centering.py

import tensorflow as tf
from scipy.special import logit


def centering_inference(c):
    """
    :param c: the number of target classes
    :return: a uniformly distributed class bias variable
    """
    return tf.Variable(
        initial_value=tf.constant(
            value=logit(1. / c),
            dtype=tf.float32,
            shape=(c,)),
        name="Bias",
        trainable=False)


def centering_loss(bias, y):
    """
    :param bias: bias variable allocated with :func:`centering_inference`
    :param y: actual targets with a shape of [batch_size, targets]
    :return:
    """
    targets = tf.to_float(tf.sparse_to_indicator(y, tf.shape(bias)[0]))
    targets = tf.reduce_mean(targets, 0)

    return tf.nn.sigmoid_cross_entropy_with_logits(
        logits=bias,
        labels=targets)


def optimize(c, y):
    with tf.variable_scope("Centering"):
        bias = centering_inference(c)
        loss = centering_loss(bias, y)

        optimizer = tf.train.AdagradOptimizer(learning_rate=0.5)
        minimizer = optimizer.minimize(loss, var_list=[bias])

        with tf.control_dependencies([minimizer]):
            return tf.identity(bias)