神经网络推导-矩阵计算

为了理清如何进行神经网络的前向传播和反向传播的推导，找了很多资料，前向传播比较简单，重点在于如何进行反向传播的梯度计算

cs231n课程推荐的计算方式是先进行单个元素求导，再逐步泛化到批量数据求梯度，参考

Derivatives, Backpropagation, and Vectorization - CS231n

Vector, Matrix, and Tensor Derivatives - CS231n

Backpropagation and Neural Networks - CS231n

自己也根据参考资料进行2层神经网络逐元素的推导

神经网络推导-单个数据

神经网络推导-批量数据

最好的方式当然是进行矩阵求导，在网上看了很多博客，比较好的有

矩阵求导术（上）

[矩阵求导]神经网络反向传播梯度计算数学原理

神经网络矩阵计算最重要的内容是进行实值标量矩阵的一阶微分以及$Jacobian$矩阵的辨识，参考《矩阵分析与应用》，有以下先导知识

导数、微分和梯度

矩阵基础

Jacobian矩阵和梯度矩阵

实值标量函数一阶微分和Jacobian矩阵辨识

使用矩阵微分能够很便捷的实现神经网络反向求导，关键部分是辨识$Jacobian$矩阵，再转换成梯度矩阵

推导一

文章[矩阵求导]神经网络反向传播梯度计算数学原理给出了一个很好的推导方式，首先给出实现代码，然后使用矩阵计算逐步解释代码

PyTorch教程Learning PyTorch with Examples给出了一个2层神经网络的numpy实现

# -*- coding: utf-8 -*-
import numpy as np

# N is batch size; D_in is input dimension;
# H is hidden dimension; D_out is output dimension.
N, D_in, H, D_out = 64, 1000, 100, 10

# Create random input and output data
x = np.random.randn(N, D_in)
y = np.random.randn(N, D_out)

# Randomly initialize weights
w1 = np.random.randn(D_in, H)
w2 = np.random.randn(H, D_out)

learning_rate = 1e-6
for t in range(500):
    # Forward pass: compute predicted y
    h = x.dot(w1)
    h_relu = np.maximum(h, 0)
    y_pred = h_relu.dot(w2)

    # Compute and print loss
    loss = np.square(y_pred - y).sum()
    print(t, loss)

    # Backprop to compute gradients of w1 and w2 with respect to loss
    grad_y_pred = 2.0 * (y_pred - y)
    grad_w2 = h_relu.T.dot(grad_y_pred)
    grad_h_relu = grad_y_pred.dot(w2.T)
    grad_h = grad_h_relu.copy()
    grad_h[h < 0] = 0
    grad_w1 = x.T.dot(grad_h)

    # Update weights
    w1 -= learning_rate * grad_w1
    w2 -= learning_rate * grad_w2

第一步：定义网络参数

• 批量数据大小$N=64$
• 输入层神经元个数$D_{in}=1000$
• 隐藏层神经元个数$H=100$
• 输出层神经元个数$D_{out}=10$

# N is batch size; D_in is input dimension;
# H is hidden dimension; D_out is output dimension.
N, D_in, H, D_out = 64, 1000, 100, 10

第二步：初始化数据、权重（该网络没有偏置向量）以及学习率

• 输入数据$x\in R^{N\times D_{in}}$
• 输出数据$y\in R^{N\times D_{out}}$
• 隐藏层权重矩阵$w1\in R^{D_{in}\times H}$
• 输出层权重矩阵$w2\in R^{H\times D_{out}}$

# Create random input and output data
x = np.random.randn(N, D_in)
y = np.random.randn(N, D_out)

# Randomly initialize weights
w1 = np.random.randn(D_in, H)
w2 = np.random.randn(H, D_out)

learning_rate = 1e-6

第三步：迭代计算，输入批量数据到神经网络，进行前向传播

$$\begin{gathered} h=x\cdot w1 \\ {h}_{relu}=max(0,h) \\ {y}_{pred}=h_{relu}\cdot w2 \end{gathered}$$

