iter_size parameter in Solver Options

[Axel Straminsky]

Hello. Recently I had some problems with training my models, which I posted in the Caffe users group: https://groups.google.com/forum/#!topic/caffe-users/ap_jBpG45Ao

Based on the answer I got, I wanted to play with the iter_size hyperparameter in the Solver, but I found it wasn't available in "Solver Options", so I guess it defaults to 1. Is there any way to modify it from DIGITS?

Thanks!

[Luke Yeager]

Interesting, I wasn't aware of that feature. Thanks for letting me know!

It looks like the "iter_size" parameter was added in this pull request on May 29th - just 10 days after we last synced NVIDIA/caffe with BVLC/caffe (see information about that here).

So, we don't have that option in NVIDIA/caffe yet. I've created an issue to track this feature in NVIDIA/caffe:

https://github.com/NVIDIA/caffe/issues/16

Once we update our fork (no promises on when that will be), it will be trivial to expose this option in DIGITS.

[Axel Straminsky]

Thanks Luke for the quick response. I guess in the meantime I will have to train with raw BVLC/caffe whenever I encounter this issue and need to play with the iter_size hyperparameter.