Hashsum of label models inconsistent between runs

[hlibbabii]

Commit example:
giganticode/bohr@3dfd353

Even if the dependencies of the label model training stage don't change, rerunning the training reruns in a label model that has a different hash. The root cause is that when the model is serialized, some of its attribute (pytorch tensors) are serialized into a different sequence of bytes every time, still not sure about the reason. So I asked a question about this:
https://discuss.pytorch.org/t/tensor-pickling-inconsistent-between-runs/122533
https://stackoverflow.com/questions/67710297/pytorch-tensor-pickling-inconsistent-between-runs