Problem about the released best_model on Porto

[LIWEIDENG0830]

Hi boathit, I plan to use your trained model on Porto. However, I meet a problem that I can't reproduce your results by using your codes and your released best_model. The performance of your experiment 1 (querysize=1000, dbsize=100000) in terms of mean rank is 12152.9840. My versions of python and pytorch are 3.6 and 1.5. I wonder know that the best_model is saved on which version and whether this fault is caused by version. P.S. I absolutely follow your script to do the experiment. Moreover, there are no any errors when I run the codes.

[mofaCatt]

Hi wei deng, I met the same problem as you, I would like to know if you have solved this. Thank you so much!

[LIWEIDENG0830]

Hi wei deng, I met the same problem as you, I would like to know if you have solved this. Thank you so much!

Hi mofaCatt, Sorry for lately reply. I solve this problem through training the model from beginning. I think there may be some problem about the uploaded model by the authors. Hope this message can solve your problem. Best regards, Liwei Deng

[Sibyl233]

Hi wei deng, I tried to train the model from the beginning but still has a gap with the author's results. Would you like to share the weight of the best model or the training strategy if possible? Take experiment-1 for example, I got results as follows:

mean rank: 2.386 with dbsize: 20000 mean rank: 3.621 with dbsize: 40000 mean rank: 5.031 with dbsize: 60000 mean rank: 6.722 with dbsize: 80000 mean rank: 8.101 with dbsize: 100000