[BUG] model.score does not return correct values when using default r2_score

[ruslanmustafin]

As per scikit-learn r2_score documentation, the function expects the arguments to be y_true followed by y_pred, however, in your implementation the order is reversed. Since R^2 is non-symmetric, it will produce a different score compared to correct order of arguments. This is hardly noticeable when predictions closely match the gt (score is close to 1.0), but will produce incorrect results otherwise.

Reproducing code example:

import numpy as np

from scipy.integrate import odeint
from sklearn.metrics import r2_score

import pysindy as ps

# Control input
def u_fun(t):
    return np.column_stack([np.sin(2 * t), t ** 2])

# Lorenz equations with control input
def lorenz_control(z, t):
    u = u_fun(t)
    return [
        10 * (z[1] - z[0]) + u[0, 0] ** 2,
        z[0] * (28 - z[2]) - z[1],
        z[0] * z[1] - 8 / 3 * z[2] - u[0, 1],
    ]

# Generate measurement data
dt = .002

t_train = np.arange(0, 10, dt)
x0_train = [-8, 8, 27]
x_train = odeint(lorenz_control, x0_train, t_train)
u_train = u_fun(t_train)

# Instantiate and fit the SINDYc model
model = ps.SINDy()
model.fit(x_train, u=u_train, t=dt)

print(model.score(x_train, u=u_train, t=dt))

x_dot_test_predicted = model.predict(x_train, u=u_train)
x_dot_test_computed = model.differentiate(x_train, t=dt)

print(r2_score(x_dot_test_computed, x_dot_test_predicted))

This should print 0.999999960658854 and 0.999999960658851 (sorry for not providing a better example, I would have to give away our internal dataset to reproduce the issue, but in our case the score changes from 0.41 to 0.05)

PySINDy/Python version information:

1.0.1 3.6.10 |Anaconda, Inc.| (default, Mar 25 2020, 23:51:54)
[GCC 7.3.0]

[briandesilva]

Thanks for pointing this out. The latest release, 1.0.2, fixes this issue.