def _logits_and_labels(self, xs_ph):
xs_ph = xs_ph * 2.0 - 1.0
batch_size = tf.shape(xs_ph)[0]
xs_tile = tf.tile(tf.expand_dims(xs_ph, 1),
[1, self.num_ensemble, 1, 1, 1])
xs_tile = tf.reshape(xs_tile, (-1, ) + self.x_shape)
with slim.arg_scope(inception_v3.inception_v3_arg_scope()):
with tf.variable_scope('RandDisc'):
xs_tile = iterative_clustering_layer(
source=xs_tile,
n_clusters=self.n_clusters,
sigma=10,
alpha=10,
noise_level_1=self.noise_level,
noise_level_2=self.noise_level)
logits, _ = inception_v3.inception_v3(xs_tile,
num_classes=self.n_class,
is_training=False,
reuse=True)
logits = tf.reshape(logits, [batch_size, self.num_ensemble, -1])
logits = tf.reduce_mean(logits, axis=1)
labels = tf.cast(tf.argmax(logits, 1), tf.int32)
return logits, labels
def load(self, session, model_path):
x_input = tf.placeholder(self.x_dtype, shape=(None, ) + self.x_shape)
with slim.arg_scope(inception_v3.inception_v3_arg_scope()):
inception_v3.inception_v3(x_input,
num_classes=self.n_class,
is_training=False,
reuse=tf.AUTO_REUSE)
saver = tf.train.Saver(slim.get_model_variables(scope='InceptionV3'))
saver.restore(session, model_path)
def _logits_and_labels(self, xs_ph):
xs_ph = xs_ph * 2.0 - 1.0
with slim.arg_scope(inception_v3.inception_v3_arg_scope()):
logits, end_points = inception_v3.inception_v3(
xs_ph,
num_classes=self.n_class,
is_training=False,
reuse=tf.AUTO_REUSE)
predicted_labels = tf.argmax(end_points['Predictions'], 1)
return logits, predicted_labels
