first of all I want to say that the new deep learning integration (d4j) works fantastic. Really great work done there!
I just wanted to ask if it is planned to integrate deconvolution/unpooling layers in the near future. That could than be used for semantic segmentation of images.
Currently, we only support layer types available by DL4J and deconvolution/unpooling is not available at the moment. As far as I know, they are not planning to add this in the near future. However, I will note that as a feature request so we could maybe add that in the future release.
Don't mean to hijack the thread but we would also like to have the ability for deconv/unpooling for sematic segmentation.
On a separate note, is it possible to use DL4J for "transfer learning" by reusing existing trained models (VGG, GoogLeNet, ResNet) from other framework(s) (e.g., Caffe/MatConvNet)?
thanks for your interest. I noted your feature request.
Transfer learning is already possible using the integration. Unfortunately, only using existing models that were created by the integration itself. Currently there is no Node to import models from other frameworks but that feature is planned for the future.