Closed zhanjw closed 5 years ago
I'm sorry, in this implementation, some points are differs from original implementation such as method of batch making, architecture and number of units, pre-training of CNN and so on. And I did not test my implementation on flicker-25k because I couldn't get experiments data from Baidu web site. If you want to reproduce the result, I think referencing https://github.com/WendellGul/DCMH is better. Thank you for asking!
thanks for your assisstance anyway, i'll try.
Hello, I would like to ask if you can reproduce the result of the paper? After 50 epoch training using your code and flicker25k datasets, my query mAP of cross-modal is only about 4.5%, and the training mAP only get 7%. I don't know what went wrong. It seems like completely impossible to fit. Can you help me?