Sample two layer perceptron in Tensorflow v1 (https://www.easy-tensorflow.com/tf-tutorials/basics/save-and-restore)

two_layer_perceptron_tensorflow.py

import tensorflow as tf


tf.reset_default_graph()

img_h = img_w = 28
img_size_flat = img_h * img_w

n_classes = 10

from tensorflow.examples.tutorials.mnist import input_data
data = input_data.read_data_sets('MNIST/', one_hot=True)

learning_rate = 0.001
batch_size = 100
num_steps = 100

x = tf.placeholder(tf.float32, shape=[None, img_size_flat], name='X')
y = tf.placeholder(tf.float32, shape=[None, n_classes], name='Y')

W = tf.get_variable('W',
                    dtype=tf.float32,
                    shape=[img_size_flat, n_classes],
                    initializer=tf.truncated_normal_initializer(stddev=0.01)
)
b = tf.get_variable('b',
                    dtype=tf.float32,
                    initializer=tf.constant(0., shape=[n_classes], dtype=tf.float32)
)

output_logits = tf.matmul(x, W) + b
y_pred = tf.nn.softmax(output_logits)

loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels=y, logits=output_logits))
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate, name='Adam-op').minimize(loss)
correct_prediction = tf.equal(tf.argmax(output_logits, 1), tf.argmax(y, 1), name='correct_pred')
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32), name='accuracy')

saver = tf.train.Saver()

sess = tf.Session()
sess.run(tf.global_variables_initializer())
for i in range(num_steps):
    x_batch, y_true_batch = data.train.next_batch(batch_size)
    feed_dict_train = {x: x_batch, y: y_true_batch}
    sess.run(optimizer, feed_dict=feed_dict_train)
    saved_path = saver.save(sess, './saved_variable')


tf.reset_default_graph()
imported_graph = tf.train.import_meta_graph('saved_variable.meta')

with tf.Session() as sess:
    imported_graph.restore(sess, './saved_variable')
    weight, bias = sess.run(['W:0', 'b:0'])
    print(weight)
    print(bias)