Closed rlouf closed 3 years ago
This PR changes the original design of the shapes in distribution, which consisted in assigning a batch shape of 1
for single number inputs. On top of not always being practical this leads to broadcasting issues for multivariate distribution. We thus took a different approach and instead of promoting the shape of the output of distributions we promote the shape of the parameters.
As a result we also had to make a small modification in sample_forward
as it changed the output shape.
@sidravi1 I had to change the initialization of MvNormal
because of tihs.
Fixes #83. I need to go through all distributions and their tests.