def main():
x = preprocess_train_data()
x = np.reshape(x, (-1, 24))
mu, sigma = scale_fit(x[:, :12])
xs_train = scale_to(x[:, :12], mu, sigma, range(0, 12))
ys_train = scale_to(x[:, 12:], mu, sigma, range(0, 12))
xs_test = scale_to(x[1000:, :12], mu, sigma, range(0, 12))
y_true = x[1000:, 12:]
# xs_train = np.vstack([xs_train[:, 0:4], xs_train[:, 4:8], xs_train[:, 8:12]])
# ys_train = np.vstack([ys_train[:, 0:4], ys_train[:, 4:8], ys_train[:, 8:12]])
print('The shape of input data is ', xs_train.shape, ys_train.shape)
models = []
for i in range(4):
model = svm.SVR(gamma='scale', epsilon=0, C=0.4)
model.fit(xs_train, ys_train[:, i])
models.append(model)
ys_pred = np.zeros(shape=(320, 4), dtype=np.float)
for i in range(4):
ys_pred[:, i] = models[i].predict(xs_test)
y_pred = scale_back(ys_pred, mu, sigma, range(0, 4))
rmse = my_metric(y_true[:, :4], y_pred)
print('rmse: %.3f'%rmse)
# ys_pred = model.predict(xs_test)
# y_pred = scale_back(ys_pred, mu, sigma, range(0, 12))
# rmse = my_metric(y_true, y_pred)
# print('rmse: %.3f'%rmse)
# import matplotlib.pyplot as plt
# for i in range(10):
# plt.plot(y_true[i], label="true", color='green')
# plt.plot(y_pred[i], label='predicted', color='red')
# # plt.plot(x[0][:12], label='origin', color='blue')
# plt.legend(loc='upper left')
# plt.show()
xs_eval = scale_to(x[:, 12:], mu, sigma, range(0, 12))
ys_eval = np.zeros(shape=(1320, 4), dtype=np.float)
for i in range(4):
ys_eval[:, i] = models[i].predict(xs_eval)
y_eval = scale_back(ys_eval, mu, sigma, range(0, 4))
y_result = np.reshape(y_eval[:, :4], (1320*4), order='F')
write_results('Results/rmse-%d-svm'%rmse, y_result)
def main():
x = preprocess_train_data()
x = np.reshape(x, (-1, 24))
mu, sigma = scale_fit(x[:, :12])
xs_train = scale_to(x[:, :12], mu, sigma, range(0, 12))
ys_train = scale_to(x[:, 12:], mu, sigma, range(0, 12))
xs_test = scale_to(x[1000:, :12], mu, sigma, range(0, 12))
y_true = x[1000:, 12:]
# xs_train = np.vstack([xs_train[:, 0:4], xs_train[:, 4:8], xs_train[:, 8:12]])
# ys_train = np.vstack([ys_train[:, 0:4], ys_train[:, 4:8], ys_train[:, 8:12]])
print('The shape of input data is ', x.shape)
model = build_cnn(input_dim=12)
model.summary()
model.fit(np.reshape(xs_train, (-1, 12, 1)),
ys_train,
batch_size=32,
epochs=300,
validation_split=0.1,
verbose=2)
ys_pred = model.predict(np.reshape(xs_test, (-1, 12, 1)))
y_pred = scale_back(ys_pred, mu, sigma, range(0, 12))
rmse = my_metric(y_true, y_pred)
print('rmse: %.3f' % rmse)
# import matplotlib.pyplot as plt
# for i in range(10):
# plt.plot(y_true[i], label="true", color='green')
# plt.plot(y_pred[i], label='predicted', color='red')
# # plt.plot(x[0][:12], label='origin', color='blue')
# plt.legend(loc='upper left')
# plt.show()
xs_eval = scale_to(x[:, 12:], mu, sigma, range(0, 12))
ys_eval = model.predict(np.reshape(xs_eval, (-1, 12, 1)))
y_eval = scale_back(ys_eval, mu, sigma, range(0, 12))
y_result = np.reshape(y_eval[:, :4], (1320 * 4), order='F')
write_results('Results/rmse-%d-cnn1d' % rmse, y_result)
def main():
x = preprocess_train_data()
x = np.reshape(x, (-1, 24))
x, base, base_lower, base_center, base_upper = smooth(x)
mu, sigma = scale_fit(x)
xs_train = scale_to(x[:, :12], mu, sigma, range(0, 12))
ys_train = scale_to(x[:, 12:], mu, sigma, range(0, 12))
xs_test = scale_to(x[:, :12], mu, sigma, range(0, 12))
y_true = x[:, 12:]
print('The shape of input data is ', x.shape)
for i in range(10):
model = build_mlp(input_dim=12)
model.summary()
model.fit(xs_train,
ys_train,
batch_size=32,
epochs=300,
validation_split=0,
verbose=2)
ys_pred = model.predict(xs_test)
y_pred = scale_back(ys_pred, mu, sigma, range(0, 12))
rmse = my_metric(y_true, y_pred)
