Concatenate and merge info, preserving conflicts as lists.

[joseppinilla]

If there is interest in merging info when concatenating samplesets since right now all info is ignored.
This preserves conflicts by listing them, but squeezes unique values.

[codecov-commenter]

Codecov Report

Merging #691 into master will increase coverage by 0.01%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #691      +/-   ##
==========================================
+ Coverage   91.85%   91.86%   +0.01%     
==========================================
  Files          60       60              
  Lines        4222     4229       +7     
==========================================
+ Hits         3878     3885       +7     
  Misses        344      344              

Impacted Files	Coverage Δ
dimod/sampleset.py	94.32% <100.00%> (+0.09%)	⬆️

---

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 2ec6944...370123d. Read the comment docs.

[arcondello]

Hmm, I am worried that the conflict resolution would lead to inconsistent/confusing results. I wonder if we could combine it with #643 by always making a list, even when the values are identical?

[randomir]

This is a very nice feature, but I share @arcondello's concern about consistency.

I suggest adding an optional info_merge_strategy argument (or similar) which would control how info fields are combined:

• "squeeze" could be your current implementation,
• "list" would always form lists,
• "drop" would skip the merge altogether,
• "inplace" or "recursive" would recursively combine all info dicts,
• custom callable would allow merge delegation to user's custom function of n dict args.

Inplace/recursive merge might be an overkill, so we can drop that one for now.

In addition, you'd probably want to make copies before merging them, due to dict's mutability.

[joseppinilla]

I was considering one option flag, e.g. squeeze_info=False which would either behave like "list" by default or "squeeze" if True.

I see now how "drop" would allow compatibility with the current implementation, but I'm not sure if adding more flexibility is necessary.

[randomir]

I prefer categorical option to boolean because it allows future expansion. Accepting a callable (in addition) is trivial and provides an absolute flexibility.

[joseppinilla]

One reason I didn't originally implement "list" is because I had to decide whether lists are always len(samplesets), e.g filled with None, or if they only list values of existing keys.

I think I'd go for the latter, even though it wouldn't allow tracking.

[randomir]

IMO, "list" should include fields from all samplesets. We might consider splitting "list" into "list-all" and "list-existing", although "squeeze" is conceptually already very similar to "list-existing". @arcondello, thoughts?

[joseppinilla]

squeeze may not be the right name since it's not the same behaviour as numpy squeeze, it only lists conflicts... so... "list-conflicts"?

[randomir]

I don't see how numpy.squeeze is relevant here, but I agree it's too general. unique sounds much better.