def removed_rss(Model, index, Y):
X = model2matix(Model)
del X[index]
X = np.transpose(np.array(X))
if len(X) <= 1:
return 999999999., None
beta = leastsquare.ls(X, Y)
Y_hat = leastsquare.predict(X, beta)
rss = np.sum((Y-Y_hat)**2)
return rss, beta
def forward(CXT, CTX, Y):
N = len(Y)
P = len(CXT)
h0 = H(np.ones(N).tolist())
Model = [h0]
while True:
min_error = 999999999.
best_h_1 = None
best_h_2 = None
best_i = None
best_p = None
best_model = None
for m in Model:
model = np.array(m.h)
for p in range(P):
for i in range(N):
X = model2matix(Model)
x_t = CXT[p,i]
t_x = CTX[p,i]
new_h_1 = x_t*model
new_h_2 = t_x*model
X.append(new_h_1.tolist())
X.append(new_h_2.tolist())
X = np.transpose(np.array(X))
#print X
beta = leastsquare.ls(X, Y)
if beta == None:
continue
error = leastsquare.RSS(X, Y, beta)
if error < min_error:
print "error:", error
min_error = error
best_h_1 = new_h_1
best_h_2 = new_h_2
best_i = i
best_p = p
best_model = m
if best_model == None:
continue
directions = copy.deepcopy(best_model.directions)
indexes = copy.deepcopy(best_model.t_idxes)
directions.append(best_p)
indexes.append(best_i)
print "d:",directions
print "ti:", indexes
Model.append(H(best_h_1.tolist(), directions, indexes, True))
Model.append(H(best_h_2.tolist(), directions, indexes, False))
if len(Model) >= 100 or min_error < 100.:
break
return Model
def draw_polynomial_pwv(train_in, train_out):
p = 3
X = polynomial(train_in)
XT = np.transpose(X)
beta = ls.ls(X, train_out)
Yhat = ls.predict(X, beta)
thegsq = ls.thegama2(Yhat, train_out, p)
beta_var = np.linalg.inv(XT.dot(X))*thegsq
print beta_var
PWV = []
for i in range(len(X)):
PWV.append(np.transpose(X[i]).dot(beta_var).dot(X[i]))
#print PWV
plt.plot(train_in, PWV)
def draw_cubic_spline_pwv(train_in, train_out):
p = 5
X = cubic(train_in)
#print X.shape
XT = np.transpose(X)
#print X
beta = ls.ls(X, train_out)
Yhat = ls.predict(X, beta)
thegsq = ls.thegama2(Yhat, train_out, p)
beta_var = np.linalg.inv(XT.dot(X))*thegsq
print beta_var
PWV = []
for i in range(len(X)):
PWV.append(np.transpose(X[i]).dot(beta_var).dot(X[i]))
#print PWV
plt.plot(train_in, PWV)
def draw_natural_spline_pwv(train_in, train_out):
p = 5
knots = np.arange(0.1, 0.9, 0.16).tolist()
knots.append(0.9)
X = nscookdata(train_in, 6, knots)
#print H
#print H.shape
XT = np.transpose(X)
beta = ls.ls(X, train_out)
Yhat = ls.predict(X, beta)
thegsq = ls.thegama2(Yhat, train_out, p)
beta_var = np.linalg.inv(XT.dot(X))*thegsq
print beta_var
PWV = []
for i in range(len(X)):
PWV.append(np.transpose(X[i]).dot(beta_var).dot(X[i]))
#print PWV
plt.plot(train_in, PWV)
product *= x_t
else:
product *= t_x
h.append(product)
if __name__ == '__main__':
data = ozone_data.load()
data_train, data_test = ozone_data.traintest(data)
X_train, Y_train = ozone_data.cook(data_train)
X_test, Y_test = ozone_data.cook(data_test)
CXT, CTX = create_basis_matrix(X_train)
Model = forward(CXT, CTX, np.array(Y_train))
#Model, beta = backward(Model, Y_train)
X = model2matix(Model)
X = np.transpose(np.array(X))
beta = leastsquare.ls(X,Y_train)
print "last error:", leastsquare.RSS(X, Y_train, beta)
print beta
fortest(X_train, Model)
#err = 0.
#for i in range(len(Y_test)):
# res = predict(beta, Model, X_test[i], X_train)
# print res, Y_test[i]
# err += (res-Y_test[i])**2
#print "test error:", err
#print "**********************************"
#ls_beta = leastsquare.ls(X_train, Y_train)
#print ls_beta
#res = leastsquare.predict(X_test, ls_beta)
#for l in range(len(Y_test)):
