论文复现 ld = genLD(b_l, sigma, ‘klloss‘, 65)

 

        for step, (b_x, b_y, b_l) in enumerate(train_loader):   # gives batch data, normalize x when iterate train_loader
        # img, label, lesion
            b_x = b_x.cuda()
            b_l = b_l.numpy()

            # generating ld
            b_l = b_l - 1
            ld = genLD(b_l, sigma, 'klloss', 65) # genLD函数生成标签分布ld，这里使用了sigma和其他参数。ld可能是一个连续的概率分布，用于后续的KL散度损失计算。

# generate label distribution
def genLD(label, sigma, loss, class_num):
    label_set = np.array(range(class_num))
    if loss == 'klloss':
        ld_num = len(label_set)
        dif_age = np.tile(label_set.reshape(ld_num, 1), (1, len(label))) - np.tile(label, (ld_num, 1))
        ld = 1.0 / np.tile(np.sqrt(2.0 * np.pi) * sigma, (ld_num, 1)) * np.exp(-1.0 * np.power(dif_age, 2) / np.tile(2.0 * np.power(sigma, 2), (ld_num, 1)))
        ld = ld / np.sum(ld, 0)

        return ld.transpose()