Bnb/obs model#310

Open

bnb32 wants to merge 34 commits intomainfrom

Commits on Mar 24, 2026

change base loss name - not always used for physics loss
bnb32
committed
Extend obs model to enable training on just sparse high res data
bnb32
committed
Remove TODO comments regarding observation handling in Sup3rGanWithObs
bnb32
committed
refact: move proxy obs sampling logic to sampler objects. removed Sup3rGanWithObs model, since all models can now train with obs with the appropriate batch handler args.
bnb32
committed
`sparse_disc` arg added to Sup3rGan - True if discriminator can handle sparse data
bnb32
committed
test fixes
bnb32
committed
test fixes
bnb32
committed
test fixes
bnb32
committed
refact: remove unused batch handler attribute checks and improve feature consistency validation
bnb32
committed
test: add tests for sampler behavior with proxy observations
bnb32
committed
make sure single sampler and dual sampler index obs and hr features correctly.
bnb32
committed
fix: hr_feature ordering with exo features
bnb32
committed
Refactor model parameter setting and update feature handling
bnb32
committed
fix: update enhancement factor calculations and model initialization in tests
bnb32
committed
fix: update model parameter settings for multi-step GAN tests
bnb32
committed
fix: add s_enhance and t_enhance parameters to multiterm loss test
bnb32
committed
small doc string clarification
bnb32
committed
doc string edit and notebook typo removed
bnb32
committed
log message about exo rasterizer selection
bnb32
committed
*removed sparse_disc argument - use train_disc arg to skip disc calcs instead.
bnb32
committed
refactor: update daily_reduction checks to use set for improved performance; enhance documentation in local_qdm_bc and Sampler class
bnb32
committed
fix: still need `sparse_disc` flag for removing obs features from discrimination.
bnb32
committed
fix: previous logic was overwriting layer.features and including layer.exo_features
bnb32
committed
fix: lock on gradient - without it multi gpu was failing suddenly.
bnb32
committed
fix: update loss assertions to check for NaN values and adjust training logic for discriminator
bnb32
committed
fix: add tensor-only gradient calculation method to improve performance and maintain device placement
bnb32
committed
fix: enhance weight checks in Sup3rGan to handle missing generator or discriminator weights
bnb32
committed
sum -> nansum for computing stats across containers. With sparse training data sometimes mean / std is nan for an obs feature in one of the containers but not others.
bnb32
committed
add attn layer to sup3r layers
bnb32
committed
fix: correct input name in CloudMask feature class
bnb32
committed
fix: set head_node to True in ForwardPassStrategy initialization and clean up unused imports
bnb32
committed
added zarr dep
bnb32
committed

Commits on Mar 25, 2026