fix_reverse_out_bound_quadratic_spline

[LiaoShiqi97]

Issue Description
quadratic spline in pyro.distributions.transforms.Spline does not consider outside_interval input at reverse.

According to the Neural spline flow(https://arxiv.org/abs/1906.04032), monotonic rational-quadratic transforms sets the boundary derivatives as 1 to match the linear tails. However, when it takes the y(sample from the target distribution ) to reverse back to the x(sample from the base distribution) in the function _monotonic_rational_spline, it does not consider the scenario where y is out of bound, which would render a negative discriminant and return Error.

Environment
For any bugs, please provide the following:

python version: 3.8.7
PyTorch version: 1.12.1+cpu
Pyro version: 1.8.2
Solution
I replace a fake discriminant for the out-of-bound input.
Origin:
discriminant = b.pow(2) - 4 * a * c
assert (discriminant >= 0).all()

Changed:
discriminant = b.pow(2) - 4 * a * c
discriminant[outside_interval_mask] = 0 # added to make sure outputs[outside_interval_mask] = inputs[outside_interval_mask]
assert (discriminant >= 0).all()

[fritzo]

I think you'll need two spaces before the comment, and possibly to move the line comment to a separate line to satisfy max line length requirements. You can locally run make lint to see lint errors.

Also you might need to use a non-inplace version like

discriminant = discriminant.masked_fill(outside_interval_mask, 0)

[LiaoShiqi97]

Thanks for your comment. I am new to pulling a request. So in this case, it is because of my non-standard comment that my request fails, right?

[fritzo]

I believe so, we use the black automatic style checker to ensure all code follows the same style. This helps everyone collaborate by making the code more uniform.

[LiaoShiqi97]

I get it. Thanks. I will edit it soon

[LiaoShiqi97]

I have re-pulled it. Please check it.