can I use your approach on deep conv ae ?
iperov opened this issue · 1 comments
iperov commented
I investigated your tf code.
so all what I need for UAE is
- transform latent space code
mean = self.encoder(x, reuse=reuse)
eps = tf.random_normal(tf.shape(mean), 0, 1, dtype=tf.float32)
z = tf.add(mean, tf.multiply(std, eps))
= self.decoder(z, reuse=reuse)
- remove all bias from decoder layers?
if I use your approach on deep conv ae, will I get better accuracy?
Should I remove biases from all conv and dense layers after latent space ?
iperov commented
ah sorry found convuae.py