Missing values imputation implementation #743
Conversation
This has a couple of advantages over the Text backend. 1. The values are stored as in a table per chain, not a file per variable. This is easier to inspect and work with directly if desired (e.g., with pd.read_csv). 2. Values are stored during sampling, not kept in memory.
The main reason to keep the Text backend around was that pandas was an optional dependency. Now that this is no longer the case, the original Text backend doesn't offer any advantages over the CSV backend. (Even if someone prefers not to write to files while sampling, they can sample with the NDArray backend and then use the CSV dump function.) Rename CSV to Text. This name is more appropriate because the values being stored as plain text is the important feature, not which delimiter is used.
Commit 85b53d4 changed the location of the test sqlite file from the PWD to /tmp. (I believe this was because, due to permission issues, the file was not being removed, and subsequent tests were failing as a result). Use tempfile.gettempdir to be more portable.
With the introduction of missing variables it becomes important to be able to use the value of an ObservedRV. However, this was not possible because ObservedRV didn't subclass TensorVariable because ObservedRV could also handle multiple observed arrays so wouldn't always be a single tensor. This splits up the two cases into ObservedRV, for single observed arrays, and MultiObservedRV for multiple observed arrays. ObservedRV now inherits from TensorVariable.
|
I'm actually not able to replicate this issue. I have the latest Theano master. ERROR: pymc3.tests.test_shared.test_deterministicTraceback (most recent call last): |
|
Managed to replicate. |
Missing values imputation implementation
2 participants
No description provided.