The MXNet deep learning framework offers HybridBlocks, which allow you to mix imperative programming (for debugging) and symbolic programming (for faster training). I have a network that I hybridized and ran, and would like to then "un-hybridize" the network to debug activation dimensions. But I couldn't find anywhere in the documentation that allows you to un-hybridize a compiled network. Is there anything in the MXNet framework that would allow me to accomplish what I'm attempting?
net.hybridize(active=False) should do the trick. Looks like we didn't mention in the hybridization tutorial. I created #18962 for tracking this.