как переписать последние слои, чтобы переобучить мою предварительно обученную модель

Question

Я восстановил предварительно обученную модель и хочу переписать последние слои.Но я не знаю, как извлечь определенные слои и переписать их.Это загруженный код:

with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options, log_device_placement=False)) as sess:

    saver = tf.train.import_meta_graph(meta_path)
    saver.restore(sess, model_path)
    graph = tf.get_default_graph()

    x = graph.get_tensor_by_name('Placeholder:0')
    y = graph.get_tensor_by_name('Placeholder_1:0')

Это предварительно обученная модель сети:

def fw_net(x, training=True):
    with tf.variable_scope('fw_net', reuse=tf.AUTO_REUSE):
        with tf.variable_scope('fc1', reuse=tf.AUTO_REUSE):
            x = ms.fc(x, 5*5*256, name='fc')
            x = ms.bn(x, training=training, name='bn')
            x = ms.activation(x, relu=True, name='relu')
            x = tf.reshape(x, [-1, 5, 5, 256])

        with tf.variable_scope('id_blk1', reuse=tf.AUTO_REUSE):
            x = id_blk(x, 256, [3, 3], training)

        with tf.variable_scope('id_blk2', reuse=tf.AUTO_REUSE):
            x = id_blk(x, 256, [3, 3], training)        # [-1, 5, 5, 256]

        with tf.variable_scope('t_conv1', reuse=tf.AUTO_REUSE):
            x = ms.t_conv2d(x, 128, [2, 2], 2, name='t_c')
            x = ms.bn(x, training=training, name='bn')
            x = ms.activation(x, relu=True, name='relu')

        with tf.variable_scope('id_blk3', reuse=tf.AUTO_REUSE):
            x = id_blk(x, 128, [3, 3], training)

        with tf.variable_scope('id_blk4', reuse=tf.AUTO_REUSE):
            x = id_blk(x, 128, [3, 3], training)        # [-1, 10, 10, 128]

        with tf.variable_scope('t_conv2', reuse=tf.AUTO_REUSE):
            x = ms.t_conv2d(x, 64, [2, 2], 2, 't_c')
            x = ms.bn(x, training=training, name='bn')
            x = ms.activation(x, relu=True, name='relu')

        with tf.variable_scope('id_blk5', reuse=tf.AUTO_REUSE):
            x = id_blk(x, 64, [3, 3], training)

        with tf.variable_scope('id_blk6', reuse=tf.AUTO_REUSE):
            x = id_blk(x, 64, [3, 3], training)         # [-1, 20, 20, 64]

        with tf.variable_scope('t_conv3', reuse=tf.AUTO_REUSE):
            x = ms.t_conv2d(x, 32, [2, 2], 2, name='t_c')
            x = ms.bn(x, training=training, name='bn')
            x = ms.activation(x, relu=True, name='relu')

        with tf.variable_scope('id_blk7', reuse=tf.AUTO_REUSE):
            x = id_blk(x, 32, [3, 3], training)

        with tf.variable_scope('id_blk8', reuse=tf.AUTO_REUSE):
            x = id_blk(x, 32, [3, 3], training)       # [-1, 40, 40, 32]

        x = tf.reshape(x, [-1, 40*40*32])

        with tf.variable_scope('output', reuse=tf.AUTO_REUSE):
            x = ms.fc(x, units=603, name='fc')
            print(x.name)

        return x

как заморозить предварительно обученную модель и переобучить переписанные слои?