Closed ngsankha closed 2 years ago
ngsankha
commented
2 years ago Constants are handled as of commit f9611a4a72877e34e3214cead2fdcf75ff7bbfd7.
The index of rows/columns are not yet enabled. It might help eliminate more programs by abstract interpretation.
ngsankha
commented
2 years ago Intermediate benchmarking results after enabling constants and running the tool on each benchmark with 20min timeout (same as AutoPandas paper). We do 17/29 benchmarks vs 17/26 on the original paper. I suspect the numbers to be better after the rows/columns domain.
SO_49581206_depth3
------------------
ERROR!
SO_13576164_depth3
------------------
ERROR!
SO_14023037_depth3
------------------
ERROR!
SO_23321300_depth3
------------------
ERROR!
SO_13807758_depth2
------------------
arg0.dropna().reset_index(drop=True)
48.4200041539998
SO_49567723_depth2
------------------
ERROR!
SO_11811392_depth3
------------------
arg0.T.reset_index().values
4.165910928999438
SO_10982266_depth3
------------------
ERROR!
SO_18172851_depth1
------------------
arg0.loc[arg1]
1.152512497999851
SO_49987108_depth2
------------------
ERROR!
SO_49583055_depth1
------------------
arg0.sort_values(by=["ID"])
4.137355733999357
SO_49583055_depth1
------------------
arg0.sort_values(by=["ID"])
4.092509506999704
SO_49572546_depth1
------------------
arg1.combine_first(arg0)
2.0150949630005925
SO_39656670_depth3
------------------
ERROR!
SO_11881165_depth1
------------------
arg0.loc[[0, 2, 4]]
1.057938928999647
SO_11941492_depth1
------------------
ERROR!
SO_21982987_depth3
------------------
ERROR!
SO_53762029_depth3
------------------
arg0.pivot_table(index=["doc_created_month", "doc_created_year", "speciality"]).cumsum()
302.4615383049986
SO_13261691_depth2
------------------
arg0.stack().unstack()
50.20894425500046
SO_12065885_depth3
------------------
arg0.loc[[2, 4, 6]]
1.0424643039987131
SO_13261175_depth1
------------------
arg0.pivot_table(values="value", index="name", columns=["type", "date"])
276.5202395289998
SO_13659881_depth2
------------------
arg0.groupby(["ip", "useragent"]).size()
1.0170146630007366
SO_14085517_depth1
------------------
arg0.sort_values(by=["SEGM1"])
195.83883650999996
SO_34365578_depth2
------------------
ERROR!
SO_11418192_depth2
------------------
arg0.query(arg2)
0.9639060649988096
SO_49592930_depth1
------------------
arg0.combine_first(arg1)
0.9801300960007211
SO_13793321_depth1
------------------
arg0.merge(arg1, on=10)
5.554725879999751
SO_13647222_depth1
------------------
ERROR!
SO_12860421_depth1
------------------
arg0.pivot_table(values="Z", index="Y", columns="Z", aggfunc=pd.Series.nunique)
683.4204104500004This is implemented now!
The one in the code base is disabled right now. A proper pandas rows/columns domain needs to track the following information:
IndexKind: There are various kind of possible indexes in Pandas. Two domains are comparable if they have the same index kinds.IndexLabels: A set of index labels. Could be numeric, date time, strings or even tuples. This behaves like a mathematical set.Implementing this correctly should give us most constant strings automatically which are now supplied upfront.