infusion.py 文件源码

def __init__(self, steps=30, learning_rate=5e-4):
        def transition(z, step):
            with tf.variable_scope('transition') as vs:
                if step > 1:
                    vs.reuse_variables()
                fc1 = tf.contrib.layers.fully_connected(z, 600, activation_fn=tf.identity)
                fc1 = tf.nn.relu(tf.contrib.layers.batch_norm(fc1))
                fc1 = tf.contrib.layers.fully_connected(fc1, 600, activation_fn=tf.identity)
                fc1 = tf.nn.relu(tf.contrib.layers.batch_norm(fc1))
                mu = tf.contrib.layers.fully_connected(fc1, 784, activation_fn=tf.sigmoid)
                sig = tf.contrib.layers.fully_connected(fc1, 784, activation_fn=tf.sigmoid)
                sig = tf.add(tf.div(sig, step ** 2), 1e-4)
                #sig = tf.add(tf.scalar_mul(0.1, sig), 1e-4)
                sig = tf.sqrt(sig)
                e = tf.random_normal(tf.shape(mu), 0, 1, dtype=tf.float32)
                z_ = tf.add(mu, tf.mul(e, sig))
                z_ = tf.minimum(tf.maximum(0.0, z_), 1.0)
            return z_, mu, sig

        self.x = tf.placeholder(tf.float32, [None, 784])
        self.alpha = tf.placeholder(tf.float32, [])
        z = tf.random_normal(tf.shape(self.x), 0, 1, dtype=tf.float32)
        z = mnist_mean + mnist_std * z
        z = tf.minimum(tf.maximum(0.0, z), 1.0)
        self.rand_init = infuse(z, self.x, self.alpha)
        self.init = tf.placeholder(tf.float32, [None, 784])

        self.z = [self.init]
        z = self.z[0]
        self.loss = 0.0
        for t in range(1, steps + 1):
            z, mu, sig = transition(z, t)
            z = infuse(z, self.x, self.alpha * t)
            dist = tf.contrib.distributions.Normal(mu=mu, sigma=sig)
            #self.loss = self.loss + tf.reduce_mean(-dist.log_pdf(self.x))
            self.loss = self.loss + tf.scalar_mul(t / float(steps), tf.reduce_mean(-dist.log_pdf(self.x)))
            self.z.append(z)

        for t in range(steps + 1, steps * 2 + 1):
            z, mu, sig = transition(z, t)
            z = infuse(z, self.x, self.alpha * t)
            self.z.append(z)

        self.trainer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(self.loss)
        self.sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
        self.saver = tf.train.Saver(tf.global_variables())