Parallel py3

[mxndrwgrdnr]

• updated synthpop for python 3
• updated the census attributes retained in starter2.py
• updated ipu.py so that rather than erroring out when max_iterations is reached, it generates a warning.
• added method synthesize_all_in_parallel() to synthesizer.py for processing sub-county geographies in parallel.

Could use an extra set of eyes at L264-265 in synthesizer.py. These two lines account for the fact that when processing geographies in parallel, draw.draw_households() cannot rely on the hh_index_start argument to set the household_id either as the index of the synthetic households table or the hh_id field of the synthetic persons table. Thus we have to adjust the households index and persons hh_id field AFTER all the tables have been generated, which is what's hapenning at L264-265.

[coveralls]

Coverage decreased (-1.2%) to 81.643% when pulling c180116 on parallel_py3 into d1de14a on master.

[janowicz]

Thanks very much for this great contribution @mxndrwgrdnr! We'll review this soon, sorry for the delay. There is an existing earlier open PR that was reveiewed, which we kept open for a bit in the case anyone from the community wanted to review, that we'll merge in first. And then we can help reconcile with this PR.

[cvanoli]

Why are you adding these variables if they are not used in the synthesis?

[mxndrwgrdnr]

They will be.

[mxndrwgrdnr]

@cvanoli @janowicz this should be good to go now folks.

[janowicz]

Awesome, thanks @mxndrwgrdnr! I'll check it out today

[cvanoli]

thanks !! @mxndrwgrdnr

[cvanoli]

Looks good to me @mxndrwgrdnr

[cvanoli]

In a conv with @janowicz we discussed that it would be better if this ignoring of errors was made optional. Could we make an ignore parameter, where, if set to True, would raise warnings instead of errors?

[cvanoli]

Idem comment to ipu.py line263.

[cvanoli]

I tested the branch with a local environment in my computer and two Future Warnings keep coming up:
1 -synthpop\ipu\ipu.py: Line191

RuntimeWarning: divide by zero encountered in double_scalars
adj = constraint / float((column * weights).sum())

2 - synthpop\census_helpers.py: Line 180:

FutureWarning: Sorting because non-concatenation axis is not aligned. A future version of pandas will change to not sort by default. To accept the future behavior, pass 'sort=False'. To retain the current behavior and silence the warning, pass 'sort=True'.
pums = pd.concat([pums, pums00], ignore_index=True)

[mxndrwgrdnr]

@janowicz the second FutureWarning that @cvanoli mentions isn't a big deal and should resolve itself in the future. The first FutureWarning, however, is happening during IPU when household weights that are very nearly zero are not close enough to zero to be caught by ipu._drop_zeros but close enough to trigger a divide by zero RuntimeWarning. It's just a warning and not an error so I don't think it is effecting performance but I wanted to bring it to attention and see if you had an opinion.