River is migrating its mini-batch methods to Narwhals so that pandas, Polars, PyArrow, and other supported eager dataframe backends can be used while preserving the caller's native backend.
Deep-river's current mini-batch paths are pandas-specific and directly use pd.DataFrame, pd.Series, .values, .index, and pandas output constructors.
Proposed scope
Start with the main supervised estimators and shared conversion boundary:
deep_river.classification.Classifierdeep_river.regression.Regressor- Shared dataframe-to-tensor utilities
The implementation should:
- Accept Narwhals-supported dataframe and series inputs.
- Convert inputs to NumPy only at the tensor boundary.
- Return predictions in the same backend as the input.
- Preserve the pandas index.
- Keep existing pandas behavior unchanged.
- Support reordered, emerging, and disappearing features.
Deep-river currently supports river>=0.25,<0.26, while the dataframe helpers on River's main branch are not part of River 0.25. The initial implementation should therefore use a small internal boundary utility and a direct Narwhals dependency instead of importing unreleased River helpers.
Autoencoders, multi-target estimators, and rolling estimators can be migrated in follow-up changes once the shared approach is established.
Tests
Add cross-backend tests covering:
- pandas, Polars, and PyArrow
- Prediction parity between backends
- Backend-preserving return types
- pandas index preservation
- Reordered columns
- Emerging and disappearing features
- Missing values
Related upstream work
River is migrating its mini-batch methods to Narwhals so that pandas, Polars, PyArrow, and other supported eager dataframe backends can be used while preserving the caller's native backend.
Deep-river's current mini-batch paths are pandas-specific and directly use
pd.DataFrame,pd.Series,.values,.index, and pandas output constructors.Proposed scope
Start with the main supervised estimators and shared conversion boundary:
deep_river.classification.Classifierdeep_river.regression.RegressorThe implementation should:
Deep-river currently supports
river>=0.25,<0.26, while the dataframe helpers on River's main branch are not part of River 0.25. The initial implementation should therefore use a small internal boundary utility and a direct Narwhals dependency instead of importing unreleased River helpers.Autoencoders, multi-target estimators, and rolling estimators can be migrated in follow-up changes once the shared approach is established.
Tests
Add cross-backend tests covering:
Related upstream work