CTR模型如果用estimator训练出来，直接导出为savedModel格式，上线时会有冗余计算，一般需要只计算模型的一个子图即可。
假设有下面这样一个计算图：
其中：
a是一个placeholder
b、d是一个variable
c是一个tf.add节点
out是一个tf.multiply节点
一般情况下，我们保存了这个图，然后再加载，输入把a以feed_dict的形式传入参数就可以对图进行计算，现在的需求是，把其中任意一个节点替换掉，比如不计算a+b，直接给一个输入c；或者d不再是常量，而是运行是传入参数。这些需求都可以用tf的API实现。具体参考代码：
保存模型：
import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()
from tensorflow.python.framework.graph_util import convert_variables_to_constants
a = tf.placeholder(dtype=tf.float32, shape=(1, 2), name='a')
b = tf.Variable(4, dtype=tf.float32, name='b')
c = tf.add(a, b, name='c')
d = tf.Variable(2, dtype=tf.float32, name='d')
out = tf.multiply(c, d, name='out')
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
graph = convert_variables_to_constants(sess, sess.graph_def, ['out'])
tf.train.write_graph(graph, '.', 'graph_placeholder.pb', as_text=False)
tf.train.write_graph(graph, '.', 'graph_placeholder_txt.pb', as_text=True)加载模型:
import numpy as np
#import tensorflow as tf
from google.protobuf import text_format
import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()
with tf.Session() as sess:
with open('./graph_placeholder.pb', 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
data = np.array([[3.],[2.]], np.float32)
output = tf.import_graph_def(graph_def, input_map={'a:0': data}, return_elements=['out:0'])
print(sess.run(output))
with tf.Session() as sess:
with open('./graph_placeholder.pb', 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
data = np.array([[3.],[2.]], np.float32)
output = tf.import_graph_def(graph_def, input_map={'c:0': 5.}, return_elements=['out:0'])
print(sess.run(output))
with tf.Session() as sess:
with open('./graph_placeholder_txt.pb', 'rb') as f:
graph_def = tf.GraphDef()
text_format.Merge(f.read(), graph_def)
data = np.array([[7.], [3.]], np.float32)
output = tf.import_graph_def(graph_def, input_map={'a:0': data, 'd:0': 9.}, return_elements=['out:0'])
print(sess.run(output))参考资料：
https://tang.su/2017/01/export-TensorFlow-network/