print('rmse: %.3f' % rmse)
# for i in range(0,1320,100):
# plt.plot(y_true[i], label="true", color='green')
# plt.plot(y_pred[i], label='predicted', color='red')
# # plt.plot(x[0][:12], label='origin', color='blue')
# plt.legend(loc='upper left')
# plt.show()
xs_eval = scale_to(x[:, 12:], mu, sigma, range(0, 12))
ys_eval = model.predict(xs_eval)
y_eval = scale_back(ys_eval, mu, sigma, range(0, 12))
# x = x + base[:, :24]
y_eval = y_eval + base[:, 24:]
# deal with long tail
for j in range(1320):
print('zoom:', j)
# print(y_eval[j][:4])
upper = np.max(y_eval[j][:4])
lower = np.min(y_eval[j][:4])
center = base_center[j][24:28]
upper_revise = base_upper[j][24:28]
lower_revise = base_lower[j][24:28]
zoom_upper = np.array([1, 1, 1, 1])
zoom_lower = np.array([1, 1, 1, 1])
if (lower < lower_revise).any():
zoom_lower = (center - lower_revise) / (center - y_eval[j][:4])
if (upper > upper_revise).any():
zoom_upper = (center - upper_revise) / (center - y_eval[j][:4])
zoom = np.vstack([zoom_lower, zoom_upper])
zoom = zoom[zoom > 0].min()
# zoom = np.min([zoom_lower, zoom_upper], axis=0)
# if base_center[j][-1] > base_center[j][0]: # k > 0
# zoom = zoom_upper
# else: # k < 0
# zoom = zoom_lower
# plt.plot(range(28), np.hstack([x[j] + base[j][:24], y_eval[j][:4]]))
y_eval[j][:4] = center - (center - y_eval[j][:4]) * zoom
# bottom = np.min(y_eval[j][:4])
# if bottom < 0:
# y_eval[j][:4] = center - (center - y_eval[j][:4]) * (center - 1) / (center - bottom)
# plt.plot(range(28), np.hstack([x[j] + base[j][:24], y_eval[j][:4]]))
# plt.plot(range(28), base_center[j][:28])
# plt.plot(range(28), base_lower[j][:28])
# plt.plot(range(28), base_upper[j][:28])
# plt.show()
# if y_eval[j][:4].min() < 0:
# print(top)
# print(y_eval[j][:4])
# print(base_revise[j][24:28])
# print(base[j][24:28])
# input()
if y_eval[:, :4].min() < -0.1:
print('Have negative number!')
y_result = np.reshape(y_eval[:, :4], (1320 * 4), order='F')
write_results('Results/mlp6-No.{}'.format(i), y_result)
def main():
x, y = preprocess_train_data()
x = np.reshape(x, (-1, 2))
y = np.reshape(y, (-1, 24))
y, base, base_lower, base_center, base_upper = smooth(y)
xs_train = np.vstack([x, x])
mu, sigma = scale_fit(y)
ys_train = scale_to(y[:, :12], mu, sigma, range(0, 12))
ys_train = np.vstack(
[ys_train, scale_to(y[:, 12:], mu, sigma, range(0, 12))])
print('The shape of input data is ', xs_train.shape, ys_train.shape)
for i in range(10):
model = build_mlp(input_dim=2)
model.summary()
model.fit([xs_train[:, :1], xs_train[:, 1:]],
ys_train,
batch_size=32,
epochs=300,
validation_split=0,
verbose=2)
xs_eval = x
ys_eval = model.predict([xs_eval[:, :1], xs_eval[:, 1:]])
y_eval = scale_back(ys_eval, mu, sigma, range(0, 12))
# x = x + base[:, :24]
y_eval = y_eval + base[:, 24:]
# deal with long tail
for j in range(1320):
print('zoom:', j)
upper = np.max(y_eval[j][:4])
lower = np.min(y_eval[j][:4])
center = base_center[j][24:28]
upper_revise = base_upper[j][24:28]
lower_revise = base_lower[j][24:28]
zoom_upper = np.array([1, 1, 1, 1])
zoom_lower = np.array([1, 1, 1, 1])
if (lower < lower_revise).any():
zoom_lower = (center - lower_revise) / (center - y_eval[j][:4])
if (upper > upper_revise).any():
zoom_upper = (center - upper_revise) / (center - y_eval[j][:4])
zoom = np.vstack([zoom_lower, zoom_upper])
zoom = zoom[zoom > 0].min()
y_eval[j][:4] = center - (center - y_eval[j][:4]) * zoom
# plt.plot(range(28), np.hstack([y[j] + base[j][:24], y_eval[j][:4]]))
# plt.plot(range(28), base_center[j][:28])
# plt.plot(range(28), base_lower[j][:28])
# plt.plot(range(28), base_upper[j][:28])
# plt.show()
if y_eval[:, :4].min() < -0.1:
print('Have negative number!')
