def main(CONFIG, return_m=False, force_restore=False):
X = np.load('../data/data_x.npy', allow_pickle=True) #Aggr, Points
Y = np.load('../data/data_y.npy', allow_pickle=True)
xs = np.load('../data/data_xs.npy', allow_pickle=True) #100 , 50
ys = np.load('../data/data_ys.npy', allow_pickle=True)
print(X[0].shape)
print(X[1].shape)
i = 0
#z_r = kmeans2(X[i].reshape([X[i].shape[0]*X[i].shape[1], 1]), 10, minit='points')[0]
z_r = X[i].reshape([X[i].shape[0]*X[i].shape[1], 1])
dataset = get_dataset(X, Y, z_r)
context = get_context(CONFIG, X, Y)
#elbo_model = gprn.models.GPAggr
elbo_model = gprn.models.GPRN_Aggr
if CONFIG['composite_w']:
context.parameters = gprn.composite_corrections.MagnitudeCorrection
#context.parameters = gprn.composite_corrections.CurvatureCorrection
m = gprn.GPRN(
model = elbo_model,
context = context,
data = dataset
)
m.optimise(context.train_flag, context.restore_flag)
prediction(CONFIG, m, X, xs)
def main(CONFIG, return_m=False, force_restore=False):
X = np.load('../data/data_with_features/processed_x_train.npy', allow_pickle=True) #[SAT, LAQN]
Y = np.load('../data/data_with_features/processed_y_train.npy', allow_pickle=True)
xs = np.load('../data/data_with_features/processed_x_test.npy', allow_pickle=True) #100 , 50
xs_grid = np.load('../data/data_with_features/processed_x_test_grid.npy', allow_pickle=True) #100 , 50
xs_small_grid = np.load('../data/data_with_features/processed_x_test_small_grid.npy', allow_pickle=True)
xs_small_fine_grid = np.load('../data/data_with_features/processed_x_test_small_fine_grid.npy', allow_pickle=True)
X[0] = np.array(X[0])
X[1] = np.array(X[1])
Y[0] = np.array(Y[0])
Y[1] = np.array(Y[1])
print(X[0].shape)
print(X[1].shape)
print(Y[0].shape)
print(Y[1].shape)
#ys = np.load('data/data_ys.npy')
num_z = 500
#num_z = 100
i = 0
z_r = kmeans2(X[i].reshape([X[i].shape[0]*X[i].shape[1], X[i].shape[2]]), num_z, minit='points')[0]
#z_r = X[i].reshape([X[i].shape[0]*X[i].shape[1], X[i].shape[2])
dataset = get_dataset(CONFIG, X, Y, z_r)
context = get_context(CONFIG, X, Y)
elbo_model = gprn.models.GPAggr
#context.parameters = gprn.composite_corrections.MagnitudeCorrection
m = gprn.GPRN(
model = elbo_model,
context = context,
data = dataset
)
m.optimise(context.train_flag, context.restore_flag)
prediction(CONFIG, m, X, xs, xs_grid, xs_small_grid, xs_small_fine_grid, plot_grid=True)
def run(CONFIG, test=None, r=1, return_m=False, force_restore=False):
test_id = test['id']
print('test_id:', test_id)
tf.reset_default_graph()
print(CONFIG)
CONFIG['num_layers'] = 2
num_z = CONFIG['z']
X = np.load('../data/data_x_{test_id}_{r}.npy'.format(test_id=test_id,
r=r),
allow_pickle=True)
Y = np.load('../data/data_y_{test_id}_{r}.npy'.format(test_id=test_id,
r=r),
allow_pickle=True)
print(np.min(X[0]), np.min(X[1]))
print(np.max(X[0]), np.max(X[1]))
#add empty Y columns
#Y[0] is task 1
Y[0] = np.concatenate([np.empty(Y[0].shape), Y[0]], axis=1)
Y[0][:, 0] = np.nan
Y[1] = np.concatenate([Y[1], np.empty(Y[1].shape)], axis=1)
Y[1][:, 1] = np.nan
XS = np.load('../data/data_xs_{test_id}.npy'.format(test_id=test_id),
allow_pickle=True)
XS_VIS = np.load(
'../data/data_x_vis_{test_id}.npy'.format(test_id=test_id),
allow_pickle=True)
rs = lambda x: x.reshape(x.shape[0] * x.shape[1], x.shape[2])
XS = rs(XS)
i = 0
x = X[i].reshape([X[i].shape[0] * X[i].shape[1], X[i].shape[2]])
if num_z is None or num_z > x.shape[0]:
z_x = x
else:
z_x = kmeans2(x, num_z, minit='points')[0]
i = 0
x = X[i].reshape([X[i].shape[0] * X[i].shape[1], X[i].shape[2]])
if num_z is None or num_z > x.shape[0]:
z_r = x
else:
z_r = kmeans2(x, num_z, minit='points')[0]
#z_r = X[i].reshape([X[i].shape[0]*X[i].shape[1], 1])
Z = np.array([z_r, z_x])
dataset = get_dataset(CONFIG, X, Y, z_r)
context = get_context(CONFIG, X, Y, test=test_id, r=1)
#elbo_model = gprn.models.GPRN_Aggr
elbo_model = gprn.models.GPRN_Aggr
#context.parameters = gprn.composite_corrections.MagnitudeCorrection
m = gprn.GPRN(model=elbo_model, context=context, data=dataset)
m.optimise(context.train_flag, context.restore_flag)
df_all, df_train, df_test = setup_data.get_site_1(test, root='../')
y_pred, y_var = m.predict(rs(X[1]), r=1)
#get task 0
target_task = 0
y_pred = np.expand_dims(y_pred[:, target_task], -1)
y_var = np.expand_dims(y_var[:, target_task], -1)
ys_pred, ys_var = m.predict(XS, r=1)
ys_pred = np.expand_dims(ys_pred[:, target_task], -1)
ys_var = np.expand_dims(ys_var[:, target_task], -1)
y_vis_pred, y_vis_var = m.predict(rs(XS_VIS), r=1)
y_vis_pred = np.expand_dims(y_vis_pred[:, target_task], -1)
y_vis_var = np.expand_dims(y_vis_var[:, target_task], -1)
if False:
y_pred = setup_data.denormalise_wrt(y_pred,
np.array(df_train['pm10']),
sphere_flag=True)
ys_pred = setup_data.denormalise_wrt(ys_pred,
np.array(df_train['pm10']),
sphere_flag=True)
y_vis_pred = setup_data.denormalise_wrt(y_vis_pred,
np.array(df_train['pm10']),
sphere_flag=True)
pred_y = np.concatenate([y_pred, y_var], axis=1)
pred_ys = np.concatenate([ys_pred, ys_var], axis=1)
pred_y_vis = np.concatenate([y_vis_pred, y_vis_var], axis=1)
std = np.nanstd(np.array(df_train['pm10']), axis=0)
y_var = y_var * std**2
ys_var = ys_var * std**2
y_vis_var = y_vis_var * std**2
np.save(
'../results/{prefix}_y_{test_id}_{r}'.format(
prefix=CONFIG['file_prefix'], r=r, test_id=test_id), pred_y)
np.save(
'../results/{prefix}_ys_{test_id}_{r}'.format(
prefix=CONFIG['file_prefix'], r=r, test_id=test_id), pred_ys)
np.save(
'../results/{prefix}_y_vis_{test_id}_{r}'.format(
prefix=CONFIG['file_prefix'], r=r, test_id=test_id), pred_y_vis)