-
Notifications
You must be signed in to change notification settings - Fork 18.7k
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
RMSprop implementation based on G. Hinton Lecture 6 #1890
Conversation
Thanks for the solver -- RMSprop has been on our list so we'll take a look when we can. Please drop the logs from versioning and instead post a gist in your PR description for us to read. |
Sorry I forgot to remove these logs but now they are gone. |
I add rmsprop example to mnist example folder also I ll give gist link as you liked. |
@erogol, I have tested it and it works quite fine :) |
@ducha-aiki yeah :) thank you... However there are some compilation issues I guess pointed by Travis. |
Are there any plans on porting this to the newest master branch? |
@@ -628,7 +628,7 @@ void NesterovSolver<Dtype>::ComputeUpdateValue() { | |||
net_params[param_id]->cpu_diff(), momentum, | |||
this->history_[param_id]->mutable_cpu_data()); | |||
|
|||
// compute udpate: step back then over step | |||
// compute uppate: step back then over step |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Well, you almost fixed the typo :-)
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
yeah :)
@erogol ,
|
I will take care of it in one of the available times but in a very buzzy period right now. |
For those interested, I've merged this PR against my own slightly modified Caffe tree which is up to date with current master, and it appears to work fine on a few applications I've tested. However, two points:
To reproduce the merge against master, you should be able to use the last two commits from https://github.com/beniz/caffe/tree/master_dd_integ_rmsprop |
I implemented RMSprop as Hinton suggests and alternatively you can apply momentum as well with RMSprop suggested by http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf
From my experiments I observed much better convergence rate compared to other implemented solver methods. Here's my another toy experiment if you like to see more empirical results. http://www.erogol.com/
Any comments are very welcome.
Here is the mnist example output, https://gist.github.com/erogol/5c5b7beae8b088bd00e9