# Forward pass: compute predicted y
h = x.dot(w1)
h_relu = np.maximum(h, 0)
y_pred = h_relu.dot(w2)

第四步：迭代计算，计算损失函数（误差平方和 - L1范数的平方）

$$loss={\parallel \begin{array}{c}{y}_{pred}-y\end{array}\parallel }^2$$

# Compute and print loss
loss = np.square(y_pred - y).sum()
print(t, loss)

第五步：迭代计算，反向传播，计算输出层输入向量梯度

设$y_{pred}-y=X$，$X$大小为$N\times D_{out}$，则

$$loss={\parallel \begin{array}{c}X\end{array}\parallel }^2=(vec(X){)}^T\cdot vec(X)$$

对损失函数$loss(y_{pred})$求输出层输入向量的微分

$$\begin{gathered} dloss=d(tr(loss))=tr(dloss)=tr(d((vec(X){)}^T\cdot vec(X))) \\ =tr(d(vec(X{)}^T)\cdot vec(X)+vec(X{)}^T\cdot dvec(X)) \\ =tr(d(vec(X{)}^T)\cdot vec(X))+tr(vec(X{)}^T\cdot dvec(X)) \\ =tr((dvec(X){)}^T\cdot vec(X))+tr(vec(X{)}^T\cdot dvec(X)) \\ =tr((vec(X){)}^T\cdot dvec(X))+tr(vec(X{)}^T\cdot dvec(X)) \\ =tr(2(vec(X){)}^T\cdot dvec(X)) \\ =tr(2X^T\cdot dX) \end{gathered}$$

所以Jacobian矩阵为$D_{X}f(X)=2X^T$，梯度矩阵为${▽}_{X}f(X)=2X=2(y_{pred}-y)$

grad_y_pred = 2.0 * (y_pred - y)

第六步：迭代计算，反向传播，计算输出层权重向量以及隐藏层输出向量梯度

$$y_{pred}=h_{relu}\cdot w2\Rightarrow d{y}_{pred}=d{h}_{relu}\cdot w2+h_{relu}\cdot dw2$$

$$\begin{gathered} dloss=tr(2X^T\cdot dX)=tr(2(y_{pred}-y{)}^T\cdot d((y_{pred}-y)))=tr(2(y_{pred}-y{)}^T\cdot d{y}_{pred}) \\ =tr(2(y_{pred}-y{)}^T\cdot (d{h}_{relu}\cdot w2+h_{relu}\cdot dw2)) \\ =tr(2(y_{pred}-y{)}^T\cdot d{h}_{relu}\cdot w2)+tr(2(y_{pred}-y{)}^T\cdot h_{relu}\cdot dw2) \\ =tr(w2\cdot 2(y_{pred}-y{)}^T\cdot d{h}_{relu})+tr(2(y_{pred}-y{)}^T\cdot h_{relu}\cdot dw2) \end{gathered}$$

输出层权重向量的Jacobian矩阵为$2(y_{pred}-y{)}^T\cdot h_{relu}$，梯度矩阵为$(h_{relu}{)}^T\cdot 2(y_{pred}-y)$

隐藏层输出向量的Jacobian矩阵为$w2\cdot 2(y_{pred}-y{)}^T$，梯度矩阵为$2(y_{pred}-y)\cdot (w2{)}^T$

grad_w2 = h_relu.T.dot(grad_y_pred)
grad_h_relu = grad_y_pred.dot(w2.T)

第七步：迭代计算，反向传播，计算隐藏层输入向量梯度

$$h_{relu}=max(0,h)\Rightarrow d{h}_{relu}=\{\begin{array}{cc}dh& h\ge 0\\ 0& h<0\end{array}=1(h\ge 0)\ast dh$$

