The Well v1.2.0 (Nov 11, 2025)
Enhancements
Metrics
Added the following metrics:
- Histogram W1 (2e035d7)
- Dynamic time warping (51e517d)
- Normalized L1 (0b7bb44)
The first two add some metrics that evaluate results between full trajectories rather than frames at varying levels of granularity. W1 can be useful for tracking whether fields have the correct distribution of intensities, though ignores local structure. DTW on the other hand can be useful for evaluating whether trajectories are simply offset by a difference in propogation speed.
Additionally:
- Added explicit epsilon parameter to metrics in the forward pass (c73b1cb)
- Updated the default video code to be slightly more adaptive and parameterizable (6911c98)
Data utilities
Tools for normalizing predictions between methods:
- Min time prediction option (ae93512) - Allows user to specify what the first step to predict is for full trajectory rollouts. Useful for comparing models with different context lengths, particularly for datasets which transition between different regimes.
- Add option to restrict number of trajectories or samples for limited training experiments (e41ded5) - Restricts data to a random subset of the original data in order to simulate low-data settings.
- Option to pass unnormalized time (eee8405) - allows passing unnormalized time - mostly useful here for writing tests. Note, be wary of using this to overfit models to specific parts of dataset with regime transitions. No data in the Well are truly time-dependent, the fact that time may be useful here is an artifact of dataset construction.
Misc
- Requirement updates to avoid wandb issue - 5b86f01
- Synced history between dev and prod.
Fixes
- Fixed HF path lookup which was broken in the normalization concatenation (e6a36c1)
Public contributions
Contributors
till-m and FloWsnr