def _build_model(self, x, y):
"""Construct the predictive model using feature and label statistics.
Args:
- x: temporal feature
- y: labels
Returns:
- model: predictor model
"""
# Parameters
dim = len(x[0, 0, :])
max_seq_len = len(x[0, :, 0])
model = tf.keras.Sequential()
model.add(
layers.Masking(mask_value=-1., input_shape=(max_seq_len, dim)))
# Stack multiple layers
for _ in range(self.n_layer - 1):
model = rnn_sequential(model,
self.model_type,
self.h_dim,
return_seq=True)
dim_y = len(y.shape)
if dim_y == 2: return_seq_bool = False
elif dim_y == 3: return_seq_bool = True
else:
raise ValueError('Dimension of y {} is not 2 or 3.'.format(
str(dim_y)))
model = rnn_sequential(model,
self.model_type,
self.h_dim,
return_seq_bool,
name='intermediate_state')
self.adam = tf.keras.optimizers.Adam(learning_rate=self.learning_rate,
beta_1=0.9,
beta_2=0.999,
amsgrad=False)
if self.task == 'classification':
if dim_y == 3:
model.add(
layers.TimeDistributed(
layers.Dense(y.shape[-1], activation='sigmoid')))
elif dim_y == 2:
model.add(layers.Dense(y.shape[-1], activation='sigmoid'))
model.compile(loss=binary_cross_entropy_loss, optimizer=self.adam)
elif self.task == 'regression':
if dim_y == 3:
model.add(
layers.TimeDistributed(
layers.Dense(y.shape[-1], activation='linear')))
elif dim_y == 2:
model.add(layers.Dense(y.shape[-1], activation='linear'))
model.compile(loss=mse_loss, optimizer=self.adam, metrics=['mse'])
return model
def _build_model(self, x, y):
"""Construct the transfer learning model using feature and label stats.
Args:
- x: temporal feature
- y: labels
Returns:
- model: transfer learning model
"""
# Parameters
dim_y = len(y.shape)
# Model initialization
model = tf.keras.Sequential()
adam = tf.keras.optimizers.Adam(learning_rate=self.learning_rate,
beta_1=0.9,
beta_2=0.999,
amsgrad=False)
# For one-shot prediction, use MLP
if dim_y == 2:
for _ in range(self.n_layer - 1):
model.add(layers.Dense(self.h_dim, activation='sigmoid'))
# For online prediction, use time-series model
elif dim_y == 3:
for _ in range(self.n_layer - 1):
model = rnn_sequential(model,
self.model_type,
self.h_dim,
return_seq=True)
# For classification
if self.task == 'classification':
if dim_y == 3:
model.add(
layers.TimeDistributed(
layers.Dense(y.shape[-1], activation='sigmoid')))
elif dim_y == 2:
model.add(layers.Dense(y.shape[-1], activation='sigmoid'))
model.compile(loss=binary_cross_entropy_loss, optimizer=adam)
# For regression
elif self.task == 'regression':
if dim_y == 3:
model.add(
layers.TimeDistributed(
layers.Dense(y.shape[-1], activation='linear')))
elif dim_y == 2:
model.add(layers.Dense(y.shape[-1], activation='linear'))
model.compile(loss=mse_loss, optimizer=adam, metrics=['mse'])
return model
def _build_model(self, x, y):
"""Construct the model using feature and label statistics.
Args:
- x: features
- y: labels
Returns:
- model: predictor model
"""
# Parameters
h_dim = self.h_dim
n_layer = self.n_layer
dim = len(x[0, 0, :])
max_seq_len = len(x[0, :, 0])
model = tf.keras.Sequential()
model.add(layers.Masking(mask_value=0.,
input_shape=(max_seq_len, dim)))
for _ in range(n_layer - 1):
model = rnn_sequential(model,
self.model_type,
h_dim,
return_seq=True)
model = rnn_sequential(model, self.model_type, h_dim, return_seq=False)
adam = tf.keras.optimizers.Adam(learning_rate=self.learning_rate,
beta_1=0.9,
beta_2=0.999,
amsgrad=False)
if self.task == 'classification':
model.add(layers.Dense(y.shape[-1], activation='sigmoid'))
model.compile(loss=binary_cross_entropy_loss, optimizer=adam)
elif self.task == 'regression':
model.add(layers.Dense(y.shape[-1], activation='linear'))
model.compile(loss=mse_loss, optimizer=adam, metrics=['mse'])
return model