Skip to content

fix pred_std scaling #523

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

Closed
Debadri-das wants to merge 1 commit into from
Closed

fix pred_std scaling #523

Show file tree
Hide file tree
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
2 changes: 2 additions & 0 deletions CHANGELOG.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -64,6 +64,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0

- Log every figure passed to `CustomMLFlowLogger.log_image` instead of silently dropping all but the first, using per-figure indexed keys (`{key}_{i}`) when more than one is supplied [\#499](https://github.com/mllam/neural-lam/pull/499) @Raj-Taware

- Scale predicted `pred_std` by `self.diff_std` in `output_std=True` graph step-predictors so it starts on the empirical one-step difference scale, avoiding early-training NLL/CRPS blow-up for variables whose physical step-diff std is much larger than `softplus(0) = ln(2)` ([#347](https://github.com/mllam/neural-lam/issues/347)) [\#523](https://github.com/mllam/neural-lam/pull/523) @Debadri-das

### Maintenance

- Add comprehensive type hints to `neural_lam/metrics.py` [\#447](https://github.com/mllam/neural-lam/pull/447) @sidhantpande
Expand Down
11 changes: 9 additions & 2 deletions neural_lam/models/step_predictors/graph/base.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -272,10 +272,17 @@ def forward(self, prev_state, prev_prev_state, forcing):
pred_delta_mean, pred_std_raw = net_output.chunk(
2, dim=-1
) # both (B, num_grid_nodes, d_f)
# NOTE: The predicted std. is not scaled in any way here
# Scale predicted std. with one-step difference std. so that
# the initial pred_std is on the empirical scale of the data,
# mirroring the diff_std scaling of pred_delta_mean below.
# Without this, NLL/CRPS explode in early training for
# variables whose physical step-diff std is much greater than
# softplus(0) = ln(2).
# linter for some reason does not think softplus is callable
# pylint: disable-next=not-callable
pred_std = torch.nn.functional.softplus(pred_std_raw)
pred_std = (
torch.nn.functional.softplus(pred_std_raw) * self.diff_std
)
else:
pred_delta_mean = net_output
pred_std = None
Expand Down