I’ve trained a model and I’m interested in using it for prediction on held-out data, on a different machine (without re-training). During training I made sure the training data is in a shared tensor. I’ve pickled the model, the trace and the shared tensor (necessary for pm.sample_posterior_predictive).
In a different session, I’ve pickle-loaded them, and changed the values of the shared tensor (following this tutorial). However, running pm.sample_posterior_predictive returns predictions for the original training data, as if the values of the shared tensor did not change.
When I’m doing shared_tensor.set_value(...) in the training session (on the same machine, with the same running python kernel), it does work; maybe the theano graph isn’t propely pickled with the model? Any way around this?