from realsafe.defense.jpeg_compression import jpeg_compression
from realsafe.utils import get_res_path
import inception_v3
MODEL_PATH = get_res_path('./imagenet/inception_v3.ckpt')
def load(session):
model = InceptionV3Jpeg()
model.load(session, MODEL_PATH)
return model
@jpeg_compression(quality=75)
class InceptionV3Jpeg(inception_v3.InceptionV3):
pass
if __name__ == '__main__':
inception_v3.download(MODEL_PATH)
indices = tf.reshape(tf.multinomial(sigma * distances, 1),
[batch_size])
weights = tf.expand_dims(tf.one_hot(indices, depth=n_samples), 2)
selected_points = tf.expand_dims(
tf.reduce_sum(weights * samples, axis=1), 1)
centroids.append(selected_points)
return tf.concat(centroids, axis=1)
def rgb_clustering(images, centroids, alpha):
_, width, height, channel = images.get_shape().as_list()
# Gaussian mixture clustering
cluster_num = centroids.get_shape().as_list()[1]
reshaped_images = tf.reshape(images, [-1, width, height, 1, channel])
reshaped_centroids = tf.reshape(centroids,
[-1, 1, 1, cluster_num, channel])
distances = tf.reduce_sum(tf.square(reshaped_centroids - reshaped_images),
axis=4)
logits = tf.clip_by_value(-alpha * distances, -200, 200)
probs = tf.expand_dims(tf.nn.softmax(logits), 4)
new_images = tf.reduce_sum(reshaped_centroids * probs, axis=3)
# update cluster centers
new_centroids = tf.reduce_sum(reshaped_images * probs, axis=[1, 2]) / (
tf.reduce_sum(probs, axis=[1, 2]) + 1e-16)
return new_images, new_centroids
if __name__ == '__main__':
from inception_v3 import download
download(MODEL_PATH)