y_result = np.reshape(y_eval[:, :4], (1320 * 4), order='F')
write_results('Results/mlp7-No.{}'.format(i), y_result)
def main():
salesVolume_list, popularity_list, comment_list, reply_list, \
car_feature_list, adcode_feature_list = preprocess_train_data()
# salesVolume_list, base = smooth(salesVolume_list)
s_mu, s_sigma = scale_fit(salesVolume_list)
s_train = scale_to(salesVolume_list[:, :12], s_mu, s_sigma, range(12))
y_train = scale_to(salesVolume_list[:, 12:], s_mu, s_sigma, range(12))
p_mu, p_sigma = scale_fit(popularity_list)
p_train = scale_to(popularity_list[:, :12], p_mu, p_sigma, range(12))
c_mu, c_sigma = scale_fit(comment_list)
c_train = scale_to(comment_list[:, :12], c_mu, c_sigma, range(12))
r_mu, r_sigma = scale_fit(reply_list)
r_train = scale_to(reply_list[:, :12], r_mu, r_sigma, range(12))
car_onehot = car_feature_list[:, :]
adcode_onehot = adcode_feature_list[:, :]
for i in range(0, 1320, 100):
plt.plot(range(1, 25),
scale_to(salesVolume_list, s_mu, s_sigma, range(12))[i])
plt.plot(range(1, 25),
scale_to(reply_list, p_mu, p_sigma, range(12))[i])
plt.show()
# for i in range(1, 6):
# s_train = np.vstack((s_train, scale_to(salesVolume_list[:, i:12+i], s_mu, s_sigma, range(i,12+i))))
# y_train = np.vstack((y_train, scale_to(salesVolume_list[:, 12+i:18+i], s_mu, s_sigma, range(i,6+i))))
# p_train = np.vstack((p_train, scale_to(popularity_list[:, i:12+i], p_mu, p_sigma, range(i,12+i))))
# c_train = np.vstack((c_train, scale_to(comment_list[:, i:12+i], c_mu, c_sigma, range(i,12+i))))
# r_train = np.vstack((r_train, scale_to(reply_list[:, i:12+i], r_mu, r_sigma, range(i,12+i))))
# car_onehot = np.vstack((car_onehot, car_feature_list[:, :]))
# adcode_onehot = np.vstack((adcode_onehot, adcode_feature_list[:, :]))
s_test = scale_to(salesVolume_list[:, :12], s_mu, s_sigma, range(12))
p_test = scale_to(popularity_list[:, :12], p_mu, p_sigma, range(12))
c_test = scale_to(comment_list[:, :12], c_mu, c_sigma, range(12))
r_test = scale_to(reply_list[:, :12], r_mu, r_sigma, range(12))
y_true = salesVolume_list[:, 12:]
for i in range(10):
model = build_mlp()
model.summary()
model.fit(
[s_train, p_train, c_train, r_train, car_onehot, adcode_onehot],
y_train,
batch_size=32,
epochs=300,
validation_split=0,
verbose=2)
ys_pred = model.predict(
[s_test, p_test, c_test, r_test, car_onehot, adcode_onehot])
y_pred = scale_back(ys_pred, s_mu, s_sigma, range(12))
rmse = my_metric(y_true, y_pred)
print('rmse: %.3f' % rmse)
# for i in range(10):
# plt.plot(y_true[i], label="true", color='green')
# plt.plot(y_pred[i], label='predicted', color='red')
# # plt.plot(x[0][:12], label='origin', color='blue')
# plt.legend(loc='upper left')
# plt.show()
s_eval = scale_to(salesVolume_list[:, 12:], s_mu, s_sigma, range(12))
p_eval = scale_to(popularity_list[:, 12:], p_mu, p_sigma, range(12))
c_eval = scale_to(comment_list[:, 12:], c_mu, c_sigma, range(12))
r_eval = scale_to(reply_list[:, 12:], r_mu, r_sigma, range(12))
ys_eval = model.predict(
[s_eval, p_eval, c_eval, r_eval, car_onehot, adcode_onehot])
y_eval = scale_back(ys_eval, s_mu, s_sigma, range(12))
y_eval = y_eval + base[:, 24:]
# deal with long tail
for j in range(1320):
print('zoom:', j)
# top = base[j][12] # > bottom bottom_revise
top = popularity_list[j][23] + base[j][23]
bottom = np.min(y_eval[j][:4])
# bottom_revise = base_revise[j][24:28] # >= 0
print(y_eval[j][:4])
if bottom < 0:
# y_eval[j][:4] = top - (top - y_eval[j][:4]) * top / (top - bottom)
# if (base_revise[j][24:28] == base[j][24:28]).all(): # k > 0
y_eval[j][:4] = top - (top - y_eval[j][:4]) * (top - 1) / (
top - bottom)
y_result = np.reshape(y_eval[:, :4], (1320 * 4), order='F')
write_results('Results/mlp5-No.{}'.format(i), y_result)