def run_LinearRegression_fit():
"""
Run the fit but now using astroML's LinearRegression
"""
x, y, y_obs, sigma = generate_points()
m = LinearRegression()
m.fit(x[:, None], y_obs, sigma)
print("Intercept={0} Slope={1}".format(m.coef_[0], m.coef_[1]))
def test_LinearRegression_simple():
"""
Test a simple linear regression
"""
x = np.arange(10.).reshape((10, 1))
y = np.arange(10.) + 1
dy = 1
clf = LinearRegression().fit(x, y, dy)
y_true = clf.predict(x)
assert_allclose(y, y_true, atol=1E-10)
def test_error_transform_diag(N=20, rseed=0):
rng = np.random.RandomState(rseed)
X = rng.rand(N, 2)
yerr = 0.05 * (1 + rng.rand(N))
y = (X[:, 0]**2 + X[:, 1]) + yerr * rng.randn(N)
Sigma = np.eye(N) * yerr**2
X1, y1 = LinearRegression._scale_by_error(X, y, yerr)
X2, y2 = LinearRegression._scale_by_error(X, y, Sigma)
assert_allclose(X1, X2)
assert_allclose(y1, y2)
def test_error_transform_diag(N=20, rseed=0):
rng = np.random.RandomState(rseed)
X = rng.rand(N, 2)
yerr = 0.05 * (1 + rng.rand(N))
y = (X[:, 0] ** 2 + X[:, 1]) + yerr * rng.randn(N)
Sigma = np.eye(N) * yerr ** 2
X1, y1 = LinearRegression._scale_by_error(X, y, yerr)
X2, y2 = LinearRegression._scale_by_error(X, y, Sigma)
assert_allclose(X1, X2)
assert_allclose(y1, y2)
def test_LinearRegression_simple():
"""
Test a simple linear regression
"""
x = np.arange(10.).reshape((10, 1))
y = np.arange(10.) + 1
dy = 1
clf = LinearRegression().fit(x, y, dy)
y_true = clf.predict(x)
assert_allclose(y, y_true, atol=1E-10)
def test_LinearRegression_fit_intercept():
np.random.seed(0)
X = np.random.random((10, 1))
y = np.random.random(10)
clf1 = LinearRegression(fit_intercept=False).fit(X, y)
clf2 = skLinearRegression(fit_intercept=False).fit(X, y)
assert_allclose(clf1.coef_, clf2.coef_)
def linearregression(M, z_spec, M_B, z_spec_B):
#apply linear regression techniques on M to find a function for z_phot
model = LinearRegression(fit_intercept=True)
res = model.fit(M, z_spec)
coeff = list(res.coef_)
print "The best fit model is:"
print "z_phot = {0:.3f} + {1:.3f} (u-g) + {2:.3f} (g-r) + {3:.3f} (r-i) + {4:.3f} (i-z)".format(
coeff[0], coeff[1], coeff[2], coeff[3], coeff[4])
#test if z_phot is predicted to have a value close to z_spec
z_phot = model.predict(M)
print "The training error on the fit is:", Training_Error(z_phot, z_spec)
#calculate the training error using file B
z_phot_B = coeff[0] + coeff[1] * M_B[:, 0] + coeff[2] * M_B[:, 1] + coeff[
3] * M_B[:, 2] + coeff[4] * M_B[:, 3]
print "The estimated error for the test file B is:", Training_Error(
z_phot_B, z_spec_B)
def test_LinearRegression_err():
"""
Test that errors are correctly accounted for
By comparing to scikit-learn LinearRegression
"""
np.random.seed(0)
X = np.random.random((10, 1))
y = np.random.random(10) + 1
dy = 0.1
y = np.random.normal(y, dy)
X_fit = np.linspace(0, 1, 10)[:, None]
clf1 = LinearRegression().fit(X, y, dy)
clf2 = skLinearRegression().fit(X / dy, y / dy)
assert_allclose(clf1.coef_[1:], clf2.coef_)
assert_allclose(clf1.coef_[0], clf2.intercept_ * dy)
def test_error_transform_full(N=20, rseed=0):
rng = np.random.RandomState(rseed)
X = rng.rand(N, 2)
# generate a pos-definite error matrix
Sigma = 0.05 * rng.randn(N, N)
u, s, v = np.linalg.svd(Sigma)
Sigma = np.dot(u * s, u.T)
