Learnable Descriptors for Visual Search