import tensorflow as tf
from network.lenetem import LeNetEm
class TripletClustering(object):
def __init__(self, ALPHA, LR):
self.ALPHA = ALPHA
self.LR = LR
self.ph_anchor = tf.placeholder(tf.float32, [None, 28*28])
self.ph_pos = tf.placeholder(tf.float32, [None, 28*28])
self.ph_neg = tf.placeholder(tf.float32, [None, 28*28])
with tf.variable_scope("towers") as scope:
anchor_tower = LeNetEm({'data':self.ph_anchor})
scope.reuse_variables()
pos_tower = LeNetEm({'data':self.ph_pos})
neg_tower = LeNetEm({'data':self.ph_neg})
self.anchor = anchor_tower.layers['ip2']
self.positive = pos_tower.layers['ip2']
self.negative = neg_tower.layers['ip2']
self.loss = self.loss_function()
self.train_op = tf.train.RMSPropOptimizer(self.LR).minimize(self.loss)
def loss_function(self):
pos_diff = self.anchor - self.positive
neg_diff = self.anchor - self.negative
pos_dist = tf.reduce_sum(tf.mul(pos_diff, pos_diff), 1)
neg_dist = tf.reduce_sum(tf.mul(neg_diff, neg_diff), 1)
triplet = tf.add(self.ALPHA, tf.add(pos_dist, tf.neg(neg_dist)))
return tf.reduce_sum(tf.nn.relu(triplet))
def train_step(self, anchors, positives, negatives, session):
loss, _ = session.run([self.loss, self.train_op], feed_dict={self.ph_anchor: anchors, self.ph_pos: positives, self.ph_neg:negatives})
return loss
def compute_embedding_for(self, images, session):
# all towers are equal so it does not matter which tower we use
embeddings = session.run(self.anchor, feed_dict={self.ph_anchor: images})
return embeddings
