ceholden/pysmoothspl

Python wrapper around R's lovely `smooth.spline`

pysmoothspl

Python wrapper around R's lovely smooth.spline

Example

import matplotlib.pyplot as plt
import numpy as np
from pysmoothspl import SmoothSpline

n = 1000
x = np.arange(n).astype(np.float)
y = 100 + 50 * np.sin(2 * np.pi * x / 365.0) + np.random.normal(0, 25, n)

spl = SmoothSpline(spar=0.2).fit(x, y)
yhat = spl.predict(x)
spl_smoother = SmoothSpline(spar=0.5).fit(x, y)
yhat_smoother = spl_smoother.predict(x)

fig, ax = plt.subplots(1, figsize=(16, 9))
ax.plot(x, y, 'ro', label='Points')
ax.plot(x, yhat, 'g-', label='Less smooth')
ax.plot(x, yhat_smoother, 'b-', label='Smoother')
ax.legend()
fig.show()

Install

You will need Cython (unless/until this repo includes the *.c files) and numpy.

Anaconda

There are two Conda environment.yaml files to help guide the installation:

1. environment.yaml contains just the needed packages for installation
2. tests/environment.yaml contains the packages needed to install and test this package

Tests

This package uses py.test to run tests, and you will also need pandas:

py.test tests/

If rpy2 is installed, the tests will actually calculate the expected values of calculations using the copy of R it is installed against. For users without rpy2, the test data also contains "cached" answers calculated using R code.

TODO

• Object oriented sklearn-esque estimator
• For function inputs: typed memory views > NumPy buffers
• Kill the GIL
• CI service tests
  • Code coverage checks with CI
• Check output against literal R output via rpy2
• Implement more features
  • Cross validation
  • Derivatives (hint: check bvalue code for commented out)
• Work back from fresh copy of R code...

pysmoothspl's Issues

Guard against 0 length inputs

Commenting for later -- code segfaults Python if given x, y = np.array([]), np.array([]). Guard against this in Python part of wrapper and raise an exception