[BUG] Force softmax returning Series

[Zeroto521]

PR Description

Please describe the changes proposed in the pull request:

115         >>> import pandas as pd
116         >>> import janitor
117         >>> s = pd.Series([0, 1, 3], name="numbers")
118         >>> s.softmax()
Expected:
    0    0.042010
    1    0.114195
    2    0.843795
    Name: numbers, dtype: float64
Got:
    array([0.04201007, 0.1141952 , 0.84379473])

This PR resolves https://github.com/pyjanitor-devs/pyjanitor/runs/7703616556?check_suite_focus=true .

PR Checklist

Please ensure that you have done the following:

1. PR in from a fork off your branch. Do not PR from <your_username>:dev, but rather from <your_username>:<feature-branch_name>.

2. If you're not on the contributors list, add yourself to AUTHORS.md.

3. Add a line to CHANGELOG.md under the latest version header (i.e. the one that is "on deck") describing the contribution.
   • Do use some discretion here; if there are multiple PRs that are related, keep them in a single line.

Automatic checks

There will be automatic checks run on the PR. These include:

• Building a preview of the docs on Netlify
• Automatically linting the code
• Making sure the code is documented
• Making sure that all tests are passed
• Making sure that code coverage doesn't go down.

Relevant Reviewers

Please tag maintainers to review.

• @ericmjl

[Zeroto521]

Actually I'd like use np.exp(s) / sum(np.exp(s)) instead of using scipy

[ericmjl]

I think the SciPy implementation comes with numerical stability checks for free. Is there a reason why you would prefer to use a NumPy implementation?

[Zeroto521]

To totally replace scipy. Since we only use it in math.py and functions/impute.py

• mode function in functions/impute.py could replace by Series.value_counts
• math.py
  • sigmoid (expit) could use np.exp
  • softmax (scipy.special.softmax) could use np.exp
  • logit (scipy.special.softmax) could use np.log
  • normal_cdf (scipy.stats.norm.cdf) and probit (scipy.stats.norm.ppf) can't replace easily.

[codecov]

Codecov Report

Merging #1139 (9fcbc4e) into dev (f748406) will decrease coverage by 0.01%.
The diff coverage is 97.14%.

@@            Coverage Diff             @@
##              dev    #1139      +/-   ##
==========================================
- Coverage   97.36%   97.34%   -0.02%     
==========================================
  Files          77       77              
  Lines        3186     3240      +54     
==========================================
+ Hits         3102     3154      +52     
- Misses         84       86       +2     

[ericmjl]

I am good to go. I think we should keep the SciPy soft Mac implementation rather than switch to numpy, for numerical stability reasons (IIRC).