def predict(self):
pred_seq_y = np.zeros_like(self.test_y_data)
# initialization
outx_table = data_processing.image_to_table(
self.test_y_data[0, :, :],
(self.G_lats - self.G_lats.mean()) / self.G_lats.std(),
(self.G_lons - self.G_lons.mean()) / self.G_lons.std(),
(((self.yday_list[0]) % 365) / 365))
for var, data in self.test_xm_data.items():
xm_img = data_processing.resolution_downward(
data[0, :, :], self.M_lats, self.M_lons, self.G_lats,
self.G_lons)
outx_table = np.concatenate(
(outx_table, xm_img.reshape(np.prod(xm_img.shape))), 1)
for var, data in self.test_xme_data.items():
xme_img = data_processing.resolution_downward(
data[0, :, :], self.Mete_lats, self.Mete_lons, self.G_lats,
self.G_lons)
outx_table = np.concatenate(
(outx_table, xme_img.reshape(np.prod(xme_img.shape))), 1)
# predict step by step
for i, day in enumerate(self.yday_list):
pred_y_flat = self.model.predict_on_batch(outx_table)
prev_g_image = pred_y_flat.reshape(self.test_y_data.shape[1:])
if i != len(self.yday_list) - 1:
outx_table = data_processing.image_to_table(
prev_g_image,
(self.G_lats - self.G_lats.mean()) / self.G_lats.std(),
(self.G_lons - self.G_lons.mean()) / self.G_lons.std(),
((day + 1 % 365) / 365))
for var, data in self.test_xm_data.items():
xm_img = data_processing.resolution_downward(
data[i + 1, :, :], self.M_lats, self.M_lons,
self.G_lats, self.G_lons)
outx_table = np.concatenate(
(outx_table, xm_img.reshape(np.prod(xm_img.shape))), 1)
for var, data in self.test_xme_data.items():
xme_img = data_processing.resolution_downward(
data[i + 1, :, :], self.Mete_lats, self.Mete_lons,
self.G_lats, self.G_lons)
outx_table = np.concatenate(
(outx_table, xme_img.reshape(np.prod(xme_img.shape))),
1)
pred_seq_y[i, :, :] = prev_g_image
seq_RMSE_mat, seq_R2_mat = self._evaluate(pred_seq_y, self.test_y_data)
np.save('seq_RMSE_mat', seq_RMSE_mat)
np.save('seq_R2_mat', seq_R2_mat)
pred_seq_y.dump('seq_pred_y')
self.test_y_data.dump('true_y')
def _generator(self, xg_data, xm_data, xme_data, yg_data, day_list):
while True:
for i, day in enumerate(day_list):
permutation = np.random.permutation(np.prod(yg_data[i].shape))
outx_table = data_processing.image_to_table(
xg_data[i, :, :],
(self.G_lats - self.G_lats.mean()) / self.G_lats.std(),
(self.G_lons - self.G_lons.mean()) / self.G_lons.std(),
(day % 365) / 365)
for var, data in xm_data.items():
xm_img = data_processing.resolution_downward(
data[i, :, :], self.M_lats, self.M_lons, self.G_lats,
self.G_lons)
outx_table = np.concatenate(
(outx_table, xm_img.reshape(
(np.prod(xm_img.shape), 1))), 1)
for var, data in xme_data.items():
xme_img = data_processing.resolution_downward(
data[i, :, :], self.Mete_lats, self.Mete_lons,
self.G_lats, self.G_lons)
outx_table = np.concatenate(
(outx_table,
xme_img.reshape((np.prod(xme_img.shape), 1))), 1)
outy = yg_data[i].reshape((np.prod(yg_data[i].shape), 1))
yield outx_table[permutation], outy[permutation]
def _flatten(self, xg_data, xm_data, xme_data, yg_data, day_list):
permut = np.random.permutation(np.prod(xg_data.shape))
x = np.zeros((np.prod(xg_data.shape),
4 + len(self.merra_var) + len(self.mete_var)))
y = np.zeros((np.prod(xg_data.shape), 1))
for i, day in enumerate(day_list):
outx_table = data_processing.image_to_table(
xg_data[i, :, :],
(self.G_lats - self.G_lats.mean()) / self.G_lats.std(),
