Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Learning rate 'Exponential Decay' policy wrong. #662

Closed
TimZaman opened this issue Mar 29, 2016 · 9 comments · Fixed by #663
Closed

Learning rate 'Exponential Decay' policy wrong. #662

TimZaman opened this issue Mar 29, 2016 · 9 comments · Fixed by #663

Comments

@TimZaman
Copy link
Contributor

Version v3.2.0-61-g0ce28b9
I often have issues with the learning rate and the decay. Sometimes it tells me it cannot process 'nan', sometimes it's correct, sometimes it jitters a bit up and down, but most often it does this. I can reproduce this only half of the time.
Base learning rate: 1.0
Policy: Exponential Decay
Gamma: 0.95
screen shot 2016-03-29 at 3 16 24 pm

Then this happens:
screen shot 2016-03-29 at 3 16 11 pm

@gheinrich
Copy link
Contributor

Thanks for the bug report. The learning rate seems to be decreasing way too quickly indeed. Are you using Torch or Caffe? I think I can reproduce your issue with Torch.

@TimZaman
Copy link
Contributor Author

Torch indeed. Did not try Caffe.

gheinrich added a commit that referenced this issue Mar 29, 2016
"exp", "inv" and "sigmoid" weren't operating as expected

close #662
@gheinrich
Copy link
Contributor

@TimZaman do you mind verifying that #663 fixes the issue? Thanks!

@TimZaman
Copy link
Contributor Author

Yes. Merci beaucoup Greg!

@TimZaman
Copy link
Contributor Author

Hmm even after the update, still something fishy here: (500 epochs used)
screenshot 2016-03-31 15 34 53
screenshot 2016-03-31 15 34 59

@TimZaman TimZaman reopened this Mar 31, 2016
@gheinrich
Copy link
Contributor

I just tried running 500 epochs of training. This is what the learning rate looked like:
lr-exp-decay-500epochs
I was using this version:

$ git describe
v3.2.0-73-gb4dd9c7

@gheinrich
Copy link
Contributor

@TimZaman were you able to verify that you were using the right version of DIGITS? The exponential decay is working for me for 500 epochs.

@gheinrich
Copy link
Contributor

No response and this is working for me on v3.2.0-73-gb4dd9c7 so I am closing the bug.

@TimZaman
Copy link
Contributor Author

i was using the most up to date version for sure. There's something going on with the learning rate for sure, i will try to hunt down the cause next time instead of posting the problem.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging a pull request may close this issue.

3 participants