vaegan.py 文件源码

def interpolate(self, x1, x2, n):
        ''' Interpolation from the latent space '''
        x1 = tf.expand_dims(x1, 0)
        x2 = tf.expand_dims(x2, 0)
        z1, _ = self._encode(x1, is_training=False)
        z2, _ = self._encode(x2, is_training=False)

        def L2norm(x):
            return tf.sqrt(tf.reduce_sum(tf.square(x), -1))

        norm1 = L2norm(z1)
        norm2 = L2norm(z2)

        theta = tf.matmul(z1/norm1, z2/norm2, transpose_b=True)

        a = tf.reshape(tf.linspace(0., 1., n), [n, 1])  # 10x1

        a1 = tf.sin((1. - a) * theta) / tf.sin(theta)
        a2 = tf.sin(a * theta) / tf.sin(theta)
        z = a1 * z1 + a2 * z2

        xh = self._generate(z, is_training=False)
        xh = tf.concat(0, [x1, xh, x2])
        return xh