python - 将相同的权重加载到新图中的多个变量

Question

我想将预训练模型中的相同变量加载到新模型中的多个变量

v1 = tf.get_variable("v1", shape=[3], initializer = tf.zeros_initializer)
inc_v1 = v1.assign(v1+1)
init_op = tf.global_variables_initializer()
saver = tf.train.Saver(v1)

with tf.Session() as sess:
  sess.run(init_op)
  sess.run(v1+1)
  save_path = saver.save(sess, "/tmp/model.ckpt")

及后记

# Create some variables.
v1 = tf.get_variable("v1", shape=[3])
v2 = tf.get_variable("v2", shape=[3])

# Add ops to save and restore all the variables.
saver = tf.train.Saver({"v1" : v1,"v1":v2})

with tf.Session() as sess:
  saver.restore(sess, "/tmp/model.ckpt")

即我希望这两个变量都从先前模型的 v1 变量进行初始化。
下面的示例崩溃，因为它表示图表不同。

Answer 1

评估原始图中变量的分配值，然后使用该值初始化新图中的新变量:

import tensorflow as tf

with tf.Graph().as_default():
    # the variable from the original graph
    v0 = tf.Variable(tf.random_normal([2, 2]))

with tf.Session(graph=v0.graph) as sess:
    sess.run(v0.initializer)
    init_val = v0.eval() # <-- evaluate the assigned value
    print('original graph:')
    print(init_val)
# original graph:
# [[-1.7466899   1.1560178 ]
#  [-0.46535382  1.7059366 ]]

# variables from new graph
with tf.Graph().as_default():
    v1 = tf.Variable(init_val) # <-- variable from new graph
    v2 = tf.Variable(init_val) # <-- variable from new graph

with tf.Session(graph=v1.graph) as sess:
    sess.run([v.initializer for v in [v1, v2]])
    print('new graph:')
    print(v1.eval())
    print(v2.eval())
# new graph:
# [[-1.7466899   1.1560178 ]
#  [-0.46535382  1.7059366 ]]
# [[-1.7466899   1.1560178 ]
#  [-0.46535382  1.7059366 ]]