(self.G_lons - self.G_lons.mean()) / self.G_lons.std(),
(day % 365) / 365)
for var, data in xm_data.items():
if var in self.merra_var:
xm_img = data_processing.resolution_downward(
data[i, :, :], self.M_lats, self.M_lons, self.G_lats,
self.G_lons)
outx_table = np.concatenate(
(outx_table, xm_img.reshape(
(np.prod(xm_img.shape), 1))), 1)
for var, data in xme_data.items():
if var in self.mete_var:
xme_img = data_processing.resolution_downward(
data[i, :, :], self.Mete_lats, self.Mete_lons,
self.G_lats, self.G_lons)
outx_table = np.concatenate(
(outx_table,
xme_img.reshape((np.prod(xme_img.shape), 1))), 1)
outy = yg_data[i].reshape((np.prod(yg_data[i].shape), 1))
x[i * np.prod(yg_data[i].shape):(i + 1) *
np.prod(yg_data[i].shape)] = outx_table
y[i * np.prod(yg_data[i].shape):(i + 1) *
np.prod(yg_data[i].shape)] = outy
return x[permut], y[permut]
def predict(self):
pred_y = self.model.predict(self.test_x_data)
pred_y = pred_y.reshape(self.test_y_data_3d.shape)
RMSE_mat, R2_mat = self.evaluate(pred_y, self.test_y_data_3d)
np.save('RMSE_mat', RMSE_mat)
np.save('R2_mat', R2_mat)
# TODO step-by-step predict
pred_seq_y = np.zeros_like(self.test_y_data_3d)
prev_image = self.test_x_data_3d[0, :, :]
for i in range(len(self.test_index[1:])):
input_x_data = data_processing.image_to_table(
prev_image, self.G_lats, self.G_lons,
(self.test_index[i + 1] % 365) / 365)
pred_seq_sub_y = self.model.predict_on_batch(input_x_data)
prev_image = pred_seq_sub_y.reshape(prev_image.shape)
pred_seq_y[i, :, :] = prev_image
seq_RMSE_mat, seq_R2_mat = self.evaluate(pred_seq_y,
self.test_y_data_3d)
np.save('seq_RMSE_mat', seq_RMSE_mat)
np.save('seq_R2_mat', seq_R2_mat)
np.save('seq_pred_y', pred_seq_y)
np.save('true_y', self.test_y_data_3d)
def predict(self):
if self.split_type == 'random_split':
pred_y = self.model.predict(self.test_x_data)
pred_y = pred_y.reshape(self.test_y_data_3d.shape)
RMSE_mat, R2_mat = self.evaluate(pred_y, self.test_y_data_3d)
np.save('RMSE_mat', RMSE_mat)
np.save('R2_mat', R2_mat)
else:
pred_y = self.model.predict(self.test_x_data)
pred_y = pred_y.reshape(self.test_y_data_3d.shape)
RMSE_mat, R2_mat = self.evaluate(pred_y, self.test_y_data_3d)
np.save('RMSE_mat', RMSE_mat)
np.save('R2_mat', R2_mat)
# TODO step-by-step predict
test_xg_data = self.g_data[self.test_index, :, :]
test_xm_data = self.m_data[self.test_index, :, :]
pred_seq_y = np.zeros_like(self.test_y_data_3d)
prev_g_image = test_xg_data[0,:,:]
prev_m_image = test_xm_data[0,:,:]
for i in range(len(self.test_index[1:])):
m_list = np.zeros((np.prod(prev_g_image.shape), 1))
input_x_data = data_processing.image_to_table(prev_g_image, self.G_lats, self.G_lons,
(self.test_index[i + 1] % 365) / 365)
m_list[:, 0] = data_processing.resolution_downward(prev_m_image, self.M_lats, self.M_lons,
self.G_lats, self.G_lons).\
reshape(np.prod(prev_g_image.shape))
input_x_data = np.concatenate((input_x_data, m_list), 1)
pred_seq_sub_y = self.model.predict_on_batch(input_x_data)
prev_g_image = pred_seq_sub_y.reshape(prev_g_image.shape)
prev_m_image = test_xm_data[i+1,:,:]
pred_seq_y[i, :, :] = prev_g_image
seq_RMSE_mat, seq_R2_mat = self.evaluate(pred_seq_y, self.test_y_data_3d)
np.save('seq_RMSE_mat', seq_RMSE_mat)
np.save('seq_R2_mat', seq_R2_mat)
pred_seq_y.dump('seq_pred_y')
self.test_y_data_3d.dump('true_y')