def exp_space_leakage(data):
""" Evaluates whether a set of features (default PROXIMITY_FEATURES) "leak"
spatial information to GBRT by producing a scatter plot of spatial
distance vs Euclidean distance in the subspace of specified features
for a collection of random pairs of data points (default n = 20k).
Plot is saved to <OUT_DIR>/space-leakage.png.
"""
dumpfile = 'space_leakage.txt'
num_samples = 20000
plot_space_leakage(data, num_samples, features=PROXIMITY_FEATURES, dumpfile=dumpfile, replot=False)
save_cur_fig('space-leakage.png', title='Spatial information leakage through proximity features', set_title_for=None)
def exp_sensitivity(data):
""" Evaluates sensitivity of GBRT and linear regression to perturbations in
training GHF. Plot is saved to <OUT_DIR>/sensitivity.png.
"""
radius = GREENLAND_RADIUS
roi_density = 11.3 # Greenland
noise_amps = np.arange(0.025, .31, .025)
ncenters = 50
dumpfile = 'sensitivity.txt'
plotfile = 'sensitivity.png'
plot_sensitivity_analysis(data, roi_density, radius, noise_amps, ncenters, dumpfile=dumpfile, replot=False)
save_cur_fig(plotfile, title='GBRT prediction sensitivity to noise in training GHF', set_title_for=None)
def exp_feature_importance(data):
""" Plots feature importances for averaged over 50 ROIs with the same
radius and sample density as the GrIS train/validation split. The max
depth of trees is increased to 8 to avoid a few influential features
taking over all trees."""
radius = GREENLAND_RADIUS
ncenters = 50
roi_density = 11.3 # Greenland
max_depth = 8
dumpfile = 'feature_importances.txt'
plot_feature_importance_analysis(data, roi_density, radius, ncenters, dumpfile=dumpfile, max_depth=max_depth, replot=False)
save_cur_fig('feature-importance.png', title='Relative feature importances in GBRT', set_title_for=None)
def exp_partial_dependence():
""" Produces partial dependence plots for GBRT. The one-way PPD is produced
for all non-categorical features. The two-way PPD is produced for all
combinations of a fixed set of top 6 features."""
X_train, y_train, _ = greenland_train_test_sets()
X_train = X_train.drop(['lat', 'lon'], axis=1)
plot_partial_dependence(X_train, y_train, n_ways=1, include_features=None)
save_cur_fig('partial-dependence-one-way.png', title='One way partial dependences', set_title_for=None)
top_features = ['age', 'G_d_2yng_r', 'd_2trench', 'litho_asth', 'ETOPO_1deg', 'moho_GReD']
plot_partial_dependence(X_train, y_train, n_ways=2, include_features=top_features)
save_cur_fig('partial-dependence-two-way.png', title='Two way partial dependences', set_title_for=None)
def exp_generalization(data):
""" Evaluates the generalization power of GBRT with increasing complexity
(number of regression tress). This is used to verify that GBRT is
robust against overfitting and to pick an appropriate number of trees
for reported results and used in all other experiments (cf. `util.GBRT_params`).
"""
radius = GREENLAND_RADIUS
ncenters = 50
roi_density = 11.3 # Greenland
ns_estimators = range(50, 750, 100) + range(750, 3001, 750)
dumpfile = 'generalization.txt'
plotfile = 'generalization.png'
plot_generalization_analysis(data, roi_density, radius, ncenters, ns_estimators, dumpfile=dumpfile, replot=False)
save_cur_fig(plotfile, title='GBRT generalization power', set_title_for=None)
def exp_error_by_density(data):
""" Evaluates prediction error (normalized rmse and r2) for GBRT, linear
regression and constant predictor by using increasingly large sample
densities in ROIs, constrained to the specified region, with radius
equal to that of Greenland. Plot is saved to <OUT_DIR>/error_by_density[<region>].png.
"""
densities = np.append(np.array([1]), np.arange(5, 51, 5))
radius = GREENLAND_RADIUS
ncenters = 50
# region constraints: 'NA-WE', 'NA', 'WE', or None (i.e all)
region = 'NA-WE'
dumpfile = 'error_by_density[%s].txt' % region
plotfile = 'error_by_density[%s].png' % region
plot_error_by_density(data, densities, radius, ncenters, region=region, dumpfile=dumpfile, replot=False)
save_cur_fig(plotfile)
def exp_error_by_radius(data):
""" Evaluates prediction error (normalized rmse and r2) for GBRT, linear
regression and constant predictor by using increasingly large radii for
ROIs, constrained to the specified region, with sample density equal to
that of Greenland. Plot is saved to <OUT_DIR>/error_by_radius[<region>].png.
"""
radius = GREENLAND_RADIUS
roi_density = 11.3 # Greenland
ncenters = 50
radii = np.arange(500, 4001, 500)
region = 'NA-WE'
dumpfile = 'error_by_radius[%s].txt' % region
plotfile = 'error_by_radius[%s].png' % region
sys.stderr.write('=> Experiment: Error by Radius (region: %s, no. centers: %d, no. radii: %d)\n' % (region, ncenters, len(radii)))
plot_error_by_radius(data, roi_density, radii, ncenters, region=region, dumpfile=dumpfile, replot=False)
save_cur_fig(plotfile)
if __name__ == '__main__':
X_train, y_train, X_test = greenland_train_test_sets()
train_lons = X_train.lon.as_matrix()
train_lats = X_train.lat.as_matrix()
X_train = X_train.drop(['lat', 'lon'], axis=1)
test_lons = X_test.lon.as_matrix()
test_lats = X_test.lat.as_matrix()
X_test = X_test.drop(['lat', 'lon'], axis=1)
# -------------------- Plot training data -------------------------
plot_training_GHF(train_lons, train_lats, y_train)
save_cur_fig('greenland_training_GHF.png', title='GHF at training set')
plot_gaussian_prescribed_GHF(train_lons, train_lats, y_train)
save_cur_fig(
'greenland_prescribed_GHF.png',
title=
'Points with prescribed GHF \n around GHF measurements (mW m$^{-2}$)')
# -------------------- Plot predicted results ----------------------
reg = train_gbrt(X_train, y_train)
y_pred = reg.predict(X_test)
plot_prediction_points(test_lons, test_lats, y_pred)
save_cur_fig('greenland_prediction_points.png',
title='GHF predicted for Greenland (mW m$^{-2}$)')
parallel_step=10.,
meridian_step=10.,
colorbar_args=COLORBAR_ARGS,
scatter_args=scatter_args)
equi(m,
center[0],
center[1],
GREENLAND_RADIUS,
lw=2,
linestyle='-',
color='black',
alpha=.5)
title = r'$GHF - \widehat{GHF}$ on validation set with ' + \
r'$\rho_{ROI}$ = %d'%roi_density
save_cur_fig('%d-diff-map.png' % roi_density, title=title)
plot_test_pred_linregress(y_test, y_pred, label='GBRT', color='b')
save_cur_fig('%d_linear_correlation.png' % roi_density,
title=r'$\rho_{ROI}$ = %i, $r^2=%.2f, RMSE=%.2f$' %
(roi_density, r2, rmse))
### Main Text Figure 1
## gbrt regression
X = data.drop(['GHF'], axis=1)
y = data.GHF
X_train, X_test, y_train, y_test = train_test_split(X,
y,
random_state=11,
test_size=0.10)