Im working on dynamic networks and I leverage onnx parser in the heuristics -translating to a networkx representation over which I operate.
I had to rewrite it between torch 0.4 and 1.0 and I just ran in with v 1.1 and I found a couple breaking changes. port.isTensor() to port.isCompleteTensor() is one, im still digging for others.
I understand that this is a WIP and that the primary goal of this is probably network IO. I understand that Deep Nets themselves are a huge WIP so it is hard to predict what will change. Networkx gives me access to graph algorithms; it is itself a bit clunky but serviceable for now. Dynamic networks and much larger multimodal networks are on the horizon.
Is there a plan to to make a complete stable object oriented graph parser? Itd be great to have.