Support fit_partial() for LightFM

CHANGELOG.md

@@ -19,6 +19,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - `load_model` function ([#213](https://github.com/MobileTeleSystems/RecTools/pull/213))
 - `model_from_config` function ([#214](https://github.com/MobileTeleSystems/RecTools/pull/214))
 - `get_cat_features` method to `SparseFeatures` ([#221](https://github.com/MobileTeleSystems/RecTools/pull/221))
+- Support `fit_partial()` for LightFM ([#223](https://github.com/MobileTeleSystems/RecTools/pull/223))
 - LightFM Python 3.12+ support ([#224](https://github.com/MobileTeleSystems/RecTools/pull/224))
 
 ### Removed

rectools/models/lightfm.py

@@ -164,20 +164,25 @@ def _from_config(cls, config: LightFMWrapperModelConfig) -> tpe.Self:
         model = model_cls(**params)
         return cls(model=model, epochs=config.epochs, num_threads=config.num_threads, verbose=config.verbose)
 
-    def _fit(self, dataset: Dataset) -> None:  # type: ignore
+    def _fit(self, dataset: Dataset) -> None:
         self.model = deepcopy(self._model)
+        self._fit_partial(dataset, self.n_epochs)
+
+    def _fit_partial(self, dataset: Dataset, epochs: int) -> None:
+        if not self.is_fitted:
+            self.model = deepcopy(self._model)
 
         ui_coo = dataset.get_user_item_matrix(include_weights=True).tocoo(copy=False)
         user_features = self._prepare_features(dataset.get_hot_user_features(), dataset.n_hot_users)
         item_features = self._prepare_features(dataset.get_hot_item_features(), dataset.n_hot_items)
         sample_weight = None if self._model.loss == "warp-kos" else ui_coo
 
-        self.model.fit(
+        self.model.fit_partial(
             ui_coo,
             user_features=user_features,
             item_features=item_features,
             sample_weight=sample_weight,
-            epochs=self.n_epochs,
+            epochs=epochs,
             num_threads=self.n_threads,
             verbose=self.verbose > 0,
         )

tests/models/test_lightfm.py

@@ -323,6 +323,33 @@ def test_second_fit_refits_model(self, dataset: Dataset) -> None:
         model = LightFMWrapperModel(model=base_model, epochs=5, num_threads=1)
         assert_second_fit_refits_model(model, dataset)
 
+    @pytest.mark.parametrize("loss", ("logistic", "bpr", "warp"))
+    @pytest.mark.parametrize("use_features_in_dataset", (False, True))
+    def test_per_epoch_partial_fit_consistent_with_regular_fit(
+        self,
+        dataset: Dataset,
+        dataset_with_features: Dataset,
+        use_features_in_dataset: bool,
+        loss: str,
+    ) -> None:
+        if use_features_in_dataset:
+            dataset = dataset_with_features
+
+        epochs = 20
+
+        base_model_1 = LightFM(no_components=2, loss=loss, random_state=1)
+        model_1 = LightFMWrapperModel(model=base_model_1, epochs=epochs, num_threads=1).fit(dataset)
+
+        base_model_2 = LightFM(no_components=2, loss=loss, random_state=1)
+        model_2 = LightFMWrapperModel(model=base_model_2, epochs=epochs, num_threads=1)
+        for _ in range(epochs):
+            model_2.fit_partial(dataset, epochs=1)
+
+        assert np.allclose(model_1.model.item_biases, model_2.model.item_biases)
+        assert np.allclose(model_1.model.user_biases, model_2.model.user_biases)
+        assert np.allclose(model_1.model.item_embeddings, model_2.model.item_embeddings)
+        assert np.allclose(model_1.model.user_embeddings, model_2.model.user_embeddings)
+
     def test_fail_when_getting_cold_reco_with_no_biases(self, dataset: Dataset) -> None:
         class NoBiasesLightFMWrapperModel(LightFMWrapperModel):
             def _get_items_factors(self, dataset: Dataset) -> Factors: