Thanks @samskalicky, this would sound like an interesting feature for Sockeye as well. In Sockeye, we use a mix of Hybrid and non-Hybrid blocks, both during training and inference. We currently do not use the export mechanism of Gluon, but keep our own configuration file to build the model graph at inference time, according to its specifications, we therefore tie code and model artifacts closely together. An export functionality that would allow us to export the inference computation after training sounds useful.
My main question about this though would be how exported HybridBlocks (i.e. SymbolBlocks) behave w.r.t changing input shapes? Before Gluon, we had to make heavy use of the BucketingModule in MXNet to account for varying shapes at inference time. -- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: https://github.com/apache/incubator-mxnet/issues/19535#issuecomment-727853073
