I got the above figure for the reconstruction loss of vanilla autoencoder. I used the learning rate of IE-2 and Adam optimization. Can we expect a graph like this for the reconstruction loss? But when I change the learning rate I got a normal loss function which decreases the loss at each iteration.