Softmax损失值及梯度推导

Softmax损失值使用的是交叉熵函数；
交叉熵损失函数的计算方式如下：

而在CS231n作业中，正确标签的值为1，因此其交叉损失值为：

也可以写成：

通过对损失函数进行求导，可以得到损失值对权重的导数为;

python 代码如下：

def softmax_loss_naive(W, X, y, reg):
    # Initialize the loss and gradient to zero.
    loss = 0.0
    dW = np.zeros_like(W)
    
    num_train = X.shape[0]
    num_classes = W.shape[1]
    F = X.dot(W)
    normalized_F = F - np.max(F, axis=1).reshape(num_train,1) 
    # 每一行减去每一行的最大值，而每一行即代表一组数据
    exp_normalized_F = np.exp(normalized_F)
    
    # compute loss
    for i in range(num_train):
        s_yi = exp_normalized_F[i][y[i]]
        sum_i = np.sum(exp_normalized_F[i])
        loss -= np.log(s_yi*1.0 / sum_i)
    
    loss /= num_train
    loss += reg*np.sum(np.square(W))
    
    # compute dW
    for i in range(num_train):
        sum_i = np.sum(exp_normalized_F[i])
        for j in range(num_classes):
            dW[:, j] += (exp_normalized_F[i][j]*1.0 / sum_i)*X[i]
            if j == y[i]:
                dW[:, j] -= X[i] # 对于j==y[i]这种情况，根据公式，还需要额外的减去X[i]
            
    
    dW /= num_train
    dW += 2*reg*W
    
    pass

    # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****

    return loss, dW

向量化计算方法

def softmax_loss_vectorized(W, X, y, reg):
    """
    Softmax loss function, vectorized version.

    Inputs and outputs are the same as softmax_loss_naive.
    """
    # Initialize the loss and gradient to zero.
    loss = 0.0
    dW = np.zeros_like(W)

    num_train = X.shape[0]
    num_classes = W.shape[1]
    F = X.dot(W)
    exp_normalized_F = np.exp(F - np.max(F, axis=1).reshape(num_train,1))
    # compute loss
    sum_i = np.sum(exp_normalized_F, axis=1)
    p_i = exp_normalized_F[range(num_train), y] / sum_i
    L_i = - np.log(p_i)
    loss = np.sum(L_i)
    
    loss /= num_train
    loss += reg*np.sum(W * W)
    
    # compute gradient
    acc_effect = exp_normalized_F / sum_i.reshape(num_train,1)
    acc_effect[range(num_train), y] -= 1.0 # 正确标签位置还需要额外减去1
    dW = X.T.dot(acc_effect)
    
    dW /= num_train
    dW += 2*reg*W
    
    pass

    # *****END OF YOUR CODE (DO NOT DELETE/MODIFY THIS LINE)*****

    return loss, dW