tf_multimodel.py

import tensorflow as tf

# model1
def model1_fn(inputs, labels, learning_rate=0.001):
    preds = tf.layers.dense(inputs, 1)
    cost = tf.losses.mean_squared_error(labels, preds)
    optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)
    return optimizer, cost, preds

# model2
def model2_fn(inputs, labels, learning_rate=0.001):
    preds = tf.layers.dense(inputs, 1)
    cost = tf.losses.mean_squared_error(labels, preds)
    optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)
    return optimizer, cost, preds

inputs1 = tf.placeholder(shape=(None, 1), dtype=tf.float32)
labels1 = tf.placeholder(shape=(None, 1), dtype=tf.float32)
op1, cost1, preds1 = model1_fn(inputs1, labels1)
saver = tf.train.Saver()

inputs2 = tf.placeholder(shape=(None, 1), dtype=tf.float32)
labels2 = tf.placeholder(shape=(None, 1), dtype=tf.float32)
op2, cost2, preds2 = model2_fn(inputs2, labels2)


# train model1
x1 = [[1], [2]]
y1 = [[3], [5]]
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    for i in range(1000):
        _, cost_, preds_ = sess.run(
            [op1, cost1, preds1], feed_dict={
                inputs1: x1,
                labels1: y1
            })

    print('Cost:', cost_)
    print('Preds:', preds_)

    saver.save(sess, 'models/test_model1')


class Predictor:
    def __init__(self):
        self.sess = tf.Session()
        saver.restore(self.sess, 'models/test_model1')

    def predict(self, inputs):
        return self.sess.run(preds1, feed_dict={inputs1: inputs})


predictor = Predictor()
preds = predictor.predict([[3], [7]])
print(preds) # [[ 7.3167486], [16.239666 ]]

x2 = preds # [[7], [15]]
y2 = [[7*2.+1.], [15*2.+1.]]

# train model2
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())

    for i in range(100):
        _, cost2_, preds2_ = sess.run(
            [op2, cost2, preds2], feed_dict={
                inputs2: x2,
                labels2: y2
            })

    print('Cost:', cost2_)
    print('Preds:', preds2_)


preds = predictor.predict([[3], [7]])
print(preds) # [[ 7.3167486], [16.239666 ]]

The predictor's parameters don't get initialized by the sess.run(tf.global_variables_initializer()) in another session.