search

tensorflow / probability

Probabilistic reasoning and statistical analysis in TensorFlow
https://www.tensorflow.org/probability/
Apache License 2.0
4.27k stars 1.11k forks source link

Is there a rsample() (reparameterization) method like in pytorch's distribution module? #1077

Closed turmeric-blend closed 4 years ago

turmeric-blend commented 4 years ago

the reparameterization concept came from Soft Actor Critic.

Exasmple of rsample()

brianwa84 commented 4 years ago

Many distributions are reparameterized, just use the sample method.

FWIW, I think this concept comes from way earlier than SAC, eg at least back to Auto-Encoding Variational Bayes, (Kingma & Welling, 2013) http://gregorygundersen.com/blog/2018/04/29/reparameterization/#kingma2013auto, and in some cases like gamma from more recent work Implicit Reparameterization Gradients, Figurnov et al).

To find out if a distribution is reparameterized, you can look at dist.reparameterization_type .

On Thu, Sep 10, 2020, 5:10 AM turmeric-blend notifications@github.com wrote:

the reparameterization concept came from Soft Actor Critic https://arxiv.org/pdf/1801.01290.pdf

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/tensorflow/probability/issues/1077, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFJFSIZI5HYWIOCKZTKMMZ3SFCJXFANCNFSM4RETNI4A .

turmeric-blend commented 4 years ago

thanks for the clarification :100: