Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Refactored rational-quadratic spline transforms to run faster #44

Open
wants to merge 1 commit into
base: master
Choose a base branch
from
Open

Refactored rational-quadratic spline transforms to run faster #44

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
66 changes: 33 additions & 33 deletions nflows/transforms/splines/rational_quadratic.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,34 +29,36 @@ def unconstrained_rational_quadratic_spline(
logabsdet = torch.zeros_like(inputs)

if tails == "linear":
unnormalized_derivatives = F.pad(unnormalized_derivatives, pad=(1, 1))
constant = np.log(np.exp(1 - min_derivative) - 1)
unnormalized_derivatives[..., 0] = constant
unnormalized_derivatives[..., -1] = constant
unnormalized_derivatives = F.pad(unnormalized_derivatives, pad=(1, 1),
value=constant)

outputs[outside_interval_mask] = inputs[outside_interval_mask]
logabsdet[outside_interval_mask] = 0
outputs += outside_interval_mask*inputs
# logabsdet += outside_interval_mask*0
else:
raise RuntimeError("{} tails are not implemented.".format(tails))

if torch.any(inside_interval_mask):
Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Also, removed this check, as I suppose most of the time it will evaluate to true as you would expect some inputs to be in the domain. Let me know if you'd prefer it added back.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I see this was added in #25. With the new implementation there won't be any crashes with all-tail inputs either. (Except the computations will essentially be wasted, but I don't suppose we're expecting many calls with all-tails inputs.)

(
outputs[inside_interval_mask],
logabsdet[inside_interval_mask],
) = rational_quadratic_spline(
inputs=inputs[inside_interval_mask],
unnormalized_widths=unnormalized_widths[inside_interval_mask, :],
unnormalized_heights=unnormalized_heights[inside_interval_mask, :],
unnormalized_derivatives=unnormalized_derivatives[inside_interval_mask, :],
inverse=inverse,
left=-tail_bound,
right=tail_bound,
bottom=-tail_bound,
top=tail_bound,
min_bin_width=min_bin_width,
min_bin_height=min_bin_height,
min_derivative=min_derivative,
)
(
inside_outputs,
inside_logabsdet,
) = rational_quadratic_spline(
# Clamp inputs to the domain to prevent out of domain errors
inputs=inputs.clamp(-tail_bound, tail_bound),
unnormalized_widths=unnormalized_widths,
unnormalized_heights=unnormalized_heights,
unnormalized_derivatives=unnormalized_derivatives,
inverse=inverse,
left=-tail_bound,
right=tail_bound,
bottom=-tail_bound,
top=tail_bound,
min_bin_width=min_bin_width,
min_bin_height=min_bin_height,
min_derivative=min_derivative,
)

outputs += inside_interval_mask*inside_outputs
logabsdet += inside_interval_mask*inside_logabsdet

return outputs, logabsdet

Expand Down Expand Up @@ -87,23 +89,21 @@ def rational_quadratic_spline(

widths = F.softmax(unnormalized_widths, dim=-1)
widths = min_bin_width + (1 - min_bin_width * num_bins) * widths
cumwidths = torch.cumsum(widths, dim=-1)
cumwidths = F.pad(cumwidths, pad=(1, 0), mode="constant", value=0.0)
cumwidths = (right - left) * cumwidths + left
cumwidths[..., 0] = left
widths *= right - left
cumwidths = F.pad(widths, pad=(1, 0), mode="constant", value=left)
cumwidths = torch.cumsum(cumwidths, dim=-1)
# Make right-most knot at the right boundary
cumwidths[..., -1] = right
widths = cumwidths[..., 1:] - cumwidths[..., :-1]

derivatives = min_derivative + F.softplus(unnormalized_derivatives)

heights = F.softmax(unnormalized_heights, dim=-1)
heights = min_bin_height + (1 - min_bin_height * num_bins) * heights
cumheights = torch.cumsum(heights, dim=-1)
cumheights = F.pad(cumheights, pad=(1, 0), mode="constant", value=0.0)
cumheights = (top - bottom) * cumheights + bottom
cumheights[..., 0] = bottom
heights *= top - bottom
cumheights = F.pad(heights, pad=(1, 0), mode="constant", value=bottom)
cumheights = torch.cumsum(cumheights, dim=-1)
# Make top-most knot at the top boundary
cumheights[..., -1] = top
heights = cumheights[..., 1:] - cumheights[..., :-1]

if inverse:
bin_idx = torchutils.searchsorted(cumheights, inputs)[..., None]
Expand Down