guxd
Wasserstein distance between prior and posterior
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
The Wasserstein distance is implicitly minimized by training a WGAN. The latent variable is implicitly sampled with the reparametrize trick.
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?
https://github.com/guxd/DialogWAE/blob/29f206af05bfe5fe28fec4448e208310a7c9258d/modules.py#L186