Changed FAQ for saving and loading objects

docs/docs/faq/question_05.md

@@ -12,7 +12,9 @@ with open("/path/to/my_posterior.pkl", "wb") as handle:
     pickle.dump(posterior, handle)
 ```
 
-Note: if you try to load a posterior that was saved under `sbi v0.14.x` or earlier under `sbi v0.15.0` or later, you have to add:
+Note: posterior objects that were saved under `sbi v0.17.2` or older can not be loaded under `sbi v0.18.0` or newer.
+
+Note: if you try to load a posterior that was saved under `sbi v0.14.x` or earlier under `sbi v0.15.x` until `sbi v0.17.x`, you have to add:
 ```python
 import sys
 from sbi.utils import user_input_checks_utils
@@ -22,35 +24,16 @@ sys.modules["sbi.user_input.user_input_checks_utils"] = user_input_checks_utils
 to your script before loading the posterior.
 
 
-`NeuralInference` objects are not picklable. There are two workarounds:
-
-- Pickle with [`dill`](https://pypi.org/project/dill/) (has to be installed with `pip install dill` first)
-```python
-import dill
-
-inference = SNPE(prior)
-# ... run inference
-
-with open("path/to/my_inference.pkl", "wb") as handle:
-    dill.dump(inference)
-```
-
-- Delete un-picklable attributes and serialize with pickle. Using this option, you will not be able to use the `retrain_from_scratch` feature and you can only use the default `SummaryWriter`.
+As of `sbi v0.18.0`, `NeuralInference` objects are also picklable.
 ```python
 import pickle
 
-inference = SNPE(prior)
 # ... run inference
-
 posterior = inference.build_posterior()
-inference._summary_writer = None
-inference._build_neural_net = None
+
 with open("/path/to/my_inference.pkl", "wb") as handle:
     pickle.dump(inference, handle)
 ```
-Then, to load:
-```python
-with open("/path/to/my_inference.pkl", "rb") as handle:
-    inference_from_disk = pickle.load(handle)
-inference_from_disk._summary_writer = inference_from_disk._default_summary_writer()
-```
+However, saving and loading the `inference` object will slightly modify the object (in order to make it serializable). 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.
+2) When the loaded object calls the `.train()` method, it generates a new tensorboard summary writer (instead of appending to the current one).