Allow a Mixture of Degrees

[JoshuaC3]

I think it would be really helpful to allow a mixture of degrees for a model. This could be passed as a list or numpy array to the degree parameter. This has real-life applications: https://pdfs.semanticscholar.org/a6cf/634e8788586ba1d37c10c744db6a2f790109.pdf

In this scenario, a mixture of [0, 1] and [1, 0] degrees would be needed. However, it might be interesting to see if mixtures of higher degrees could be useful.

[cjekel]

I think this would be a nice addition! I won't be able to work on it any time soon though.

[JoshuaC3]

I'll take a look at it and do a PR at some point.

[cjekel]

Hey, I just got an email about this. Going to try an implementation today. It will work easiest with degrees 1 and 0 since they have the same number of parameters.

[cjekel]

Check out the mixed_degree branch.

Documentation on setting degree as list

"""
        degree : int, list, optional
            The degree of polynomial to use. The default is degree=1 for
            linear models. Use degree=0 for constant models. Use a list for
            mixed degrees (only supports degrees 1 or 0). List should be read
            from left to right, degree=[1, 0, 1] corresponds to a mixed degree
            model, where the left most segment has degree=1, the middle
            segment degree=0, and the right most segment degree=1.
"""

Mixed degree example:

import numpy as np
import matplotlib.pyplot as plt
import pwlf

x = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
y = [1, 2, 3, 4, 4.25, 3.75, 4, 5, 6, 7]

degree_list = [1, 0, 1]

my_pwlf = pwlf.PiecewiseLinFit(x, y, degree=degree_list)

breaks = my_pwlf.fit(3)

# generate predictions
x_hat = np.linspace(min(x), max(x), 1000)
y_hat = my_pwlf.predict(x_hat)

plt.figure()
plt.plot(x, y, 'o')
plt.plot(x_hat, y_hat)
plt.show()

Current issues:

• Does not check that the number of degrees matches the number of line segments
• Only supports degrees 1 and 0
• Slopes are not calculated correctly
• Needs unit tests

There are probably more issues with this...

The entire assemble_with_breaks should probably be rewritten to accommodate this feature in a cleaner manner.

[cjekel]

Here is the commit that added the changes: ef58aaa

[cyriltw]

I've been playing around with the latest commit and with the code segment, but I keep getting an error. Its on python 3.6 and I tried converting it to an np array as well.

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-6-42eb67ecfacf> in <module>()
     11 degree_list = [1, 0, 1]
     12 
---> 13 my_pwlf = pwlf.PiecewiseLinFit(x, y, degree=degree_list)
     14 
     15 breaks = my_pwlf.fit(3)

/usr/local/lib/python3.6/dist-packages/pwlf/pwlf.py in __init__(self, x, y, disp_res, lapack_driver, degree)
    164         self.break_n = np.max(self.x_data)
    165 
--> 166         if degree < 12 and degree >= 0:
    167             # I actually don't know what the upper degree limit should
    168             self.degree = int(degree)

TypeError: '<' not supported between instances of 'list' and 'int'

[cjekel]

@cyriltw You'll need to install that experimental branch first.

git clone git@github.com:cjekel/piecewise_linear_fit_py.git
cd piecewise_linear_fit_py
git checkout mixed_degree
pip install . --force-reinstall --no-deps

[cyriltw]

@cjekel you beat me to it, for some reason the branch wasn't checked out so causing the trouble. Now it works. Doing more testing. Thanks for including this