my trained model cannot achieve the same performance compared with the provided trained model

[pipipopo]

I use the default training setting in the code. For StanfordOnlineProduct dataset I got:
stanford online products mean recall@ 1 : 0.632077
stanford online products mean recall@ 10 : 0.785517
stanford online products mean recall@ 100 : 0.891618
stanford online products mean recall@ 1000 : 0.959900

How can I train a model with comparable performance?

[PkuRainBow]

Please check the test code https://github.com/PkuRainBow/Hard-Aware-Deeply-Cascaded-Embedding_release/blob/master/src_code/test_deep_fashion.py.

[pipipopo]

I mean training. Test is OK. For the provided model I got
stanford online products mean recall@ 1 : 0.695367
stanford online products mean recall@ 10 : 0.843716
stanford online products mean recall@ 100 : 0.928379
stanford online products mean recall@ 1000 : 0.975735

But when I train by myself, the performance of final model after 60000 iters is:
stanford online products mean recall@ 1 : 0.632077
stanford online products mean recall@ 10 : 0.785517
stanford online products mean recall@ 100 : 0.891618
stanford online products mean recall@ 1000 : 0.959900

So a gap exsits. I use the default solver prototxt and batch size...

[PkuRainBow]

@pipipopo Get it. Try to change the hard_ratio from {1.0, 0.5, 0.2} to {0.5, 0.2, 0.1}. Also please pay attention to the sampling methods. You need to sample 2 big classes firstly and then sample small classes, which is detailed in https://github.com/PkuRainBow/Hard-Aware-Deeply-Cascaded-Embedding_release/blob/master/src_code/sample_stanford_products.py

[zimenglan-sysu-512]

hi @PkuRainBow,
why try to change the hard_ratio?

[zhengxiawu]

@pipipopo I trained on CARS, have the same question...

[PkuRainBow]

@zhengxiawu Please try hard ratio settings : {0.5, 0.2, 0.1}, which can mine better hard examples.