def call(self, inputs, training=None):
if self.input_features is not None:
features = self.input_features(inputs)
features = self.input_bn(features, training=training)
else:
features = inputs
batch_size = tf.shape(features)[0]
self._step_feature_selection_masks = []
self._step_aggregate_feature_selection_mask = None
# Initializes decision-step dependent variables.
output_aggregated = tf.zeros([batch_size, self.output_dim])
masked_features = features
mask_values = tf.zeros([batch_size, self.num_features])
aggregated_mask_values = tf.zeros([batch_size, self.num_features])
complemantary_aggregated_mask_values = tf.ones(
[batch_size, self.num_features])
total_entropy = 0.0
entropy_loss = 0.
for ni in range(self.num_decision_steps):
# Feature transformer with two shared and two decision step dependent
# blocks is used below.
transform_f1 = self.transform_f1(masked_features, training=training)
transform_f1 = glu(transform_f1, self.feature_dim)
transform_f2 = self.transform_f2(transform_f1, training=training)
transform_f2 = (glu(transform_f2, self.feature_dim) +
transform_f1) * tf.math.sqrt(0.5)
transform_f3 = self.transform_f3(transform_f2, training=training)
transform_f3 = (glu(transform_f3, self.feature_dim) +
transform_f2) * tf.math.sqrt(0.5)
transform_f4 = self.transform_f4(transform_f3, training=training)
transform_f4 = (glu(transform_f4, self.feature_dim) +
transform_f3) * tf.math.sqrt(0.5)
if (ni > 0):
decision_out = tf.nn.relu(transform_f4[:, :self.output_dim])
# Decision aggregation.
output_aggregated += decision_out
# Aggregated masks are used for visualization of the
# feature importance attributes.
scale_agg = tf.reduce_sum(decision_out, axis=1, keepdims=True)
scale_agg = scale_agg / tf.cast(self.num_decision_steps - 1, tf.float32)
aggregated_mask_values += mask_values * scale_agg
features_for_coef = (transform_f4[:, self.output_dim:])
if (ni < (self.num_decision_steps - 1)):
# Determines the feature masks via linear and nonlinear
# transformations, taking into account of aggregated feature use.
mask_values = self.transform_coef(features_for_coef, training=training)
mask_values *= complemantary_aggregated_mask_values
mask_values = sparsemax(mask_values, axis=-1)
# Relaxation factor controls the amount of reuse of features between
# different decision blocks and updated with the values of
# coefficients.
complemantary_aggregated_mask_values *= (
self.relaxation_factor - mask_values)
# Entropy is used to penalize the amount of sparsity in feature
# selection.
total_entropy += tf.reduce_mean(
tf.reduce_sum(
-mask_values * tf.math.log(mask_values + self.epsilon), axis=1)) / (
tf.cast(self.num_decision_steps - 1, tf.float32))
# Add entropy loss
entropy_loss = total_entropy
# Feature selection.
masked_features = tf.multiply(mask_values, features)
# Visualization of the feature selection mask at decision step ni
# tf.summary.image(
# "Mask for step" + str(ni),
# tf.expand_dims(tf.expand_dims(mask_values, 0), 3),
# max_outputs=1)
mask_at_step_i = tf.expand_dims(tf.expand_dims(mask_values, 0), 3)
self._step_feature_selection_masks.append(mask_at_step_i)
else:
# This branch is needed for correct compilation by tf.autograph
entropy_loss = 0.
# Adds the loss automatically
self.add_loss(self.sparsity_coefficient * entropy_loss)
# Visualization of the aggregated feature importances
# tf.summary.image(
# "Aggregated mask",
# tf.expand_dims(tf.expand_dims(aggregated_mask_values, 0), 3),
# max_outputs=1)
agg_mask = tf.expand_dims(tf.expand_dims(aggregated_mask_values, 0), 3)
self._step_aggregate_feature_selection_mask = agg_mask
return output_aggregated
def test_sparsemax():
x = tf.random.uniform([5, 20], -1., 1.)
out = custom_objs.sparsemax(x, axis=-1)
row_zero_sum = tf.reduce_sum(tf.cast(tf.equal(out, 0.0), tf.float32), axis=-1)
assert tf.reduce_all(row_zero_sum > 0.)