Classification/ConvNets/image_classification/smoothing.py#L18
class LabelSmoothing(nn.Module):
    """NLL loss with label smoothing.
    """
    def __init__(self, smoothing=0.0): # 平滑因子
        """Constructor for the LabelSmoothing module.
        :param smoothing: label smoothing factor
        """
        super(LabelSmoothing, self).__init__()
        self.confidence = 1.0 - smoothing
        self.smoothing = smoothing

    def forward(self, x, target):
        logprobs = torch.nn.functional.log_softmax(x, dim=-1) # x: (batch size * class數量)，即log(p(k))
        nll_loss = -logprobs.gather(dim=-1, index=target.unsqueeze(1)) # target: (batch size) 數字標籤
        # 相當於取出logprobs中的真實標籤的那個位置的logit的負值
        nll_loss = nll_loss.squeeze(1) # (batch size * 1)再squeeze成batch size，即log(p(k))δk,y，δk,y表示除了k=y時該值為1，其餘為0
        smooth_loss = -logprobs.mean(dim=-1) # 在class維度取均值，就是對每個樣本x的所有類的logprobs取了平均值。
        # smooth_loss = -log(p(k))u(k) = -log(p(k))∗ 1/k

        loss = self.confidence * nll_loss + self.smoothing * smooth_loss # (batch size)
        # loss = (1−ϵ)log(p(k))δk,y + ϵlog(p(k))u(k)
        return loss.mean() # −∑ k=1~K [(1−ϵ)log(p(k))δk,y+ϵlog(p(k))u(k)]



if __name__=="__main__":
    # Wangleiofficial
    crit = LabelSmoothingLoss(smoothing=0.3, reduction="mean")
    predict = torch.FloatTensor([[0, 0.2, 0.7, 0.1, 0],
                                 [0, 0.9, 0.2, 0.2, 1],
                                 [1, 0.2, 0.7, 0.9, 1]])

    v = crit(Variable(predict),
             Variable(torch.LongTensor([2, 1, 0])))
    print(v)

    # NVIDIA
    crit = LabelSmoothing(smoothing=0.3)
    predict = torch.FloatTensor([[0, 0.2, 0.7, 0.1, 0],
                                 [0, 0.9, 0.2, 0.2, 1],
                                 [1, 0.2, 0.7, 0.9, 1]])
    v = crit(Variable(predict),
             Variable(torch.LongTensor([2, 1, 0])))
    print(v)

# tensor(1.3883)
# tensor(1.3883)

上面程式碼可以直接跑出LS的結果

兩個LS的實現方法來源於

# Wangleiofficial
# https://github.com/pytorch/pytorch/issues/7455#issuecomment-720100742

# NVIDIA
# https://github.com/NVIDIA/DeepLearningExamples/blob/8d8b21a933fff3defb692e0527fca15532da5dc6/PyTorch/Classification/ConvNets/image_classification/smoothing.py#L18

主要講解NVIDIA裡的實現過程

class LabelSmoothing(nn.Module):
    """NLL loss with label smoothing.
    """
    def __init__(self, smoothing=0.0): # 平滑因子
        """Constructor for the LabelSmoothing module.
        :param smoothing: label smoothing factor
        """
        super(LabelSmoothing, self).__init__()
        self.confidence = 1.0 - smoothing
        self.smoothing = smoothing

    def forward(self, x, target):
        logprobs = torch.nn.functional.log_softmax(x, dim=-1) # x: (batch size * class數量)，即log(p(k))
        nll_loss = -logprobs.gather(dim=-1, index=target.unsqueeze(1)) # target: (batch size) 數字標籤
        # 相當於取出logprobs中的真實標籤的那個位置的logit的負值
        nll_loss = nll_loss.squeeze(1) # (batch size * 1)再squeeze成batch size，即log(p(k))δk,y，δk,y表示除了k=y時該值為1，其餘為0
        smooth_loss = -logprobs.mean(dim=-1) # 在class維度取均值，就是對每個樣本x的所有類的logprobs取了平均值。
        # smooth_loss = -log(p(k))u(k) = -log(p(k))∗ 1/k

        loss = self.confidence * nll_loss + self.smoothing * smooth_loss # (batch size)
        # loss = (1−ϵ)log(p(k))δk,y + ϵlog(p(k))u(k)
        return loss.mean() # −∑ k=1~K [(1−ϵ)log(p(k))δk,y+ϵlog(p(k))u(k)]

LS通俗易懂的推導

https://blog.csdn.net/weixin_41811314/article/details/115863126保姆級程式碼

狄拉克delta

講解NLLLoss()和CrossEntropyLoss()的區別

https://blog.csdn.net/qq_22210253/article/details/85229988/

使用NLLLoss()的時候（即Negative Log Likelihood Loss），需要在網路的最後輸出層加上啟用函式，這也正是likelihood的定義，即輸出一個概率；而使用CrossEntropyLoss()的時候，網路的最後輸出不要加啟用，在 CrossEntropyLoss()中會幫我們完成此操作

NLLLoss (Negative Log Likelihood Loss）翻譯為 “負對數似然損失函式”，但並不計算對數，只是利用對數運算後的值（故需要和LogSofmax組合），進一步結合真實標籤計算“負對數似然值”。 “似然”的含義為：所求概率分佈與真實概率分佈的相似程度。在分類任務中，“似然”的含義就是預測值與真實標籤（one-hot型別）的接近程度（實質沒有改變）。

pytorch中的CrossEntropyLoss中是結合了logsoftmax和Nllloss

softmax，取自然對數後，與Label對應的那個值拿出來，再取相反數，再求均值（除以樣本數）。

Pytorch中的torch.gather函式詳解，從抽象到具體

https://blog.csdn.net/weixin_41811314/article/details/115869024

torch.gather(input, dim, index, *, sparse_grad=False, out=None)

dim=1時，out[i][j][k] = input[i][index[i][j][k]][k]

out[i][j][k]就是先獲取index[i][j][k]，設為a，dim為1，結果即為input[i][a][k]

input是個倉庫

更巨集觀的理解

也就是說，對於out中的每一個{m}位置，我們都去input這個倉庫裡找到對應的那一個點！
找到那一個點之後，我們從這個點出發，發射一條只照亮"dim"這條線的光線，在這條光線上再根據index找到我們真正需要的那個點就行了！