I have two instances of the same model, living on different devices. The inference results differ. Is it possible to compare the models “step-by-step” to figure out which calculation is faulty?
One possible approach would be to use forward hooks and to compare the intermediate outputs of each layer. This post gives you an example on how to register the hook and how to print it.
You could use different
dicts or store the desired outputs using different keys etc.
This sounds very good, thank you!