I'm not sure if this is a proper place to ask such a question - if not, please guide me to a place where I can find some help.
Is it possible now to implement Variational Autoencoders in Julia using MXNet.jl ?
Cost/error function is composed of two components there - one is placed at the very end of computation graph which is probably possible to implement but the other cost function component is "attached" somewhere in the middle of the graph - (KL Divergence between parametrized Gaussian and Normal Distribution)
Little bit more generic question could emerge here: Is it possible to have a computation graph that is not a chain (single path of computations) but more complex structure like the network for VAE ?
Hello there,
I'm not sure if this is a proper place to ask such a question - if not, please guide me to a place where I can find some help.
Is it possible now to implement Variational Autoencoders in Julia using MXNet.jl ?
Cost/error function is composed of two components there - one is placed at the very end of computation graph which is probably possible to implement but the other cost function component is "attached" somewhere in the middle of the graph - (KL Divergence between parametrized Gaussian and Normal Distribution)
For details take a look at: http://arxiv.org/pdf/1606.05908v1.pdf (Figure 4 - right part of it)
Little bit more generic question could emerge here: Is it possible to have a computation graph that is not a chain (single path of computations) but more complex structure like the network for VAE ?
Thanks a lot,