def _mlp_layer(self, input, fc_type, hidden_units, dropouts, scope_name, reuse=False):
if fc_type == "fc":
output = dense_block(input, hidden_units=hidden_units, dropouts=dropouts,
densenet=False, scope_name=scope_name,
reuse=reuse,
training=self.training, seed=self.params["random_seed"])
elif fc_type == "densenet":
output = dense_block(input, hidden_units=hidden_units, dropouts=dropouts,
densenet=True, scope_name=scope_name,
reuse=reuse,
training=self.training, seed=self.params["random_seed"])
elif fc_type == "resnet":
output = resnet_block(input, hidden_units=hidden_units, dropouts=dropouts,
cardinality=1, dense_shortcut=True, training=self.training,
reuse=reuse,
seed=self.params["random_seed"],
scope_name=scope_name)
return output
def _score_fn_inner(self, x, reuse=False):
# deep
hidden_units = [self.params["fc_dim"] * 4, self.params["fc_dim"] * 2, self.params["fc_dim"]]
dropouts = [self.params["fc_dropout"]] * len(hidden_units)
out = dense_block(x, hidden_units=hidden_units, dropouts=dropouts, densenet=False, reuse=reuse,
training=self.training, seed=self.params["random_seed"])
# score
score = tf.layers.dense(out, 1, activation=None,
kernel_initializer=tf.glorot_uniform_initializer(seed=self.params["random_seed"]))
return score
def multilayer_perceptron(x):
hidden_units = [fc_dim]
dropouts = [fc_dropout] * len(hidden_units)
out_unit = dense_block(x,
hidden_units=hidden_units,
dropouts=dropouts,
densenet=False,
reuse=False,
training=True,
seed=random_seed,
bn=True)
return out_unit