Add extrapolation to the Bernstein polynomial transformation #37
Conversation
|
Just mentioning @oduerr here to notify him of the PR. |
zuko/transforms.py
Outdated
| rank = self.theta.dim() | ||
| if rank > 1: | ||
| # add singleton batch dimensions | ||
| dims = [...] + [None] * (rank - 1) | ||
| x = x[dims] |
There was a problem hiding this comment.
This should not be necessary as torch.distributions.Beta.log_prob broadcasts.
There was a problem hiding this comment.
Test fail if I remove this. Couldn't make it work without.
What would you suggest?
zuko/transforms.py
Outdated
|
|
||
| left_bound = x <= self.eps | ||
| right_bound = x >= 1 - self.eps | ||
| x_safe = torch.where(left_bound | right_bound, 0.5, x) |
There was a problem hiding this comment.
I am afraid that the 0.5 might not work with all supported versions of PyTorch / CUDA tensors. To be tested. Also, left_bound and right_bound are not great names as they are out-of-bound indicators.
There was a problem hiding this comment.
Would this be better:
x_safe = torch.where(lower_bound | upper_bound, 0.5 * torch.ones_like(x), x)
francois-rozet
changed the title
1. When to log-Scale 2. #sigmoid(17) is 1.0! --> clamping to 1e-6 1e+6
adds additional reference to BNF density forecasting paper
MArpogaus
force-pushed
the
bpf_extrapolation
branch
from
3096cb4
to
13ead7d
Compare
|
Hello @MArpogaus, I built the documentation with the new components and there was a few errors, so I made a few edits + some reformulation. I also limited the domain to [-5, 5] (like NSF) instead of [-10, 10], which should improve the results of Otherwise everything seems fine to me. Great job! I am ready to merge the PR, so tell me if you agree with my edits. |
|
Hello @francois-rozet, thanks for your final review. Your edits seam absolutely reasonable to me and improve the clarity of the doc strings. |
francois-rozet
changed the title
This adresses #36.