Deep learning classification

User 117 | 10/20/2014, 11:13:15 AM

Hello,

in the guidelines for the new deep learning module it is mentioned that deep learning models can only be created for "image type" input.

Could you clarify how GL detects whether the underlying data are of image type?

Is there any way to create a deep learning model for classification when the features are numerical but do not come from images?

Comments

User 14 | 10/20/2014, 5:23:05 PM

Hi,

graphlab.Image is a valid datatype for SArray, along with int, float, array.array, dict. You can create an SArray/SFrame of images using graphlab.imageanalysis.loadimages. You can convert numerical data to image data using SArray.pixelarrayto_image, and image data to numerical data using SArray.astype().

DeepLearning does accept numerical data, but ConvolutionLayer can only be applied to image data for now.