def residuals(samples, i, j, cond_set):
cond_set = list(cond_set)
g = pygam.GAM()
g.fit(samples[:, cond_set], samples[:, i])
residuals_i = g.deviance_residuals(samples[:, cond_set], samples[:, i])
g.fit(samples[:, cond_set], samples[:, j])
residuals_j = g.deviance_residuals(samples[:, cond_set], samples[:, j])
return residuals_i, residuals_j
def gammer(self, n_splines=21):
self.model = pygam.GAM(n_splines=n_splines).fit(
self.x_train, self.y_train)
self.train_preds = self.model.predict(self.x_train)
self.test_preds = self.model.predict(self.x_test)
train_mse = mse(self.y_train, self.train_preds)
test_mse = mse(self.y_test, self.test_preds)
print(train_mse, test_mse)
print("Train R2", r2_score(self.y_train, self.train_preds))
print("Test R2", r2_score(self.y_test, self.test_preds))
def gamSplineSens(preinterpsurfaces):
sumerrors=[]
for splines in range(4,16):
error=[]
for k in preinterpsurfaces.keys():
if int(k)>2000:
surface = preinterpsurfaces[k]
X = np.zeros((len(surface["lons"]),2))
X[:,0]=surface["lons"]
X[:,1]=surface["lats"]
#for d in Bar("Interpolating: ").iter(surface["data"].keys()):
d = "pres"
notnan = ~np.isnan(surface["data"][d])
if np.count_nonzero(notnan)>10:
gam = pygam.GAM(pygam.te(0,1,n_splines=[splines,splines])).fit(X[notnan],np.asarray(surface["data"][d])[notnan])
#random_gam = pygam.LinearGAM(pygam.s(0) + pygam.s(1) ).gridsearch(X, surface["data"][d])
error += list(np.log10(np.abs(surface["data"][d]-gam.predict(X))))
sns.distplot(error,kde_kws={"fill":False,"label": str(splines)})
#plt.plot(range(4,16),sumerrors)
plt.legend()
plt.show()
def gammerMulti(self):
train_preds = []
test_preds = []
for i in range(5, 50):
print("Iter ", i)
model = pygam.GAM(n_splines=i).fit(self.x_train, self.y_train)
train_preds.append(mse(self.y_train, model.predict(self.x_train)))
test_preds.append(mse(self.y_test, model.predict(self.x_test)))
print(train_preds)
print(test_preds)
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(12, 4))
fig.suptitle("Train Error vs. Test Error", fontsize=14, y=1)
plt.subplots_adjust(top=0.93, wspace=0)
ax1.plot(range(5, 50), test_preds, linestyle='-', marker='o')
ax1.set_title("Test Set")
ax1.set_xlabel("N_Splines")
ax1.set_ylabel("MSE")
ax2.plot(range(5, 50), train_preds, linestyle='-', marker='o')
ax2.set_title("Train Set")
ax2.set_xlabel("N_Splines")
ax2.set_ylabel("")
plt.show()
mse_train = mean_squared_error(y[train], lm.predict(X[train]))
mse_test = mean_squared_error(y[test], lm.predict(X[test]))
print('MSE train: %.4f, MSE test: %.4f' % (mse_train, mse_test))
mse_train_lm.append(mse_train)
mse_test_lm.append(mse_test)
print('Mean MSE train: %.4f (%.4f), Mean MSE test: %.4f (%.4f)' %
(np.mean(mse_train_lm), np.std(mse_train_lm), np.mean(mse_test_lm),
np.std(mse_test_lm)))
#%% GAM
# Modules
import pygam
pygam.GAM(pygam.s(0), distribution='normal', link='identity')
#%% Ridge
mse_train_ridge = list()
mse_test_ridge = list()
for train, test in kf_cv.split(X, y):
lasso = Ridge()
lasso.fit(X[train], y[train])
mse_train = mean_squared_error(y[train], ridge.predict(X[train]))
mse_test = mean_squared_error(y[test], ridge.predict(X[test]))
print('MSE train: %.4f, MSE test: %.4f' % (mse_train, mse_test))
mse_train_ridge.append(mse_train)
mse_test_ridge.append(mse_test)
#%% Network
def gam_reg(target_samples, cond_samples):
g = pygam.GAM()
g.fit(cond_samples, target_samples)
print(g.coef_.shape)
