search

DagnyT / hardnet

Hardnet descriptor model - "Working hard to know your neighbor's margins: Local descriptor learning loss"
MIT License
514 stars 100 forks source link

how to evaluate the Liberty using Notredame #10

Closed hurricane2018 closed 6 years ago

hurricane2018 commented 6 years ago

image

The test accuracy is always change. Did you get the average precision of the test data ?

hurricane2018 commented 6 years ago

I mean, If we train 10 epoch, we can get a precision. If we train 20 epoches, we can get another result. How did you solve this problem ?

hurricane2018 commented 6 years ago

how many epoches did you use to evaluate a dataset ?

DagnyT commented 6 years ago

Hi! To get exactly same results from run to run, please, set up torch.backends.cudnn.deterministic = True, num_workers=0. We've trained for 10 epochs and evaluated results after the last one, also, take a look at the adjust_learning_rate function, which determines the learning rate policy. If you have any other questions, don't hesitate to ask.

hurricane2018 commented 6 years ago

Yeah, thank you for your reply.

But I get different results every time. I mean I can get a result this money, but tomorrow a slight different result . I mean every time when I run your code, I can get a different result.

So did you run your program several times, and average them ?

2018-04-16 20:33 GMT-04:00 Anastasiia notifications@github.com:

Hi! To get exactly same results from run to run, please, set up torch.backends.cudnn.deterministic = True, num_workers=0. We've trained for 10 epochs and evaluated results after the last one, also, take a look at the adjust_learning_rate function, which determines the learning rate policy. If you have any other questions, don't hesitate to ask.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/DagnyT/hardnet/issues/10#issuecomment-381792037, or mute the thread https://github.com/notifications/unsubscribe-auth/AifZ-bLQEsds4dmzxA3mWY47CiY_fLBXks5tpThFgaJpZM4TVQHx .

DagnyT commented 6 years ago

We didn't average the results during the evaluation. To get deterministic results from run to run, just add torch.backends.cudnn.deterministic = True to your code (check the last commit, I've added setup to get exactly same results from run to run).