def predict(X, kernelMeans, kernelSigma, kernelWeights):
# vector prediction
if(isinstance(X[0], list)):
n = len(X)
Yest = []
for i in range(n):
Yest.append(ft.output(X[i], kernelMeans, kernelSigma, kernelWeights))
Yest = np.array(Yest)
return Yest
# scalar prediction
else:
return ft.output(X, kernelMeans, kernelSigma, kernelWeights)
def updateWeights(X, y, num_kernels, kernelMeans, kernelSigma, kernelWeights):
# phase 2
B = np.identity(num_kernels)
e = y - ft.output(X, kernelMeans, kernelSigma, kernelWeights)
B, kernelSigma = ft.Phase2(X, y, e, num_kernels, B, kernelMeans,
kernelSigma, kernelWeights)
# phase 3
B = np.identity(num_kernels)
e = y - ft.output(X, kernelMeans, kernelSigma, kernelWeights)
B, kernelWeights = ft.Phase3(X, y, e, num_kernels, B, kernelMeans,
kernelSigma, kernelWeights)
return kernelMeans, kernelSigma, kernelWeights
def rolling_forecast(teX, teY, teYdate, num_kernels, kernelMeans, kernelSigma,
kernelWeights, formatter, locater):
"""
model test, rolling forecast
"""
# forecast and update
n = len(teX)
Yest = []
for i in range(n):
# forecast
Yhat = ft.output(teX[i], kernelMeans, kernelSigma, kernelWeights)
Yest.append(Yhat)
# update
kernelMeans, kernelSigma, kernelWeights = \
updateWeights(teX[i], teY[i],
num_kernels,
kernelMeans, kernelSigma, kernelWeights)
# evaluate
f = open('result.txt', 'w')
err, rmse, rsq, mae = ft.loss_with_prediction_array(teY, Yest)
print(format('rmse: %f, R2: %f, MAE: %f') % (rmse, rsq, mae))
f.write(format('rmse: %f, R2: %f, MAE: %f') % (rmse, rsq, mae) + '\n')
"""
plot
"""
dates = [datetime.datetime.strptime(d, "%Y-%m-%d").date() for d in teYdate]
plt.gca().xaxis.set_major_formatter(formatter)
plt.gca().xaxis.set_major_locator(locater)
pre = teY - err
plt.plot(dates, teY, 'r')
plt.plot(dates, pre, 'b')
plt.legend(["Test Data", "Prediction"])
plt.savefig("./kernel" + str(num_kernels) + "_prediction_graph.png")
plt.show()
f.close()
return rmse, rsq, mae
def rolling_forecast(teX, teY, num_kernels, kernelMeans, kernelSigma, kernelWeights, loop):
"""
model test, rolling forecast
"""
# forecast and update
n = len(teX)
Yest = []
for i in range(n):
# forecast
Yhat = ft.output(teX[i], kernelMeans, kernelSigma, kernelWeights)
Yest.append(Yhat)
# update
kernelMeans, kernelSigma, kernelWeights = \
updateWeights(teX[i], teY[i],
num_kernels,
kernelMeans, kernelSigma, kernelWeights,
loop)
# evaluate
f = open('result.txt', 'w')
err, rmse, rsq, mae = ft.loss_with_prediction_array(teY, Yest)
print(format('rmse: %f, R2: %f, MAE: %f') % (rmse, rsq, mae))
f.write(format('rmse: %f, R2: %f, MAE: %f') % (rmse, rsq, mae) + '\n')
# plot
pre = teY - err
plt.plot(teY, 'r')
plt.plot(pre, 'b')
plt.legend(["Test Data", "Prediction"])
plt.savefig("./kernel" + str(num_kernels) + "_prediction_graph.png")
plt.show()
f.close()
return rmse, rsq, mae
def rolling_forecast(teX, teY, num_kernels, kernelMeans, kernelSigma,
kernelWeights):
"""
model test, rolling forecast
"""
# forecast and update
n = len(teX)
Yest = []
for i in range(n):
# forecast
Yhat = ft.output(teX[i], kernelMeans, kernelSigma, kernelWeights)
Yest.append(Yhat)
# update
kernelMeans, kernelSigma, kernelWeights = \
