SGDA Requires Validation Set
BrianMiner opened this issue · 7 comments
Can a validation set be used? It is a requirement for SGDA but doesnt appear to be an option
I don't have much experience with SGDA - actually never used it - since I've only used MCMC. Hence, I haven't implemented the validation set option.
Could you give me some info on the structure of the validation set, I could quickly implement that option.
I believe it is simply a data set of the same structure as train and
test. If you look at the libfm manual under sdga you will see a
validation set mentioned as a requirement. If memory serves, SGDA uses
the validation set to "adapt" the learning rate of SGD.
On 12/28/2015 04:47 AM, João Ferreira Loff wrote:
I don't have much experience with SGDA - actually never used it -
since I've only used MCMC. Hence, I haven't implemented the validation
set option.Could you give me some info on the structure of the validation set, I
could quickly implement that option.—
Reply to this email directly or view it on GitHub
#2 (comment).
I've uploaded a new version with this change (also updated README).
Could you please test this change? (make sure its version 0.8.1 . Use pip --no-cache-dir install pywFM)
I'm wasn't sure if the validation set should be placed as a FM class parameter or as a FM.run function parameter. My intuition said that its was best to include it in the FM.run since the validation set changes according to the train/test data. Feel free to correct me.
Thank you,
PS: if you feel like it, adding a simple-sgda example would be awesome!
@BrianMiner Have you been able to test this change?
Thank you,
I have not yet, I will try asap!
hey @jfloff, thanks for the wrapper. I've noticed that parameters were being ignored for methods sgd / sgda, which were resulting in the algorithms not converging
fix for this is within the init, you're casting some decimal values into integers with %d, so if I were to put in r2_regularization=0.01, its translated as 0. there's also a spot when you're building the args for the subprocess where the learning rate could again be reduced to an integer as well.
switching those formats to something like %.5f fixes this and the call to libfm runs as expected.
edit: also, adding the -regular argument for sgda's is unnecessary since its trying to optimize those values after each iteration