source code

Question

source code

Opened this issue 6 years ago · 2 comments

Dear Yingzhen,

Thanks for sharing the dataset. Would you like to share the source code of ICML 2018 paper with us?

best,
Dong

Answer 1 · 2018-06-18T07:29:41.000Z

I would love to but unfortunately I can't: this is a project largely done at Disney research so the detailed implementation belongs to them. Sadly, open-sourcing the dataset processing script is the best thing I can do.

Answer 2 · 2018-07-27T07:05:17.000Z

Hi, I am trying to reproduce this model as the baseline and curious about basic network architectures applied in the experiment.

p(f) is from N(0, I) -- q(f | x_{1:T}) is a Gaussian with parameters computed by a Bi-LSTM. Do you add any MLPs to compute means and variances?

p(z_{t} | z_{<t}) is Gaussian with parameters computed by a LSTM
-- for full q, q(z_{1:T} | f, x_{1:T}) is Gaussian, where z_{t} = LSTM(x_{t}, z_{<t}, f). Is this right?

p(x_{t}) | z_{t}, f): concatenate [z_{t}, f] and feed it into a MLP.