PandasParallelLFApplier does not preserve the order of the rows

[hcvazquez]

Issue description

I'm using PandasParallelLFApplier to apply labeling functions to a pandas dataframe with 5000 rows.

Code example/repro steps

Using PandasParallelLFApplier

# Apply the LFs to the unlabeled training data
applier = PandasParallelLFApplier(lfs)
topic_labeling = applier.apply(df[:5000])
topic_labeling

output: array([
       [-1, -1, -1, -1],
       [-1, -1, -1, -1],
       [-1, -1, -1, -1],
       ...,
       [-1, -1, -1, -1],
       [-1, -1, -1, -1],
       [-1, -1, -1,  1]])

Same code using PandasLFApplier

output: array([
       [-1, -1, -1, -1],
       [-1,  1, -1, -1],
       [-1, -1, -1, -1],
       ...,
       [-1, -1, -1, -1],
       [-1, -1, -1, -1],
       [-1, -1, -1,  1]])

Second row is different.

Expected behavior

I would expect the same result for both. Labeling coverage and overlaps is the same for both. Because of that the problem has to be the order of the rows.

System info

• How you installed Snorkel (conda, pip, source):
• Build command you used (if compiling from source):
• OS:
• Python version: 3.6.8
• Snorkel version: 0.9.3
• Versions of any other relevant libraries: dask==2.8.1 pandas==0.25.3 numpy==1.16.4

[henryre]

Hi @hcvazquez, great question. This is due to index sorting, and isn't reflected well in the docs right now (but on our list to update). This was discussed on the Spectrum thread here: https://spectrum.chat/snorkel/help/how-to-use-the-pandasparallelapplier~cf50f563-28e6-418c-93a3-337384566c13

[henryre]

Closing for now, feel free to re-open!