Conv. auto-encoders using mirroring

[patnicolas]

I am working on automating the generation of convolutional auto-encoders/GANs (Java/Scala). The plan is to generate the corresponding de-convolutional (Transposed) automatically from existing block builders.
I am new to DJL but I have not seen such capability in 0.17.0? If not, is there a plan add such functionality in a future release?
I enjoy this 'lego' design to build neural nets..

[zachgk]

We do have a Deconvolution class that you can use. The use of GAN may be a bit tricky with Trainers, but you can simply avoid the Trainer and use the underlying pieces (GradientCollector) directly.

Is there a particular area that you are having problems?

[patrick-nicolas]

Very sorry for this late reply. I was able to resolve the issue looking at the Gradient collector code. Great library!