@alex.veuthey I don’t. What I do is I forward propagate model A (an autoencoder), and subsequently use the latent hidden layer as input for model B (everything online?). I finally got time to keep working again on this and am to the point where I concatenate the parameters of the models in one single list.