Open lemonadeww opened 5 years ago
为什么不回答这个问题呢?
Hi, sorry for the delay response, cifa_KNN is generated by the fc7 layer of VGG-19. The noise does not inject during the training process. In the training stage, we still use the continuous value and the approximate sign function focuses on forcing the continuous vector to be close to binary codes so as to minimize the quantization error, however, which can not fully constrain the continuous vector into hash codes.
请问cifar_KNN.npz的特征是怎么得到的呢,由VGG-19提取的吗? 另外,代码和论文似乎也不对应,噪声并不是添加的二进制,训练的一直是连续的哈希,到最后用sign函数保存哈希码。