Speeding up GoogLeNet

[Thomas Wood]

Hi

I have been training my own networks using GoogLeNet's architecture. I'd like to know if anyone has any tips for speeding up the execution of the network, especially when running on mobile platforms. 

Maybe there's an alternative to GoogLeNet that runs faster? Or any kind of modification that I should make to the prototxt to speed it up? Or do I need to modify the Caffe source code? 

I'm interested to hear your ideas!

Thanks

[Joshua Slocum]

You're probably SOL unless you're ready for a substantial software engineering project. A lot of optimization work has already gone in to making Caffe and the libraries it operates with run fast. 

One option is to try to use the GoogLeNet to train a deeper but thinner network with fewer parameters. You could also do some ablation studies to see if you can remove parts of the GoogLeNet without significantly compromising its performance on your tasks. 

--
You received this message because you are subscribed to the Google Groups "Caffe Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to caffe-users...@googlegroups.com.
To post to this group, send email to caffe...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/caffe-users/cc0f798b-7e0a-40ff-aec0-cf38715dbad8%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Thomas Wood]

OK thanks! I'll look into what modifications I can do to the network for now, and will try to get into modifying the Caffe source code if it's possible.

I suspected modifying Caffe would be a large project. In particular I had looked at a paper called Compressing Neural Networks with the Hashing Trick (http://arxiv.org/pdf/1504.04788.pdf) which suggests I could get 2x speed improvement by switching from the array representation to a hash representation.