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

Open
Debadri-das wants to merge 1 commit into
base: main
Choose a base branch
from
+3 −2
Open

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
1 change: 1 addition & 0 deletions CHANGELOG.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,6 +21,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0

### Fixed

- Fixed unscaled `pred_std` when `output_std=True` in `BaseGraphModel` which caused exponential loss blowup during probabilistic model training [\#347](https://github.com/mllam/neural-lam/issues/347)
- Initialize `da_forcing_mean` and `da_forcing_std` to `None` when forcing data is absent, fixing `AttributeError` in `WeatherDataset` with `standardize=True` [\#369](https://github.com/mllam/neural-lam/issues/369) @Sir-Sloth-The-Lazy

- Ensure proper sorting of `analysis_time` in `NpyFilesDatastoreMEPS._get_analysis_times` independent of the order in which files are processed with glob [\#386](https://github.com/mllam/neural-lam/pull/386) @Gopisokk
Expand Down
4 changes: 2 additions & 2 deletions neural_lam/models/base_graph_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -348,10 +348,10 @@ def predict_step(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.
# 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