Doubts

[OleguerCanal]

I don't fully understand the following things from the paper:
Implementation

• In the re-parametrization trick part, do we estimate the means and variances of z distribution or just means and assume 1 std dev? If estimating both, do they have a 2-headed encoder, one head for means and one for variances?

Method

• What if the total system energy is not constant? (e.g. friction)

[fedetask]

Other things that are unclear:
Implementation

• How is the result of Hamiltonian network convolutional layers turned into a scalar? Should we use a linear layer at the end?
• What is the shape of the latent encoding z? The paper says it has 48 channels but doesn't say anything about the H and W dimensions. Should we assume it is 32x32?

[fedetask]

• Does the decoder need a final sigmoid activation? I would say so since we want values to always lie in [0, 1]. But maybe this goes against the use of L2 loss.