I have a question related to Pytorch 1.0. If I want to convert to ONNX, previous versions only supported a method which essentially exported a trace of the actual flow. This was fine for certain topologies (e.g. CNNs), but ones with loops (e.g. NMT) would end up exporting unrolled loops.
I read something about a “script mode” and that there is some support now for simple loops (among other new features). Does this mean it will be able to capture Python branch/loop constructs or do we have to recode these constructs into “PyTorch loops” (if there is such a thing) in order to properly capture the loop construct (according to ONNX standard) without unrolling loops?