Wasserstein distance between prior and posterior

[sadegh-aa]

Hi,

I'm trying to find the part of the code that attempts to compute the Wasserstein distance between prior and posterior (as in Eq. 5 in your ICLR paper), but couldn't find it. Would you please point to the part of the code for this distance?

Moreover, I found that the latent variables are computed directly from the model (e.g., a fully connected layer) rather than predicting \mu and \sigma and then sampling from that distribution, as stated in Eq. 3 and Eq. 4. Would you please clarify this?

Thanks

[guxd]

The Wasserstein distance is implicitly minimized by training a WGAN.
The latent variable is implicitly sampled with the reparametrize trick.

[sadegh-aa]

If the latent variable is sampled with the reparam trick, it should be a mu and sigma and epsilon, where
latent = mu + sigma \odot epsilon

But, as far as I understood, your model directly generates a latent variable. Is that correct?

[guxd]

DialogWAE/modules.py

Line 186 in 29f206a

z = epsilon * std + mu