updateWeights(teX[i], teY[i],
num_kernels,
kernelMeans, kernelSigma, kernelWeights)
# evaluate
err, rmse, rsq, mae = ft.loss_with_prediction_array(teY, Yest)
return Yest, rmse, rsq, mae
def predict(X, kernelMeans, kernelSigma, kernelWeights):
n = len(X)
Yest = []
for i in range(n):
Yest.append(ft.output(X[i], kernelMeans, kernelSigma, kernelWeights))
return Yest
def train(trX, trY, teX, teY, epochs, num_kernels, kernelMeans, kernelSigma,
kernelWeights):
"""model training"""
log = open('./log.txt', 'w')
'''
phase 2 & phase 3
learning kernel parameter
'''
# init
kernelMeans = kernelMeans[:num_kernels]
kernelSigma = kernelSigma[:num_kernels]
kernelWeights = kernelWeights[:num_kernels]
# history
epochs_arr = []
training_err = []
testing_err = []
min_err = sys.float_info.max
best_kernelMeans = None
best_kernelSigma = None
best_kernelWeights = None
best_epoch = None
for epoch in range(1, epochs + 1):
# phase 2
B = np.identity(num_kernels)
for i in range(len(trX)):
x = trX[i]
y = trY[i]
e = y - ft.output(x, kernelMeans, kernelSigma, kernelWeights)
if i % 100 == 0:
err, rmse, rsq, mae = ft.loss(trX, trY, kernelMeans,
kernelSigma, kernelWeights)
log.write(
format('Phase 2 step rmse = %f, rsq = %f\n') % (rmse, rsq))
B, kernelSigma = ft.Phase2(x, y, e, num_kernels, B, kernelMeans,
kernelSigma, kernelWeights)
# phase 3
B = np.identity(num_kernels)
for i in range(len(trX)):
x = trX[i]
y = trY[i]
e = y - ft.output(x, kernelMeans, kernelSigma, kernelWeights)
if i % 100 == 0:
err, rmse, rsq, mae = ft.loss(trX, trY, kernelMeans,
kernelSigma, kernelWeights)
log.write(
format('Phase 3 step rmse = %f, rsq = %f\n') % (rmse, rsq))
B, kernelWeights = ft.Phase3(x, y, e, num_kernels, B, kernelMeans,
kernelSigma, kernelWeights)
# check current epoch
err, rmse, rsq, mae = ft.loss(trX, trY, kernelMeans, kernelSigma,
kernelWeights)
terr, trmse, trsq, trmae = ft.loss(teX, teY, kernelMeans, kernelSigma,
kernelWeights)
training_err.append(rmse)
testing_err.append(trmse)
epochs_arr.append(epoch)
print("EPOCH {}: training rmse {}, test rmse {}".format(
epoch, rmse, trmse))
# update kernel if it is best
# metric is rmse
if (trmse < min_err):
min_err = trmse
best_epoch = epoch
best_kernelMeans = kernelMeans
best_kernelSigma = kernelSigma
best_kernelWeights = kernelWeights
print("EPOCH {} selected.".format(best_epoch))
plt.plot(epochs_arr, testing_err)
plt.savefig("./kernel" + str(num_kernels) + "_training_graph.png")
plt.show()
log.close()
return num_kernels, best_kernelMeans, best_kernelSigma, best_kernelWeights
def train(data, trX, trY, teX, teY, te_index, epochs, num_kernels, kernelMeans,
kernelSigma, kernelWeights, tau, E, P, target_P, mode):
"""model training"""
log = open('./log.txt', 'w')
'''
phase 2 & phase 3
learning kernel parameter
'''
# init
kernelMeans = kernelMeans[:num_kernels]
kernelSigma = kernelSigma[:num_kernels]
kernelWeights = kernelWeights[:num_kernels]
# history
epochs_arr = []
training_err = []
testing_err = []
max_rsq = 0
best_kernelMeans = None
best_kernelSigma = None
best_kernelWeights = None
best_epoch = None
best_Yest = None
f = open('result.txt', 'w')