激活函数是逐个元素操作，所以使用Hadamard积

$$\begin{gathered} dloss=tr(w2\cdot 2(y_{pred}-y{)}^T\cdot d{h}_{relu}) \\ =tr(w2\cdot 2(y_{pred}-y{)}^T\cdot 1(h\ge 0)\ast dh) \\ =tr((2(y_{pred}-y)\cdot (w2{)}^T{)}^T\cdot 1(h\ge 0)\ast dh) \\ =tr((2(y_{pred}-y)\cdot (w2{)}^T{)}^T\ast 1(h\ge 0{)}^T\cdot dh) \end{gathered}$$

所以Jacobian矩阵为$(2(y_{pred}-y)\cdot (w2{)}^T{)}^T\ast 1(h\ge 0{)}^T$，梯度矩阵为

$${▽}_{h}f(h)=1(h\ge 0)\cdot 2(y_{pred}-y)\cdot (w2{)}^T$$

grad_h = grad_h_relu.copy()
grad_h[h < 0] = 0

第八步：迭代计算，反向传播，计算隐藏层权重向量梯度

$$h=x\cdot w1\Rightarrow dh=x\cdot dw1$$

$$\begin{gathered} dloss=tr((2(y_{pred}-y)\cdot (w2{)}^T{)}^T\ast 1(h\ge 0{)}^T\cdot dh) \\ =tr((2(y_{pred}-y)\cdot (w2{)}^T{)}^T\ast 1(h\ge 0{)}^T\cdot x\cdot dw1) \end{gathered}$$

所以Jacobian矩阵为$(2(y_{pred}-y)\cdot (w2{)}^T{)}^T\ast 1(h\ge 0{)}^T\cdot x$，梯度矩阵为

$$\begin{gathered} {▽}_{w1}f(w1)=((2(y_{pred}-y)\cdot (w2{)}^T{)}^T\ast 1(h\ge 0{)}^T\cdot x{)}^T \\ =x^T\cdot 1(h\ge 0)\ast 2(y_{pred}-y)\cdot (w2{)}^T \end{gathered}$$

grad_w1 = x.T.dot(grad_h)

第九步：迭代计算，反向传播，更新权重矩阵

# Update weights
w1 -= learning_rate * grad_w1
w2 -= learning_rate * grad_w2

推导二

cs231n课程Putting it together: Minimal Neural Network Case Study中实现了一个2层神经网络

N = 100 # number of points per class
D = 2 # dimensionality
K = 3 # number of classes
X = np.zeros((N*K,D)) # data matrix (each row = single example)
y = np.zeros(N*K, dtype='uint8') # class labels
for j in xrange(K):
  ix = range(N*j,N*(j+1))
  r = np.linspace(0.0,1,N) # radius
  t = np.linspace(j*4,(j+1)*4,N) + np.random.randn(N)*0.2 # theta
  X[ix] = np.c_[r*np.sin(t), r*np.cos(t)]
  y[ix] = j

# initialize parameters randomly
h = 100 # size of hidden layer
W = 0.01 * np.random.randn(D,h)
b = np.zeros((1,h))
W2 = 0.01 * np.random.randn(h,K)
b2 = np.zeros((1,K))

# some hyperparameters
step_size = 1e-0
reg = 1e-3 # regularization strength

# gradient descent loop
num_examples = X.shape[0]
for i in xrange(10000):
  
  # evaluate class scores, [N x K]
  hidden_layer = np.maximum(0, np.dot(X, W) + b) # note, ReLU activation
  scores = np.dot(hidden_layer, W2) + b2
  
  # compute the class probabilities
  exp_scores = np.exp(scores)
  probs = exp_scores / np.sum(exp_scores, axis=1, keepdims=True) # [N x K]
  
  # compute the loss: average cross-entropy loss and regularization
  correct_logprobs = -np.log(probs[range(num_examples),y])
  data_loss = np.sum(correct_logprobs)/num_examples
  reg_loss = 0.5*reg*np.sum(W*W) + 0.5*reg*np.sum(W2*W2)
  loss = data_loss + reg_loss
  if i % 1000 == 0:
    print "iteration %d: loss %f" % (i, loss)
  
