Interpretation of each dimension on the shape about variational-autoencoder HOT 1 CLOSED

Good question! That's correct.

By default the variational distribution returns a single sample. But you could use more samples to reduce the gradient variance.

I tried to make it clearer by adding this explicitly:

    z, log_q_z = variational(x, n_samples=1)
    log_p_x_and_z = model(z, x)
    # average over sample dimension
    elbo = (log_p_x_and_z - log_q_z).mean(1)

https://github.com/altosaar/variational-autoencoder/blob/master/train_variational_autoencoder_pytorch.py#L229

from variational-autoencoder.