for epoch in range(1, epochs + 1):
# phase 2
B = np.identity(num_kernels)
for i in range(len(trX)):
x = trX[i]
y = trY[i]
e = y - ft.output(x, kernelMeans, kernelSigma, kernelWeights)
if i % 100 == 0:
err, rmse, rsq, mae = ft.loss(trX, trY, kernelMeans,
kernelSigma, kernelWeights)
log.write(
format('Phase 2 step rmse = %f, rsq = %f\n') % (rmse, rsq))
B, kernelSigma = ft.Phase2(x, y, e, num_kernels, B, kernelMeans,
kernelSigma, kernelWeights)
# phase 3
B = np.identity(num_kernels)
for i in range(len(trX)):
x = trX[i]
y = trY[i]
e = y - ft.output(x, kernelMeans, kernelSigma, kernelWeights)
if i % 100 == 0:
err, rmse, rsq, mae = ft.loss(trX, trY, kernelMeans,
kernelSigma, kernelWeights)
log.write(
format('Phase 3 step rmse = %f, rsq = %f\n') % (rmse, rsq))
B, kernelWeights = ft.Phase3(x, y, e, num_kernels, B, kernelMeans,
kernelSigma, kernelWeights)
err, rmse, rsq, mae = ft.loss(trX, trY, kernelMeans, kernelSigma,
kernelWeights)
terr, trmse, trsq, trmae = ft.loss(teX, teY, kernelMeans, kernelSigma,
kernelWeights)
print("EPOCH {}: training r2 {}, test r2 {}".format(epoch, rsq, trsq))
if epoch == 1:
max_rsq = trsq
best_epoch = epoch
best_kernelMeans = kernelMeans
best_kernelSigma = kernelSigma
best_kernelWeights = kernelWeights
# check epoch 75 150 300
if epoch == 75 or epoch == 150 or epoch == 300:
Yest, termse, tersq, temae = evaluate(data, teX, teY, te_index,
kernelMeans, kernelSigma,
kernelWeights, tau, E, P,
target_P, mode)
print("Evaluation {}: rmse {} r2 {},MAE {}".format(
epoch, termse, tersq, temae))
f.write(
format('epoch : %d, rmse: %f, R2: %f, MAE: %f') %
(epoch, termse, tersq, temae) + '\n')
# update kernel if it is best
# metric is rsq
if (tersq > max_rsq):
max_rsq = tersq
best_Yest = Yest
best_epoch = epoch
best_kernelMeans = kernelMeans
best_kernelSigma = kernelSigma
best_kernelWeights = kernelWeights
f.close()
print("EPOCH {} selected.".format(best_epoch))
log.close()
return best_Yest, num_kernels, best_kernelMeans, best_kernelSigma, best_kernelWeights, best_epoch
def GKFN(trX, trY, teX, teY, alpha, loop, Kernel_Num):
"""model training"""
#initial model parameter
m = 0 # kernelnumber
kernelMeans = None
kernelSigma = None
kernelWeights = None
invPSI = None
#initial kernel recruiting
#첫번쨰 커널, 두번째 커널: y값이 가장 큰 index와 가장 작은 index
idx1 = np.argmax(trY)
x1 = trX[idx1]
y1 = trY[idx1]
e1 = y1
idx2 = np.argmin(trY)
x2 = trX[idx2]
y2 = trY[idx2]
e2 = y2
m += 2
kernelWeights = np.array([e1, e2])
kernelMeans = np.array([x1, x2])
dist = np.sqrt(np.sum(np.square(x1 - x2))) #x1,x2사이 거리
sig1, sig2 = alpha * dist, alpha * dist
kernelSigma = np.array([sig1, sig2])
initial_PSI = None
initial_PSI = np.ndarray(shape=(2, 2))
initial_PSI[0][0] = ft.GaussianKernel(x1, kernelMeans[0], sig1)
initial_PSI[0][1] = ft.GaussianKernel(x1, kernelMeans[1], sig2)
initial_PSI[1][0] = ft.GaussianKernel(x2, kernelMeans[0], sig1)
initial_PSI[1][1] = ft.GaussianKernel(x2, kernelMeans[1], sig2)
invPSI = lin.inv(initial_PSI)
init_y = np.array([y1, y2])
kernelWeights = np.matmul(invPSI, init_y)
#Phase 1
estv = ft.EstimatedNoiseVariance(trY)
# print(np.sqrt(estv))
trainerr = []
validerr = []
