I have been thinking for several days about this problems.
As before, if I initialized one weight for the encoder part, and then transpose it for the decoder part. whethere it would be the tied weights?
Today, I did a experiments, and saw that it was true. See the code below:
x = tf.placeholder(dtype=tf.float32, shape=[None, 4])
w = tf.Variable(tf.truncated_normal(shape=[4, 2], mean=0.0, stddev=0.1, dtype=tf.float32))
b = tf.Variable(tf.constant(0.0, shape=[2], dtype=tf.float32))
z = tf.nn.relu(tf.matmul(x, w) + b)
w_inv = tf.transpose(w)
b_inv = tf.Variable(tf.constant(0.0, shape=[4], dtype=tf.float32))
rec = tf.nn.relu(tf.matmul(z, w_inv) + b_inv)
loss = tf.squared_difference(x, rec)
train_op = tf.train.AdamOptimizer().minimize(loss)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
all_variables = tf.trainable_variables()
for v in all_variables:
print(v)
And the outputs:
<tf.Variable 'Variable:0' shape=(4, 2) dtype=float32_ref>
<tf.Variable 'Variable_1:0' shape=(2,) dtype=float32_ref>
<tf.Variable 'Variable_2:0' shape=(4,) dtype=float32_ref>
It is clearly, here the tied weights are successful.