SVM损失函数

Question

def svm_loss_naive(W, X, y):
"""
SVM loss function, naive implementation calculating loss for each sample
using loops.    

Inputs:
- X: A numpy array of shape (n, m) containing data(samples).
- y: A numpy array of shape (m, ) containing labels
- W: A numpy array of shape (p, n) containing weights.  
   
"""
# Compute the loss
num_classes = W.shape[0] # classes weights are in row wise fashion
num_samples = X.shape[1] # samples of unknown images are in column-wise fashion
loss = 0.0
delta = 1 # SVM parameter
for i in range(num_samples):
    scores = np.dot(W, X[:,i])
    correct_class_score = scores[y[i]] 
    for j in range(num_classes):       
        if j == y[i]:
            continue
        margin = max(0, scores[j] - correct_class_score + delta )
        loss = loss + margin

# Average loss
loss = loss / num_samples

return loss

根据我对python代码的理解

1. 我们首先通过将第1行的权重乘以第1样本列

来计算第1类的得分。

然后，我们将获取存储在数组y

中的ith示例的correct_class_score。

，

1. ，那么，我们在迭代类的数量(假设是3)，我不明白的是j == yi在做什么？我的意思是，当j在0到2之间时，当j等于yi时，yi只是ith示例correct_class_score

的索引。

我预先理解的其余代码

Answer 1

这是SVM丢失(数据丢失)函数的定义：

​

​

在内部求和中，明确排除了等于易的j指数。