def gp_adjust(lon, lat, gp_x, gp_y):
x,y,z = sphere.spherical_to_cartesian((lon, lat))
az_x, az_y = sphere.polar_to_az(lon, lat)
xc = gp_x.predict([x,y,z])
yc = gp_y.predict([x,y,z])
corr_touch_lon, corr_touch_lat = sphere.az_to_polar(az_x+xc, az_y+yc)
return corr_touch_lon, corr_touch_lat
def train_gp_sp(calibration, n=None):
"""Train a pair of squared-exponential Gaussian processes to predict offsets
in azimuthal equidistant space. The GPs are trained on x, y inputs in the
azimuthal space, and there is one GP for the x' and y' outputs.
Returns a pair of GP objects (gp_x, gp_y) that perform the prediction."""
unique_targets = calibration.groupby(["target_x", "target_y"]).mean()
if n is not None:
unique_targets = unique_targets.loc[random.sample(unique_targets.index, n)]
target_x = unique_targets["target_az_x"]
target_y = unique_targets["target_az_y"]
corr_x = unique_targets["touch_az_x"]
corr_y = unique_targets["touch_az_y"]
resid_x = np.array(target_x - corr_x)
resid_y = np.array(target_y - corr_y)
x,y,z = sphere.spherical_to_cartesian((unique_targets["touch_lon"], unique_targets["touch_lat"]))
target = np.vstack((x,y,z)).transpose()
gp_x = gaussian_process.GaussianProcess(theta0=1e-2, thetaL=1e-7, thetaU=3e1, nugget=1e-3, random_start=10)
gp_y = gaussian_process.GaussianProcess(theta0=1e-2, thetaL=1e-7, thetaU=3e1, nugget=1e-3, random_start=10)
gp_x.fit(target, resid_x)
gp_y.fit(target, resid_y)
return gp_x, gp_y
