Open cxy0616 opened 3 years ago
If I remember correctly I used 25000 iterations for AID since the dataset is larger than UCMD
Thank you for your reply. I think the selection method of triples in the code is randomly selected offline. Each batch_size has 30 triples. Have you tried to select batch_all or batch_hard online?
Yes they are always randomly selected. But it's an interesting suggestion to do hard mining. Please let me know if you find improvements using batch_all or batch_hard. Thanks.
Did you iterate the network model 10,000 times on the UCMD and AID datasets? If not, can you tell me the number of iterations when converging on these two data sets?