[FEA] cudf.Series equivalent of pd.Series.to_dict or pd.Series.iteritems or pd.Series.items

[paul-tqh-nguyen]

Is your feature request related to a problem? Please describe.

It's useful to convert a series into a dictionary when interoperating with other Python libraries.

We can do this with a pandas series in a few ways.

pnguyen@machine:/tmp$ python3 -c "

import pandas as pd

df = pd.DataFrame()
df['keys'] = [111, 222, 333]
df['vals'] = ['a', 'b', 'c']
df = df.set_index('keys')

dict1 = df.vals.to_dict()
dict2 = {k:v for k,v in df.vals.iteritems()}
dict3 = {k:v for k,v in df.vals.items()}

print()
print(f'dict1 {repr(dict1)}')
print(f'dict2 {repr(dict2)}')
print(f'dict3 {repr(dict3)}')
print()

"

dict1 {111: 'a', 222: 'b', 333: 'c'}
dict2 {111: 'a', 222: 'b', 333: 'c'}
dict3 {111: 'a', 222: 'b', 333: 'c'}

pnguyen@machine:/tmp$ 

The above approaches don't seem to be supported in cudf.

pnguyen@machine:/tmp$ python3 -c "

import cudf

df = cudf.DataFrame()
df['keys'] = [111, 222, 333]
df['vals'] = ['a', 'b', 'c']
df = df.set_index('keys')

dict = df.vals.to_dict()

"
Traceback (most recent call last):
  File "<string>", line 10, in <module>
AttributeError: 'Series' object has no attribute 'to_dict'
pnguyen@machine:/tmp$ python3 -c "

import cudf

df = cudf.DataFrame()
df['keys'] = [111, 222, 333]
df['vals'] = ['a', 'b', 'c']
df = df.set_index('keys')

dict = {k:v for k,v in df.vals.iteritems()}

"
Traceback (most recent call last):
  File "<string>", line 10, in <module>
AttributeError: 'Series' object has no attribute 'iteritems'
pnguyen@machine:/tmp$ python3 -c "

import cudf

df = cudf.DataFrame()
df['keys'] = [111, 222, 333]
df['vals'] = ['a', 'b', 'c']
df = df.set_index('keys')

dict = {k:v for k,v in df.vals.items()}

"
Traceback (most recent call last):
  File "<string>", line 10, in <module>
AttributeError: 'Series' object has no attribute 'items'
pnguyen@machine:/tmp$ 

This is my current workaround:

pnguyen@machine:/tmp$ python3 -c "

import cudf

df = cudf.DataFrame()
df['keys'] = [111, 222, 333]
df['vals'] = ['a', 'b', 'c']
df = df.set_index('keys')

dict = {i.item():df.vals.loc[i.item()] for i in df.vals.index.values}

print()
print(f'dict {repr(dict)}')
print()

"

dict {111: 'a', 222: 'b', 333: 'c'}

pnguyen@machine:/tmp$ 

Describe the solution you'd like

Ideally, there would be support for cudf.Series.to_dict or cudf.Series.iteritems or cudf.Series.items so that a simple pd -> cudf replacement would work.

Describe alternatives you've considered

I'm currently using the workaround shown above.

Additional context

[kkraus14]

@paul-tqh-nguyen we intentional haven't implemented these methods because there's a HUGE performance penalty to leaving the GPU that we don't want users to fall into implicitly. That being said we should throw more explicit exceptions here saying that they're not implemented for performance reasons, but if you need them to go via PyArrow.

[paul-tqh-nguyen]

Yes, that makes complete sense. Thanks for the insight! It'll help me figure out a smarter way to deal with my problem.