  # compute the gradient on scores
  dscores = probs
  dscores[range(num_examples),y] -= 1
  dscores /= num_examples
  
  # backpropate the gradient to the parameters
  # first backprop into parameters W2 and b2
  dW2 = np.dot(hidden_layer.T, dscores)
  db2 = np.sum(dscores, axis=0, keepdims=True)
  # next backprop into hidden layer
  dhidden = np.dot(dscores, W2.T)
  # backprop the ReLU non-linearity
  dhidden[hidden_layer <= 0] = 0
  # finally into W,b
  dW = np.dot(X.T, dhidden)
  db = np.sum(dhidden, axis=0, keepdims=True)
  
  # add regularization gradient contribution
  dW2 += reg * W2
  dW += reg * W
  
  # perform a parameter update
  W += -step_size * dW
  b += -step_size * db
  W2 += -step_size * dW2
  b2 += -step_size * db2

第一步：设置批量输入数据和输出数据

• 批量数据大小$N=100$
• 数据维数$D=2$
• 类别数$K=3$
• 输入数据$X\in R^{N\times D}$
• 输出数据$y\in R^{N\times K}$

N = 100 # number of points per class
D = 2 # dimensionality
K = 3 # number of classes
X = np.zeros((N*K,D)) # data matrix (each row = single example)
y = np.zeros(N*K, dtype='uint8') # class labels
for j in xrange(K):
  ix = range(N*j,N*(j+1))
  r = np.linspace(0.0,1,N) # radius
  t = np.linspace(j*4,(j+1)*4,N) + np.random.randn(N)*0.2 # theta
  X[ix] = np.c_[r*np.sin(t), r*np.cos(t)]
  y[ix] = j

第二步：初始化权重参数

• 隐藏层神经元个数$h=100$
• 隐藏层权重矩阵$W\in R^{D\times h}$
• 隐藏层偏置向量$b\in R^{1\times h}$
• 输出层权重矩阵$W2\in R^{h\times K}$
• 输出层偏置向量$b2\in R^{1\times K}$

# initialize parameters randomly
h = 100 # size of hidden layer
W = 0.01 * np.random.randn(D,h)
b = np.zeros((1,h))
W2 = 0.01 * np.random.randn(h,K)
b2 = np.zeros((1,K))

第三步：设置学习率和正则化强度

# some hyperparameters
step_size = 1e-0
reg = 1e-3 # regularization strength

第四步：迭代计算，输入批量数据到神经网络，进行前向传播

$$\begin{gathered} hiddenLayer=max(X\cdot W+b,0) \\ scores=hiddenLayer\cdot W2+b2 \end{gathered}$$

# evaluate class scores, [N x K]
hidden_layer = np.maximum(0, np.dot(X, W) + b) # note, ReLU activation
scores = np.dot(hidden_layer, W2) + b2

第四步：迭代计算，计算损失值

$$\begin{gathered} expScores=exp(scores) \\ probs=\frac{expScores}{expScores\cdot 1} \\ correctLogProbs=-\ln prob{s}_y\in R^{N\times 1} \\ dataLoss=\frac{1}{N}{1}^T\cdot correctLogProbs \\ regLoss=0.5\cdot reg\cdot ||W|{|}^2+0.5\cdot reg\cdot ||W2|{|}^2 \\ loss=dataLoss+regLoss \end{gathered}$$

$1$表示求和向量：$[1,1,...{]}^T$

$prob{s}_y$表示每行正确类别的概率

# compute the class probabilities
exp_scores = np.exp(scores)
probs = exp_scores / np.sum(exp_scores, axis=1, keepdims=True) # [N x K]

