投放模型时如何修改图形，例如替换占位符

如何解决投放模型时如何修改图形，例如替换占位符

方案如下： 有一个正常的训练模型，我想替换一些输入节点，然后导出以进行服务。我试图用import_graph_def（）替换张量，但是图的新副本 不会复制变量，也无法正确存储所需的参数。 简单的代码如下： tensorflow == 1.14.0

import tensorflow as tf

input = tf.placeholder(dtype=tf.int32,shape=[None],name="id")
embedding = tf.get_variable(name="id_emb",shape=[5,10],initializer=tf.constant_initializer(1.))

lookup_output = tf.nn.embedding_lookup(embedding,input,name='lookup')
layer = tf.layers.dense(lookup_output,100)
y = tf.reduce_sum(layer,name='pred')

# just for display -- you don't need to create a Session for serialization
sess = tf.Session()
sess.run(tf.global_variables_initializer())
for var in tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES):
    print("raw var_list:%s" % var)
print("in raw graph: ",sess.run(y,{input: [3]}))

# serialize the graph
graph_def = tf.get_default_graph().as_graph_def()

# create new graph,replace the placeholder and export for serving
with tf.Graph().as_default():
    new_input = tf.placeholder(shape=(None,10),name='new_input',dtype=tf.float32)
    tf.import_graph_def(graph_def,input_map={'lookup/Identity:0': new_input},name='')

    with tf.Session() as new_sess:
        # print resource in the graph
        for var in tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES):
            # nothing print,there is no variable copy.
            print("new graph var_list:%s" % var)
        for idx,tensor in enumerate(tf.contrib.graph_editor.get_tensors(tf.get_default_graph())):
            # tensor is ok
            print("tensor_list[%s],tensor:%s" % (idx,tensor))

        # # ERROR: Attempting to use uninitialized value dense/bias
        # output = new_sess.graph.get_tensor_by_name('pred:0')
        # print("in replaced graph: ",#       new_sess.run(output,feed_dict={new_input: [[1.,1.,1.]]}))

        # # ERROR: no variables saved!!!
        # builder = tf.saved_model.builder.SavedModelBuilder('output')
        # builder.add_meta_graph_and_variables(new_sess,tags=['serving'])
        # builder.save()

# go on train work
output = sess.graph.get_tensor_by_name('pred:0')
print("go on raw graph: ",sess.run(output,{input: [3]}))

谁知道如何解决或提出任何其他建议，非常感谢。