Changing conv input tensors from 'Constant' to 'Initializer' and populating the graph initializer

[ecenurustun]

I have a previously exported model (<path_to_file>/model.onnx) and after loading it I realized that its graph initializer list (model.graph.initializer) is empty because weights of Conv layers are somehow saved as "Constant" instead of "Initializer". Please see the Netron screenshot below for one such tensor. And onnx2torch fails because converter will iterate over the empty initializer list. So my question is, how can I add these tensors to graph initializer? I can access their names through node.input[1] but how about their values? Will it require onnxruntime? Thank you.

(I asked this question here too in case someone already has a solution.)