residuals = g.deviance_residuals(cond_samples, target_samples)
return residuals
@author: Christian Winkler
"""
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
from pygam import LinearGAM
import pygam
example_data = pd.read_csv("example_data.csv")
y = example_data['head'].values
X = example_data['age'].values
#gam = LinearGAM()
#gam = LinearGAM(n_splines=10).gridsearch(X, y)
gam = pygam.GAM(n_splines=10).gridsearch(X, y)
gam.fit(X, y) # your fitted model
samples = gam.sample(X, y, quantity='y', n_draws=500, sample_at_X=X)
# sampels is shape (500, len(y))
percentiles = np.percentile(samples, q=[2.5, 97.5], axis=0)
#percentiles is now shape (2, len(y))
plt.figure(figsize=(10, 8))
plt.scatter(X, y)
plt.plot(X, percentiles[0])
plt.plot(X, percentiles[1])
# plotting
plt.savefig("pygam_example.png")
def interpolateSurface(surface,region,coord="xy",debug=True,interpmethod="gam",smart=False,splines=10):
#print("######")
interpsurf={}
X = np.zeros((len(surface["lons"]),2))
X[:,0]=surface["lons"]
X[:,1]=surface["lats"]
if smart:
xi,yi,neighbors,finalids = smartMesh(surface["x"],surface["y"],region,coord)
else:
xi,yi,neighbors,finalids = generateMaskedMesh(surface["lons"],surface["lats"],region,coord)
interpdata={}
interpsurf["x"] =xi
interpsurf["y"] =yi
interpsurf["ids"] =finalids
if len(xi) != len(finalids):
print("OH NOOOOOO")
if interpmethod=="gam":
for d in surface["data"].keys():
notnan = ~np.isnan(surface["data"][d])
if np.count_nonzero(notnan)>10:
gam = pygam.GAM()
lams = np.logspace(-3, 3, 11)
# lams = [lams,lams]
# sclfactor = np.nanmin(np.nanmin(np.asarray(surface["data"][d])[notnan]))
# scores = gam.gridsearch(X[notnan],np.asarray(surface["data"][d])[notnan]/sclfactor,return_scores=True,objective="GCV")
df = pd.DataFrame({'lon': X[:,0][notnan], 'lat': X[:,1][notnan], 'd':(np.asarray(surface["data"][d])[notnan])})
r_dataframe = pandas2ri.py2rpy(df)
tps=mgcv.gamm(ro.Formula('d~te(lon,lat,bs=\"tp\" )'),data=r_dataframe)
sgrid = pd.DataFrame({'lon': xi.T, 'lat': yi.T})
griddata = mgcv.predict_gam(dollar(tps,'gam'),sgrid,se="TRUE")
interpdata[d] = griddata[0]
# print(scores
# if d == "pv":
# with open('pv.csv', 'w', newline='') as csvfile:
# fieldnames = ['lon', 'lat', 'pv']
# writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
# writer.writeheader()
# for l in range(len(X[notnan][:,0])):
# writer.writerow({'lon': X[notnan][l,0] , 'lat': X[notnan][l,1], 'pv':np.log10((np.asarray(surface["data"][d]))[notnan][l])})
# gam.summary()
# print(gam.terms[0].lam)
#gam = pygam.GAM(pygam.te(0,1,n_splines=[splines,splines])).fit(X[notnan],np.asarray(surface["data"][d])[notnan])
Xgrid = np.zeros((yi.shape[0],2))
Xgrid[:,0] = xi
Xgrid[:,1] = yi
#interpdata[d] = gam.predict(Xgrid)*sclfactor
else:
interpdata[d] = np.asarray([np.nan]*len(xi))
elif interpmethod in ["linear","nearest"] :
for d in Bar("Interpolating: ").iter(surface["data"].keys()):
notnan = np.logical_and(~np.isnan(X[:,0]),~np.isnan(surface["data"][d]))
if np.count_nonzero(notnan)>10:
Xgrid = np.zeros((yi.shape[0],2))
Xgrid[:,0] = xi
Xgrid[:,1] = yi
f = nstools.griddata(X[notnan],np.asarray(surface["data"][d])[notnan],Xgrid,method=interpmethod)
if np.isinf(f).any():
print("oh no!")
interpdata[d] = f
else:
interpdata[d] = np.asarray([np.nan]*len(xi))
else:
interpdata = surfaceSnap(surface,xi,yi)
interpsurf["data"] = interpdata
interpsurf["data"]["ids"] = finalids
interpsurf = addLatLonToSurface(interpsurf)
return interpsurf,neighbors