def rnn_predict(self, x, m, t, median_vals):
"""Impute missing data using RNN block.
Args:
- x: incomplete data
- m: mask matrix
- t: time matrix
- median_vals: median imputation for variables without observations
Returns:
- imputed_x: imputed data by rnn block
"""
# Output Initialization
imputed_x = np.zeros([self.no, self.seq_len, self.dim])
# For each feature
for f in tqdm(range(self.dim)):
temp_input = np.dstack((x[:, :, f], m[:, :, f], t[:, :, f]))
temp_input_reverse = np.flip(temp_input, 1)
forward_input = np.zeros([self.no, self.seq_len, 3])
forward_input[:, 1:, :] = temp_input[:, :(self.seq_len - 1), :]
backward_input = np.zeros([self.no, self.seq_len, 3])
backward_input[:, 1:, :] = temp_input_reverse[:, :(self.seq_len -
1), :]
save_file_name = self.save_file_directory + '/rnn_feature_' + str(
f + 1) + '/'
# Load saved model
graph = tf.Graph()
with graph.as_default():
with tf.compat.v1.Session() as sess:
sess.run(tf.compat.v1.global_variables_initializer())
tf.compat.v1.saved_model.loader.load(
sess, [tf.compat.v1.saved_model.SERVING],
save_file_name)
fw_input = graph.get_tensor_by_name('inputs:0')
bw_input = graph.get_tensor_by_name('inputs_rev:0')
output = graph.get_tensor_by_name(
'map/TensorArrayStack/TensorArrayGatherV3:0')
imputed_data = sess.run(output,
feed_dict={
fw_input: forward_input,
bw_input: backward_input
})
imputed_x[:, :, f] = (1-m[:,:,f]) * np.transpose(np.squeeze(imputed_data)) + \
m[:,:,f] * x[:,:,f]
# Initial poitn interpolation for better performance
imputed_x = initial_point_interpolation(x, m, t, imputed_x,
median_vals)
return imputed_x
def rnn_fc_predict(self, x, m, t, median_vals):
"""Impute missing data using RNN and FC.
Args:
- x: incomplete data
- m: mask matrix
- t: time matrix
- median_vals: median imputation for variables without observations
Returns:
- fc_imputed_x: imputed data using RNN and FC
"""
# rnn imputation results
rnn_imputed_x = self.rnn_predict(x, m, t, median_vals)
print('Finish M-RNN imputations with RNN')
# Reshape the data for FC predict
x = np.reshape(x, [self.no * self.seq_len, self.dim])
rnn_imputed_x = np.reshape(rnn_imputed_x,
[self.no * self.seq_len, self.dim])
m = np.reshape(m, [self.no * self.seq_len, self.dim])
save_file_name = self.save_file_directory + '/fc_feature/'
# Load saved data
graph = tf.Graph()
with graph.as_default():
with tf.compat.v1.Session() as sess:
sess.run(tf.compat.v1.global_variables_initializer())
tf.compat.v1.saved_model.loader.load(
sess, [tf.compat.v1.saved_model.SERVING], save_file_name)
x_input = graph.get_tensor_by_name('Placeholder:0')
target = graph.get_tensor_by_name('Placeholder_1:0')
mask = graph.get_tensor_by_name('Placeholder_2:0')
outputs = graph.get_tensor_by_name('Sigmoid_1:0')
fc_imputed_x = sess.run(outputs,
feed_dict={
x_input: x,
target: rnn_imputed_x,
mask: m
})
# Reshape imputed data to 3d array
fc_imputed_x = np.reshape(fc_imputed_x,
[self.no, self.seq_len, self.dim])
m = np.reshape(m, [self.no, self.seq_len, self.dim])
x = np.reshape(x, [self.no, self.seq_len, self.dim])
fc_imputed_x = fc_imputed_x * (1 - m) + x * m
fc_imputed_x = initial_point_interpolation(x, m, t, fc_imputed_x,
median_vals)
print('Finish M-RNN imputations with RNN and FC')
return fc_imputed_x
def rnn_predict(self, x, m, t, median_vals):
"""Impute missing data using RNN block.
Args:
- x: incomplete data
- m: mask matrix
- t: time matrix
- median_vals: median imputation for variables without observations
Returns:
- imputed_x: imputed data by rnn block
"""
# Check `x` dimensions.
self.no_predict, seq_len, dim = x.shape
assert seq_len == self.seq_len, "Sequence length dimension (1) of `x` passed to `rnn_predict()` "\
f"differs from that of `x` used in `rnn_train()` ({seq_len} vs {self.seq_len,})"
assert dim == self.dim, "Features dimension (2) of `x` passed to `rnn_predict()` "\
f"differs from that of `x` used in `rnn_train()` ({dim} vs {self.dim})"
# Output Initialization
imputed_x = np.zeros([self.no_predict, self.seq_len, self.dim])
# For each feature
for f in tqdm(range(self.dim)):
temp_input = np.dstack((x[:, :, f], m[:, :, f], t[:, :, f]))
temp_input_reverse = np.flip(temp_input, 1)
forward_input = np.zeros([self.no_predict, self.seq_len, 3])
forward_input[:, 1:, :] = temp_input[:, : (self.seq_len - 1), :]
backward_input = np.zeros([self.no_predict, self.seq_len, 3])
backward_input[:, 1:, :] = temp_input_reverse[:, : (self.seq_len - 1), :]
save_file_name = self.save_file_directory + "/rnn_feature_" + str(f + 1) + "/"
# Load saved model
graph = tf.Graph()
with graph.as_default():
with tf.compat.v1.Session() as sess:
sess.run(tf.compat.v1.global_variables_initializer())
tf.compat.v1.saved_model.loader.load(sess, [tf.compat.v1.saved_model.SERVING], save_file_name)
fw_input = graph.get_tensor_by_name("inputs:0")
bw_input = graph.get_tensor_by_name("inputs_rev:0")
output = graph.get_tensor_by_name("map/TensorArrayStack/TensorArrayGatherV3:0")
imputed_data = sess.run(output, feed_dict={fw_input: forward_input, bw_input: backward_input})
imputed_x[:, :, f] = (1 - m[:, :, f]) * np.transpose(np.squeeze(imputed_data)) + m[:, :, f] * x[
:, :, f
]
# Initial point interpolation for better performance
imputed_x = initial_point_interpolation(x, m, t, imputed_x, median_vals)
return imputed_x