# draw y from this error distribution
y = (X[:, 0]**2 + X[:, 1])
y = rng.multivariate_normal(y, Sigma)
X2, y2 = LinearRegression._scale_by_error(X, y, Sigma)
# check that the form entering the chi^2 is correct
assert_allclose(np.dot(X2.T, X2), np.dot(X.T, np.linalg.solve(Sigma, X)))
assert_allclose(np.dot(y2, y2), np.dot(y, np.linalg.solve(Sigma, y)))
def test_LinearRegressionwithErrors():
"""
Test for small errors agrees with fit with y errors only
"""
from astroML.linear_model import LinearRegressionwithErrors
np.random.seed(0)
X = np.random.random(10) + 1
dy = np.random.random(10) * 0.1
y = X * 2 + 1 + (dy - 0.05)
dx = np.random.random(10) * 0.01
X = X + (dx - 0.005)
clf1 = LinearRegression().fit(X[:, None], y, dy)
clf2 = LinearRegressionwithErrors().fit(np.atleast_2d(X), y, dy, dx)
assert_allclose(clf1.coef_, clf2.coef_, 0.2)
def test_error_transform_full(N=20, rseed=0):
rng = np.random.RandomState(rseed)
X = rng.rand(N, 2)
# generate a pos-definite error matrix
Sigma = 0.05 * rng.randn(N, N)
u, s, v = np.linalg.svd(Sigma)
Sigma = np.dot(u * s, u.T)
# draw y from this error distribution
y = (X[:, 0] ** 2 + X[:, 1])
y = rng.multivariate_normal(y, Sigma)
X2, y2 = LinearRegression._scale_by_error(X, y, Sigma)
# check that the form entering the chi^2 is correct
assert_allclose(np.dot(X2.T, X2),
np.dot(X.T, np.linalg.solve(Sigma, X)))
assert_allclose(np.dot(y2, y2),
np.dot(y, np.linalg.solve(Sigma, y)))
#------------------------------------------------------------
# Generate data
z_sample, mu_sample, dmu = generate_mu_z(100, random_state=0)
cosmo = Cosmology()
z = np.linspace(0.01, 2, 1000)
mu_true = np.asarray(map(cosmo.mu, z))
#------------------------------------------------------------
# Define our classifiers
basis_mu = np.linspace(0, 2, 15)[:, None]
basis_sigma = 3 * (basis_mu[1] - basis_mu[0])
subplots = [221, 222, 223, 224]
classifiers = [
LinearRegression(),
PolynomialRegression(4),
BasisFunctionRegression('gaussian', mu=basis_mu, sigma=basis_sigma),
NadarayaWatson('gaussian', h=0.1)
]
text = [
'Straight-line Regression', '4th degree Polynomial\n Regression',
'Gaussian Basis Function\n Regression', 'Gaussian Kernel\n Regression'
]
# number of constraints of the model. Because
# Nadaraya-watson is just a weighted mean, it has only one constraint
n_constraints = [2, 5, len(basis_mu) + 1, 1]
#------------------------------------------------------------
# Plot the results
def plot_regressions(ksi,
eta,
x,
y,
sigma_x,
sigma_y,
add_regression_lines=False,
alpha_in=1,
beta_in=0.5,
basis='linear'):
figure = plt.figure(figsize=(8, 6))
ax = figure.add_subplot(111)
ax.scatter(x, y, alpha=0.5)
ax.errorbar(x, y, xerr=sigma_x, yerr=sigma_y, alpha=0.3, ls='')
ax.set_xlabel('x')
ax.set_ylabel('y')
x0 = np.linspace(np.min(x) - 0.5, np.max(x) + 0.5, 20)
# True regression line
if alpha_in is not None and beta_in is not None:
if basis == 'linear':
y0 = alpha_in + x0 * beta_in
elif basis == 'poly':
y0 = alpha_in + beta_in[0] * x0 + beta_in[1] * x0 * x0 + beta_in[
2] * x0 * x0 * x0
ax.plot(x0, y0, color='black', label='True regression')
else:
y0 = None
if add_regression_lines:
for label, data, *target in [['fit no errors', x, y, 1],
['fit y errors only', x, y, sigma_y],
['fit x errors only', y, x, sigma_x]]:
