About reproduction about openibl

[yxgeee]

1. The loss seems normal. The convergence may be slow in later epochs.
2. The best model selected by validation results may not achieve the optimal performance on the test set. The model reported in the paper was selected from the last epoch of the 4th generation. Since there may exist training randomness, it is recommended to test the five checkpoints in the last generation and choose the best-performing one.
3. If you use the default settings on one GPU, only one triplet will be adopted for training in each mini-batch. Try to modify --tuple-size in the training scripts to adopt more triplets on one GPU for training. In my experiments, I adopted 4 GPUs and one triplet on each GPU, thus a batch of 4 triplets is used. In case the GPU memory is not enough for 4 triplets on only one 2080TI, maybe you need to decrease the learning rate to fit your batch size.

[zhangpj]

@yxgeee All right, thanks for your suggestion, I will have a try.

[zhangpj]

@yxgeee Hi, in your paper, you also evaluated SFRS on Oxford 5k, Paris 6k and Holidays datasets, can you share source codes that evaluate SFRS on those datasets, or tell me how do you evaluate SFRS on Oxford 5k, Paris 6k and Holidays datasets?

[yxgeee]

My colleague helps me test SFRS on retrieval datasets, and maybe I will merge these code in this repo after re-organization. We strictly follow all the same settings (e.g. image size, augmentation, etc.) as SARE and NetVLAD. So you could also refer to their code for evaluation details.

[zhangpj]

Ok, thank you.