# compute the loss: average cross-entropy loss and regularization
correct_logprobs = -np.log(probs[range(num_examples),y])
data_loss = np.sum(correct_logprobs)/num_examples
reg_loss = 0.5*reg*np.sum(W*W) + 0.5*reg*np.sum(W2*W2)
loss = data_loss + reg_loss
if i % 1000 == 0:
    print "iteration %d: loss %f" % (i, loss)

第五步：迭代计算，反向传播，计算输出层输入向量梯度

$$\begin{gathered} score{s}_y=scores\ast Y\cdot 1 \\ expscore{s}_y=exp(scores\ast Y\cdot 1)expscores=exp(scores)expscore{s}_{sum}=exp(scores)\cdot 1 \\ prob{s}_y=\frac{expscore{s}_y}{expscore{s}_{sum}}probs=\frac{expscores}{expscore{s}_{sum}} \end{gathered}$$

$$\begin{gathered} dataloss=-\frac{1}{N}{1}^T\cdot \ln (prob{s}_y)=-\frac{1}{N}{1}^T\cdot \ln \frac{expscore{s}_y}{expscore{s}_{sum}} \\ =-\frac{1}{N}{1}^T\cdot (\ln expscore{s}_y-\ln expscore{s}_{sum}) \\ =-\frac{1}{N}{1}^T\cdot (scores\ast Y\cdot 1-\ln expscore{s}_{sum}) \end{gathered}$$

$$\begin{gathered} d(dataloss)=tr(d(-\frac{1}{N}(1^T\cdot scores\ast Y\cdot 1-1^T\cdot \ln expscore{s}_{sum}))) \\ =tr(d(-\frac{1}{N}(1^T\cdot scores\ast Y\cdot 1)))-tr(d(-\frac{1}{N}(1^T\cdot \ln expscore{s}_{sum}))) \end{gathered}$$

$$\begin{gathered} tr(d(-\frac{1}{N}(1^T\cdot scores\ast Y\cdot 1)))=tr(-\frac{1}{N}(1^T\cdot dscores\ast Y\cdot 1)) \\ =tr(-\frac{1}{N}(dscore{s}^T\cdot Y))=tr(-\frac{1}{N}{Y}^T\cdot dscores) \end{gathered}$$

$$\begin{gathered} tr(d(-\frac{1}{N}(1^T\cdot \ln expscore{s}_{sum})))=tr(-\frac{1}{N}(1^T\cdot expscore{s}_{sum}^{-1}\cdot dexpscore{s}_{sum})) \\ =tr(-\frac{1}{N}\frac{(1^T\cdot dexpscore{s}_{sum})}{expscore{s}_{sum}})=tr(-\frac{1}{N}\frac{(1^T\cdot exp(scores)\ast dscores\cdot 1)}{expscore{s}_{sum}}) \\ =tr(-\frac{1}{N}\frac{exp(scores{)}^T\cdot dscores}{expscore{s}_{sum}})=tr(-\frac{1}{N}(\frac{exp(scores)}{expscore{s}_{sum}}{)}^T\cdot dscores)=tr(-\frac{1}{N}prob{s}^T\cdot dscores) \end{gathered}$$

$$\begin{gathered} \Rightarrow d(dataloss)=tr(-\frac{1}{N}{Y}^T\cdot dscores)-tr(-\frac{1}{N}prob{s}^T\cdot dscores) \\ =tr(\frac{1}{N}(prob{s}^T-Y^T)\cdot dscores) \end{gathered}$$

所以$Jacobian$矩阵为$D_{scores}f(scores)=prob{s}^T-Y^T$，梯度矩阵为${▽}_{scores}f(scores)=probs-Y$

• $Y$大小为$N\times K$，每行仅正确类别位置为1，其余为0
• $1$是求和向量，$[1,1,...{]}^T$

计算softmax分类的交叉熵损失关于输出层输入向量梯度，这一部分想了好久，主要问题是关于矩阵除法和逐元素除法（标量除法）的分别，感觉还是先对单个数据进行求梯度再泛化比较方便