linreg = LinearRegression()
linreg.fit(data[:, None], *target)
if label == 'fit x errors only' and y0 is not None:
x_fit = linreg.predict(y0[:, None])
ax.plot(x_fit, y0, label=label)
else:
y_fit = linreg.predict(x0[:, None])
ax.plot(x0, y_fit, label=label)
# TLS
X = np.vstack((x, y)).T
dX = np.zeros((len(x), 2, 2))
dX[:, 0, 0] = sigma_x
dX[:, 1, 1] = sigma_y
def min_func(beta):
return -TLS_logL(beta, X, dX)
beta_fit = optimize.fmin(min_func, x0=[-1, 1])
m_fit, b_fit = get_m_b(beta_fit)
x_fit = np.linspace(-10, 10, 20)
ax.plot(x_fit, m_fit * x_fit + b_fit, label='TLS')
ax.set_xlim(np.min(x) - 0.5, np.max(x) + 0.5)
ax.set_ylim(np.min(y) - 0.5, np.max(y) + 0.5)
ax.legend()
fh.close()
except:
print("Loading pickled data failed!", sys.exc_info()[0])
data = None
return data
d = pickle_from_file('points_example1.pkl')
x = d['x']
yobs = d['y']
sigma = d['sigma']
M = x[:, None]
model = LinearRegression(fit_intercept=True)
res = model.fit(M, yobs, sigma)
model.predict(M)
print(res.coef_)
def lnprob(theta, x, yobs, sigma):
a, b = theta
model = b * x + a
inv_sigma2 = 1.0 / (sigma**2)
return -0.5 * (np.sum((yobs - model)**2 * inv_sigma2))
p_init = res.coef_
ndim, nwalkers = 2, 100
pos = [p_init + 1e-4 * np.random.randn(ndim) for i in range(nwalkers)]
widths = 0.2
X = gaussian_basis(z_sample[:, np.newaxis], centers, widths)
#------------------------------------------------------------
# Set up the figure to plot the results
fig = plt.figure(figsize=(5, 2.7))
fig.subplots_adjust(left=0.1, right=0.95,
bottom=0.1, top=0.95,
hspace=0.15, wspace=0.2)
regularization = ['none', 'l2', 'l1']
kwargs = [dict(), dict(alpha=0.005), dict(alpha=0.001)]
labels = ['Linear Regression', 'Ridge Regression', 'Lasso Regression']
for i in range(3):
clf = LinearRegression(regularization=regularization[i],
fit_intercept=True, kwds=kwargs[i])
clf.fit(X, mu_sample, dmu)
w = clf.coef_[1:]
fit = clf.predict(gaussian_basis(z[:, None], centers, widths))
# plot fit
ax = fig.add_subplot(231 + i)
ax.xaxis.set_major_formatter(plt.NullFormatter())
# plot curves for regularized fits
if i == 0:
ax.set_ylabel('$\mu$')
else:
ax.yaxis.set_major_formatter(plt.NullFormatter())
curves = 37 + w * gaussian_basis(z[:, np.newaxis], centers, widths)
curves = curves[:, abs(w) > 0.01]
widths = 0.2
X = gaussian_basis(z_sample[:, np.newaxis], centers, widths)
#------------------------------------------------------------
# Set up the figure to plot the results
fig = plt.figure(figsize=(5, 2.7))
fig.subplots_adjust(left=0.1, right=0.95,
bottom=0.1, top=0.95,
hspace=0.15, wspace=0.2)
regularization = ['none', 'l2', 'l1']
kwargs = [dict(), dict(alpha=0.005), dict(alpha=0.001)]
labels = ['Linear Regression', 'Ridge Regression', 'Lasso Regression']
for i in range(3):
clf = LinearRegression(regularization=regularization[i],
fit_intercept=True, kwds=kwargs[i])
clf.fit(X, mu_sample, dmu)
w = clf.coef_[1:]
fit = clf.predict(gaussian_basis(z[:, None], centers, widths))
# plot fit
ax = fig.add_subplot(231 + i)
ax.xaxis.set_major_formatter(plt.NullFormatter())
# plot curves for regularized fits
if i == 0:
ax.set_ylabel('$\mu$')
else:
ax.yaxis.set_major_formatter(plt.NullFormatter())
curves = 37 + w * gaussian_basis(z[:, np.newaxis], centers, widths)
curves = curves[:, abs(w) > 0.01]
