pot.py 文件源码

def discriminator(self, opts, input_, prefix='DISCRIMINATOR', reuse=False):
        """Discriminator for the GAN objective

        """
        num_units = opts['d_num_filters']
        num_layers = opts['d_num_layers']
        nowozin_trick = opts['gan_p_trick']
        # No convolutions as GAN happens in the latent space
        with tf.variable_scope(prefix, reuse=reuse):
            hi = input_
            for i in range(num_layers):
                hi = ops.linear(opts, hi, num_units, scope='h%d_lin' % (i+1))
                hi = tf.nn.relu(hi)
            hi = ops.linear(opts, hi, 1, scope='final_lin')
        if nowozin_trick:
            # We are doing GAN between our model Qz and the true Pz.
            # We know analytical form of the true Pz.
            # The optimal discriminator for D_JS(Pz, Qz) is given by:
            # Dopt(x) = log dPz(x) - log dQz(x)
            # And we know exactly dPz(x). So add log dPz(x) explicitly 
            # to the discriminator and let it learn only the remaining
            # dQz(x) term. This appeared in the AVB paper.
            assert opts['latent_space_distr'] == 'normal'
            sigma2_p = float(opts['pot_pz_std']) ** 2
            normsq = tf.reduce_sum(tf.square(input_), 1)
            hi = hi - normsq / 2. / sigma2_p \
                    - 0.5 * tf.log(2. * np.pi) \
                    - 0.5 * opts['latent_space_dim'] * np.log(sigma2_p)
        return hi