[Tensorflow] Save and load model weight to/from numpy array csv file

save_and_copy_weight.py

import tensorflow as tf
import numpy as np

# (3,4) shape weight define
W = tf.Variable([[1,2,3,4], [5,6,7,8], [9,10,11,12]], name='W')

with tf.Session() as sess:
  sess.run(tf.global_variables_initializer())
  weight = sess.run(W)
  print("Original weight:\n", weight)
  np.savetxt("weight.csv", weight, delimiter=",") # save [3,4] weights
  
tf.reset_default_graph()
call = np.loadtxt("weight.csv", delimiter= ",").astype(np.float32) # load the saved weights

# Using tf.assign for initialize the variable to numpy array
W2 = tf.get_variable("W2", [3,2])
M = tf.placeholder(tf.float32, shape = (3,2))
W2_init = W2.assign(M)

with tf.Session() as sess:
  sess.run(tf.global_variables_initializer())
  weight2 = sess.run(W2_init, feed_dict={M: call[:,:2]}) # Different variable shape, so cut [3,2] from [3,4] weights
  print("\nLoaded weight:\n", weight2)