版权声明:本文为博主原创文章,遵循 CC 4.0 BY-SA 版权协议,转载请附上原文出处链接和本声明。
本文链接:https://blog.csdn.net/chaipp0607/article/details/101946040
Cross Entropy是分类问题中常见的一种损失函数,我们在之前的文章提到过二值交叉熵的证明和交叉熵的作用,下面解释一下交叉熵损失的求导。 首先一个模型的最后一层神经元的输出记为f0...fif_{0}...f_{i}f0...fi, 输出经过softmax激活之后记为p0...pip_{0}...p_{i}p0...pi,那么: pi=efi∑k=0C−1efkp_{i} = frac{e^{f_{i}}}{sum_{k=0}^{C-1} e^{f_{k}}}pi=∑k=0C−1efkefi 类别的实际标签记为y0...yiy_{0}...y_{i}y0...yi,那么交叉熵损失L为: L=−∑i=0C−1yilogpiL = -sum_{i=0}^{C-1} y_{i}log^{p_{i}}L=−i=0∑C−1yilogpi 上式中的logloglog是一种简写,为了后续的求导方便,一般我们认为logloglog的底是eee,即logloglog为lnlnln。 那么LLL对第iii个神经元的输出fif_{i}fi求偏导∂L∂fifrac{partial L}{partial f_{i}}∂fi∂L: 根据复合函数求导原则: ∂L∂fi=∑j=0C−1∂Lj∂pj∂pj∂fifrac{partial L}{partial f_{i}} = sum_{j=0}^{C-1} frac{partial L_{j}}{partial p_{j}}frac{partial p_{j}}{partial f_{i}}∂fi∂L=j=0∑C−1∂pj∂Lj∂fi∂pj 在这里需要说明,在softmax中我们使用了下标iii和kkk,在交叉熵中使用了下标iii,但是这里的两个iii并不等价,因为softmax的分母中包含了每个神经元的输出fff,也就是激活后所有的ppp对任意的fif_{i}fi求偏导都不为0,同时LLL中又包含了所有的ppp,所以为了避免重复我们需要为ppp引入一个新的下标jjj,jjj有0...C−10...C-10...C−1这C种情况。 那么依次求导:
∂Lj∂pj=∂(−yjlogpj)∂(pj)frac{partial L_{j}}{partial p_{j}}= frac{partial (-y_{j}log^{p_{j}})}{partial (p_{j})}∂pj∂Lj=∂(pj)∂(−yjlogpj)
由于默认一般我们认为logloglog的底是eee,即logloglog为lnlnln,所以:
∂Lj∂pj=∂(−yjlogpj)∂(pj)=−yjpjfrac{partial L_{j}}{partial p_{j}}= frac{partial (-y_{j}log^{p_{j}})}{partial (p_{j})} =-frac{y_{j}}{p_{j}}∂pj∂Lj=∂(pj)∂(−yjlogpj)=−pjyj
接着要求∂pj∂fifrac{partial p_{j}}{partial f_{i}}∂fi∂pj的值,在这里可以发现,每一个pjp_{j}pj中都包含fif_{i}fi,所以∂pj∂fifrac{partial p_{j}}{partial f_{i}}∂fi∂pj都不是0,但是j=ij=ij=i和j≠ij neq ij=i的时候,∂pj∂fifrac{partial p_{j}}{partial f_{i}}∂fi∂pj结果又不相同,所以这里需要分开讨论:
- 首先j=ij=ij=i时: ∂pj∂fi=∂pi∂fi=∂efi∑k=0C−1efk∂fifrac{partial p_{j}}{partial f_{i}} = frac{partial p_{i}}{partial f_{i}} = frac{partial frac{e^{f_{i}}}{sum_{k=0}^{C-1} e^{f_{k}}}}{partial f_{i}} ∂fi∂pj=∂fi∂pi=∂fi∂∑k=0C−1efkefi =(efi)′∑k=0C−1efk−efi(∑k=0C−1efk)′(∑k=0C−1efk)2= frac{ (e^{f_{i}})' sum_{k=0}^{C-1} e^{f_{k}} - e^{f_{i}}(sum_{k=0}^{C-1} e^{f_{k}})' }{(sum_{k=0}^{C-1} e^{f_{k}})^{2}} =(∑k=0C−1efk)2(efi)′∑k=0C−1efk−efi(∑k=0C−1efk)′ =efi∑k=0C−1efk−(efi)2(∑k=0C−1efk)2=efi∑k=0C−1efk−(efi∑k=0C−1efk)2= frac{ e^{f_{i}}sum_{k=0}^{C-1} e^{f_{k}} - (e^{f_{i}})^2 }{(sum_{k=0}^{C-1} e^{f_{k}})^{2}}= frac{ e^{f_{i}} }{sum_{k=0}^{C-1} e^{f_{k}}} - (frac{ e^{f_{i}} }{sum_{k=0}^{C-1} e^{f_{k}}})^2=(∑k=0C−1efk)2efi∑k=0C−1efk−(efi)2=∑k=0C−1efkefi−(∑k=0C−1efkefi)2 =pi−(pi)2=pi(1−pi) = p_{i}-(p{i})^2 = p_{i}(1-p_{i})=pi−(pi)2=pi(1−pi)
- 然后j≠ijneq ij=i时: ∂pj∂fi=∂efj∑k=0C−1efk∂fifrac{partial p_{j}}{partial f_{i}}= frac{partial frac{e^{f_{j}}}{sum_{k=0}^{C-1} e^{f_{k}}}}{partial f_{i}} ∂fi∂pj=∂fi∂∑k=0C−1efkefj =(efj)′∑k=0C−1efk−efj(∑k=0C−1efk)′(∑k=0C−1efk)2= frac{ (e^{f_{j}})' sum_{k=0}^{C-1} e^{f_{k}} - e^{f_{j}}(sum_{k=0}^{C-1} e^{f_{k}})' }{(sum_{k=0}^{C-1} e^{f_{k}})^{2}} =(∑k=0C−1efk)2(efj)′∑k=0C−1efk−efj(∑k=0C−1efk)′ =−efiefj(∑k=0C−1efk)2=−efi∑k=0C−1efkefj∑k=0C−1efk= frac{ - e^{f_{i}} e^{f_{j}} }{(sum_{k=0}^{C-1} e^{f_{k}})^{2}} = - frac{ e^{f_{i}} }{sum_{k=0}^{C-1} e^{f_{k}}} frac{ e^{f_{j}} }{sum_{k=0}^{C-1} e^{f_{k}}}=(∑k=0C−1efk)2−efiefj=−∑k=0C−1efkefi∑k=0C−1efkefj =−pipj = -p_{i}p_{j}=−pipj
对于最后的偏导数,需要把上述两个部分加起来: ∂L∂fi=∑j=iC−1∂Lj∂pj∂pj∂fi ∑j≠iC−1∂Lj∂pj∂pj∂fifrac{partial L}{partial f_{i}} = sum_{j=i}^{C-1} frac{partial L_{j}}{partial p_{j}}frac{partial p_{j}}{partial f_{i}} sum_{jneq i}^{C-1} frac{partial L_{j}}{partial p_{j}}frac{partial p_{j}}{partial f_{i}}∂fi∂L=j=i∑C−1∂pj∂Lj∂fi∂pj j=i∑C−1∂pj∂Lj∂fi∂pj =−yipipi(1−pi) ∑j≠iC−1−pipj(−yjpj)=-frac{y_{i}}{p_{i}}p_{i}(1-p_{i}) sum_{jneq i}^{C-1}-p_{i}p_{j}(-frac{y_{j}}{p_{j}})=−piyipi(1−pi) j=i∑C−1−pipj(−pjyj) =−yi(1−pi) ∑j≠iC−1piyj=-y_{i}(1-p_{i}) sum_{jneq i}^{C-1}p_{i}y_{j}=−yi(1−pi) j=i∑C−1piyj =yipi−yi ∑j≠iC−1piyj=y_{i}p_{i}-y_{i} sum_{jneq i}^{C-1}p_{i}y_{j}=yipi−yi j=i∑C−1piyj
在上式中,j≠ijneq ij=i的情况中刚好缺了j=ij=ij=i,所以可以继续改写为: =∑j=0C−1piyj−yi=sum_{j=0}^{C-1}p_{i}y_{j} - y_{i} =j=0∑C−1piyj−yi =pi∑j=0C−1yj−yi=p_{i}sum_{j=0}^{C-1}y_{j} - y_{i} =pij=0∑C−1yj−yi 而∑j=0C−1yj=1sum_{j=0}^{C-1}y_{j} = 1∑j=0C−1yj=1,所以: =pi∑j=0C−1yj−yi=pi−yi=p_{i}sum_{j=0}^{C-1}y_{j} - y_{i} = p_{i}-y_{i} =pij=0∑C−1yj−yi=pi−yi