# compute the gradient on scores
dscores = probs
dscores[range(num_examples),y] -= 1
dscores /= num_examples

第六步：迭代计算，反向传播，计算输出层权重矩阵、偏置向量以及隐藏层输出向量梯度

$$\begin{gathered} scores=hiddenLayer\cdot W2+b2 \\ dscores=dhiddenLayer\cdot W2+hiddenLayer\cdot dW2+db2 \end{gathered}$$

$$\begin{gathered} d(dataloss)=tr(\frac{1}{N}(prob{s}^T-Y^T)\cdot dscores) \\ =tr(\frac{1}{N}(prob{s}^T-Y^T)\cdot (dhiddenLayer\cdot W2+hiddenLayer\cdot dW2+db2)) \\ =tr(\frac{1}{N}(prob{s}^T-Y^T)\cdot dhiddenLayer\cdot W2) \\ +tr(\frac{1}{N}(prob{s}^T-Y^T)\cdot hiddenLayer\cdot dW2)+tr(\frac{1}{N}(prob{s}^T-Y^T)\cdot db2) \end{gathered}$$

求输出层权重矩阵梯度

$$d(dataloss)=tr(\frac{1}{N}(prob{s}^T-Y^T)\cdot hiddenLayer\cdot dW2)$$

$$\begin{gathered} D_{W2}f(W2)=\frac{1}{N}(prob{s}^T-Y^T)\cdot hiddenLayer \\ {▽}_{W2}f(W2)=\frac{1}{N}hiddenLaye{r}^T\cdot (probs-Y) \end{gathered}$$

求输出层偏置向量梯度

$$d(dataloss)=tr(\frac{1}{N}\sum _{i=1}^N(prob{s}_i^T-Y_i^T)\cdot db2)$$

$$\begin{gathered} D_{b2}f(b2)=\frac{1}{N}\sum _{i=1}^N(prob{s}_i^T-Y_i^T) \\ {▽}_{b2}f(b2)=\frac{1}{N}\sum _{i=1}^N(prob{s}_i-Y_i) \end{gathered}$$

对偏置向量还需要注意维数，求和批量数据的偏置向量梯度

求隐藏层输出向量梯度

$$d(dataloss)=tr(\frac{1}{N}(prob{s}^T-Y^T)\cdot dhiddenLayer\cdot W2)=tr(\frac{1}{N}W2\cdot (prob{s}^T-Y^T)\cdot dhiddenLayer)$$

$$\begin{gathered} D_{hiddenLayer}f(hiddenLayer)=\frac{1}{N}W2\cdot (prob{s}^T-Y^T) \\ {▽}_{hiddenLayer}f(hiddenLayer)=\frac{1}{N}(probs-Y)\cdot (W2{)}^T \end{gathered}$$

# backpropate the gradient to the parameters
# first backprop into parameters W2 and b2
dW2 = np.dot(hidden_layer.T, dscores)
db2 = np.sum(dscores, axis=0, keepdims=True)
# next backprop into hidden layer
dhidden = np.dot(dscores, W2.T)

第七步：迭代计算，反向传播，计算隐藏层输入向量梯度

$$\begin{gathered} hiddenLaye{r}_{in}=X\cdot W+b \\ hiddenLayer=max(0,hiddenLaye{r}_{in}) \\ dhiddenLayer=1(hiddenLaye{r}_{in}\ge 0)\ast dhiddenLaye{r}_{in} \end{gathered}$$

$$\begin{gathered} d(dataloss)=tr(\frac{1}{N}W2\cdot (prob{s}^T-Y^T)\cdot dhiddenLayer) \\ =tr(\frac{1}{N}W2\cdot (prob{s}^T-Y^T)\cdot 1(hiddenLaye{r}_{in}\ge 0)\ast dhiddenLaye{r}_{in}) \\ =tr(\frac{1}{N}(W2\cdot (prob{s}^T-Y^T){)}^T\ast 1(hiddenLaye{r}_{in}\ge 0{)}^T\cdot dhiddenLaye{r}_{in}) \end{gathered}$$

