Make inference object serializable

[michaeldeistler]

Resolves #476

Saving the inference object was always a bit messy. But dill used to work.

I just tried and could no longer save the inference object with dill. The workaround in this PR is to set the non-serializable objects to None before saving. This allows to save the inference object with pickle. These modifications lead to the following two changes in behaviour:

1. Retraining from scratch is not supported, i.e. .train(..., retrain_from_scratch=True) does not work after loading.
2. When the loaded object calls the .train() method, it generates a new tensorboard summary writer (instead of appending to the current one).

[codecov-commenter]

Codecov Report

Merging #617 (23f3225) into main (4b3d260) will decrease coverage by 0.13%.
The diff coverage is 16.66%.

@@            Coverage Diff             @@
##             main     #617      +/-   ##
==========================================
- Coverage   66.77%   66.63%   -0.14%     
==========================================
  Files          66       66              
  Lines        4307     4319      +12     
==========================================
+ Hits         2876     2878       +2     
- Misses       1431     1441      +10     

Flag	Coverage Δ
unittests	66.63% <16.66%> (-0.14%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
sbi/inference/base.py	72.66% <16.66%> (-4.87%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 4b3d260...23f3225. Read the comment docs.

[janfb]

Thanks! Looks good in itself, but I am not sure this is a long-term solution. Isn't there a way to serialise the full object without this workaround? We can merge this for now but I think it would be good to keep an issue open so that we find a sustainable solution later.

[janfb]

could this be relevant?
uqfoundation/dill#348

[michaeldeistler]

I don't think the tensorboard writer can be serialized. See here

uqfoundation/dill#348 seems relevant, but they do not provide a fix either, no?