# 커널 수를 늘려가며 학습을 합니다.
while (True):
err, rmse, rsq = ft.loss(trX, trY, kernelMeans, kernelSigma,
kernelWeights)
# verr, vrmse, vrsq = ft.loss(vaX, vaY, kernelMeans, kernelSigma, kernelWeights)
# print(format('train: Phase1 : m = %d, rmse = %f, rsq = %f \nvalidation Phase1 : m = %d, rmse = %f, rsq = %f') % (m, rmse, rsq, m, vrmse, vrsq))
trainerr.append(rmse)
# validerr.append(vrmse)
if m > Kernel_Num:
break
# if (rmse**2) < estv:
# break
# if rsq > 0.9:
# break
## if np.abs(temp-rsq) < 1e-5:
# break
#
# temp = rsq
if m % 10 == 0:
print(m)
idx = np.argmax(np.abs(err), axis=0)
x = trX[idx]
y = trY[idx]
e = err[idx]
m, kernelMeans, kernelSigma, kernelWeights, invPSI = ft.Phase1(
x, y, e, m, alpha, kernelMeans, kernelSigma, kernelWeights, invPSI)
# # 커널수에 따른 에러
# plt.plot(trainerr,'r')
# plt.plot(validerr,'b')
# plt.xticks(np.arange(0,100,5)) #x축 눈금
# plt.show()
# 커널 몇개를 할것인가?
m = Kernel_Num
kernelMeans = kernelMeans[:m]
kernelSigma = kernelSigma[:m]
kernelWeights = kernelWeights[:m]
#Phase 2 & Phase3 : kernel parameter 학습
for i in range(loop):
B = None
B = np.identity(m)
for i in range(len(trX)):
x = trX[i]
y = trY[i]
e = y - ft.output(x, kernelMeans, kernelSigma, kernelWeights)
if i % 100 == 0:
err, rmse, rsq = ft.loss(trX, trY, kernelMeans, kernelSigma,
kernelWeights)
print(format('Phase 2 step rmse = %f, rsq = %f') % (rmse, rsq))
B, kernelSigma = ft.Phase2(x, y, e, m, B, kernelMeans, kernelSigma,
kernelWeights)
B = None
B = np.identity(m)
for i in range(len(trX)):
x = trX[i]
y = trY[i]
e = y - ft.output(x, kernelMeans, kernelSigma, kernelWeights)
if i % 100 == 0:
err, rmse, rsq = ft.loss(trX, trY, kernelMeans, kernelSigma,
kernelWeights)
print(format('Phase 3 step rmse = %f, rsq = %f') % (rmse, rsq))
B, kernelWeights = ft.Phase3(x, y, e, m, B, kernelMeans,
kernelSigma, kernelWeights)
"""model test"""
err, rmse, rsq = ft.loss(teX, teY, kernelMeans, kernelSigma, kernelWeights)
print(format('rmse: %f, R2: %f') % (rmse, rsq))
#pre = teY - err
#plt.plot(teY,'r')
#plt.plot(pre,'b')
#plt.show()
return rmse, rsq
