I need to build some CNN's using very large numbers of photos, to classify which photos contain certain features.

I've run into a number of road blocks.

As much as I love both Linux and virtualization, I've found that Linux-only solutions like Berkeley's Caffe and NYU's pre-trained OverFeat don't work for large scale data aren't viable because virtualization has a lot of problems in an HPC and big data scenario, especially when the tools depend heavily on specific hardware and drivers to be effective (even if you have the hardware and drivers, getting it to work on a VM is something different).

I've found a number of solutions in R, Python, and ImageJ which work well on the small scale or with specialized types of photos like medical images, satellite or microscopic images, but not for real world photos with lots of "noise" and features which can vary a good bit between photos.

Is there a Windows-friendly or platform-agnostic scalable solution that's free and minimally complicated?

  • $\begingroup$ Have you tried installing Microsofts caffe port github.com/Microsoft/caffe ? I only tried cpu only mode without python though. $\endgroup$
    – Emil
    Apr 10 '16 at 10:25
  • $\begingroup$ What about mxnet? Though I haven't tried it myself, I heard it is very good mxnet.io $\endgroup$
    – Gil
    Feb 19 '17 at 8:53
  • $\begingroup$ @Gil It's a good idea, I've actually used mxnet for a few years via R, but for image classification it's not very straightforward or complete, like using TensorFlow, overfeat, caffe etc. Though having said that, it is possible, especially within Python and leveraging some additional libraries. $\endgroup$
    – Hack-R
    Mar 15 '17 at 13:37
  • $\begingroup$ @Emil That's a great idea, I had used Caffe on Linux but didn't realize there was a Windows port. I'm going to try it today. Feel free to make that an answer if you like. $\endgroup$
    – Hack-R
    Mar 15 '17 at 13:38

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.