def test_l2_penalty():
g = tf.Graph()
with g.as_default():
w1 = tf.Variable(tf.ones((4, 4)))
w2 = tf.Variable(tf.ones((3, 3)))
b = tf.Variable(tf.ones((3, 3)), name='bias')
l2 = modeling.l2_penalty(0.5)
init()
l2_val = sess.run(l2)
assert l2_val == 12.5
def cost(y, logits, regularize=False, l2_weight=0.01):
batch_size = logits.get_shape().as_list()[0]
with tf.name_scope('cost'):
with tf.name_scope('xentropy'):
xentropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(
logits, tf.cast(modeling.onehot(y, NUM_CLASSES), 'float')))
if regularize:
l2 = modeling.l2_penalty(l2_weight)
tot_cost = xentropy
if regularize:
tot_cost = tf.add(xentropy, l2)
tf.scalar_summary('cost', tot_cost)
return tot_cost