$$\begin{gathered} D_{hiddenLaye{r}_{in}}f(hiddenLaye{r}_{in})=\frac{1}{N}(W2\cdot (prob{s}^T-Y^T){)}^T\ast 1(hiddenLaye{r}_{in}\ge 0{)}^T \\ {▽}_{hiddenLaye{r}_{in}}f(hiddenLaye{r}_{in})=\frac{1}{N}((probs-Y)\cdot (W2{)}^T)\ast 1(hiddenLaye{r}_{in}\ge 0) \end{gathered}$$

# backprop the ReLU non-linearity
dhidden[hidden_layer <= 0] = 0

第七步：迭代计算，反向传播，计算隐藏层权重向量和偏置向量梯度

$$\begin{gathered} hiddenLaye{r}_{in}=X\cdot W+b \\ dhiddenLaye{r}_{in}=X\cdot dW+db \end{gathered}$$

$$\begin{gathered} d(dataloss)=tr(\frac{1}{N}(W2\cdot (prob{s}^T-Y^T){)}^T\ast 1(hiddenLaye{r}_{in}\ge 0{)}^T\cdot dhiddenLaye{r}_{in}) \\ =tr(\frac{1}{N}(W2\cdot (prob{s}^T-Y^T){)}^T\ast 1(hiddenLaye{r}_{in}\ge 0{)}^T\cdot (X\cdot dW+db)) \end{gathered}$$

求隐藏层权重向量梯度

$$d(dataloss)=tr(\frac{1}{N}(W2\cdot (prob{s}^T-Y^T){)}^T\ast 1(hiddenLaye{r}_{in}\ge 0{)}^T\cdot X\cdot dW)$$

$$\begin{gathered} D_{W}f(W)=\frac{1}{N}(W2\cdot (prob{s}^T-Y^T){)}^T\ast 1(hiddenLaye{r}_{in}\ge 0{)}^T\cdot X \\ {▽}_{W}f(W)=\frac{1}{N}{X}^T\cdot ((probs-Y)\cdot (W2{)}^T)\ast 1(hiddenLaye{r}_{in}\ge 0) \end{gathered}$$

求隐藏层偏置向量梯度

$$d(dataloss)=tr(\frac{1}{N}\sum _{i=1}^N(W2\cdot (prob{s}^T-Y^T){)}^T\ast 1(hiddenLaye{r}_{in}\ge 0{)}^T\cdot db)$$

$$\begin{gathered} D_{W}f(W)=\frac{1}{N}\sum _{i=1}^N(W2\cdot (prob{s}^T-Y^T){)}^T\ast 1(hiddenLaye{r}_{in}\ge 0{)}^T \\ {▽}_{W}f(W)=\frac{1}{N}\sum _{i=1}^N((probs-Y)\cdot (W2{)}^T)\ast 1(hiddenLaye{r}_{in}\ge 0) \end{gathered}$$

对偏置向量还需要注意维数，求和批量数据的偏置向量梯度

# finally into W,b
dW = np.dot(X.T, dhidden)
db = np.sum(dhidden, axis=0, keepdims=True)

第八步：迭代计算，反向传播，计算正则化梯度

$$\begin{gathered} regLoss=0.5\cdot reg\cdot ||W|{|}^2+0.5\cdot reg\cdot ||W2|{|}^2 \\ d(regLoss)=reg\cdot W\cdot dW+reg\cdot W2\cdot dW2 \end{gathered}$$

# add regularization gradient contribution
dW2 += reg * W2
dW += reg * W

第九步：迭代计算，反向传播，更新权重矩阵和偏置向量

# perform a parameter update
W += -step_size * dW
b += -step_size * db
W2 += -step_size * dW2
b2 += -step_size * db2

相关资料

1. The Matrix Cookbook - Mathematics
2. Matrix calculus