Make TableConfig and FeatureConfig unhashable.

`TableConfig` contains a keras `Optimizer` and a keras `Initializer`, which are not hashable and even mutable in the case of the optimizer. The "unsafe hash" was actually unsafe. Instead of using `TableConfig`s as keys directly, we use the id of the `TableConfig`, which is correct because we are detecting reused instances.

Also:
- renamed TensorFlow config conversion functions to be more consistent with the JAX ones
- added unit tests for the TensorFlow config conversion functions

Author: hertschuh

keras_rs/src/layers/embedding/base_distributed_embedding.py

@@ -822,13 +822,13 @@ def _default_device_init(
         table_stacking: str | Sequence[Sequence[str]],
     ) -> None:
         del table_stacking
-        table_to_embedding_layer: dict[TableConfig, EmbedReduce] = {}
+        table_config_id_to_embedding_layer: dict[int, EmbedReduce] = {}
         self._default_device_embedding_layers: dict[str, EmbedReduce] = {}
 
         for path, feature_config in feature_configs.items():
-            if feature_config.table in table_to_embedding_layer:
+            if id(feature_config.table) in table_config_id_to_embedding_layer:
                 self._default_device_embedding_layers[path] = (
-                    table_to_embedding_layer[feature_config.table]
+                    table_config_id_to_embedding_layer[id(feature_config.table)]
                 )
             else:
                 embedding_layer = EmbedReduce(
@@ -838,7 +838,9 @@ def _default_device_init(
                     embeddings_initializer=feature_config.table.initializer,
                     combiner=feature_config.table.combiner,
                 )
-                table_to_embedding_layer[feature_config.table] = embedding_layer
+                table_config_id_to_embedding_layer[id(feature_config.table)] = (
+                    embedding_layer
+                )
                 self._default_device_embedding_layers[path] = embedding_layer
 
     def _default_device_build(
@@ -1013,8 +1015,8 @@ def get_config(self) -> dict[str, Any]:
 
         # The serialized `TableConfig` objects.
         table_config_dicts: list[dict[str, Any]] = []
-        # Mapping from `TableConfig` to index in `table_config_dicts`.
-        table_config_indices: dict[TableConfig, int] = {}
+        # Mapping from `TableConfig` id to index in `table_config_dicts`.
+        table_config_id_to_index: dict[int, int] = {}
 
         def serialize_feature_config(
             feature_config: FeatureConfig,
@@ -1024,17 +1026,17 @@ def serialize_feature_config(
             # key.
             feature_config_dict = feature_config.get_config()
 
-            if feature_config.table not in table_config_indices:
+            if id(feature_config.table) not in table_config_id_to_index:
                 # Save the serialized `TableConfig` the first time we see it and
                 # remember its index.
-                table_config_indices[feature_config.table] = len(
+                table_config_id_to_index[id(feature_config.table)] = len(
                     table_config_dicts
                 )
                 table_config_dicts.append(feature_config_dict["table"])
 
             # Replace the serialized `TableConfig` with its index.
-            feature_config_dict["table"] = table_config_indices[
-                feature_config.table
+            feature_config_dict["table"] = table_config_id_to_index[
+                id(feature_config.table)
             ]
             return feature_config_dict
 

keras_rs/src/layers/embedding/distributed_embedding_config.py

@@ -10,7 +10,7 @@
 
 
 @keras_rs_export("keras_rs.layers.TableConfig")
-@dataclasses.dataclass(eq=True, unsafe_hash=True, order=True)
+@dataclasses.dataclass(order=True)
 class TableConfig:
     """Configuration for one embedding table.
 
@@ -88,7 +88,7 @@ def from_config(cls, config: dict[str, Any]) -> "TableConfig":
 
 
 @keras_rs_export("keras_rs.layers.FeatureConfig")
-@dataclasses.dataclass(eq=True, unsafe_hash=True, order=True)
+@dataclasses.dataclass(order=True)
 class FeatureConfig:
     """Configuration for one embedding feature.
 

keras_rs/src/layers/embedding/jax/distributed_embedding.py

@@ -690,19 +690,22 @@ def _default_device_set_tables(
             raise ValueError("Layer must first be built before setting tables.")
 
         if "default_device" in self._placement_to_path_to_feature_config:
-            table_to_embedding_layer = {}
+            table_name_to_embedding_layer = {}
             for (
                 path,
                 feature_config,
             ) in self._placement_to_path_to_feature_config[
                 "default_device"
             ].items():
-                table_to_embedding_layer[feature_config.table] = (
+                table_name_to_embedding_layer[feature_config.table.name] = (
                     self._default_device_embedding_layers[path]
                 )
 
-            for table, embedding_layer in table_to_embedding_layer.items():
-                table_values = tables.get(table.name, None)
+            for (
+                table_name,
+                embedding_layer,
+            ) in table_name_to_embedding_layer.items():
+                table_values = tables.get(table_name, None)
                 if table_values is not None:
                     if embedding_layer.lora_enabled:
                         raise ValueError("Cannot set table if LoRA is enabled.")

keras_rs/src/layers/embedding/jax/distributed_embedding_test.py

@@ -561,12 +561,18 @@ def loss_fn(y_true, y_pred):
         # Setup a model with a zero initializer but otherwise the same
         # feature configs to test restore. Keep the same embedding layer name to
         # ensure the correct weights are restored.
+        table_config_id_to_table_config_with_zero_init = {
+            id(table_config): dataclasses.replace(
+                table_config, initializer="zeros"
+            )
+            for table_config in table_configs
+        }
         feature_configs_with_zero_init = {
             feature_config.name: dataclasses.replace(
                 feature_config,
-                table=dataclasses.replace(
-                    feature_config.table, initializer="zeros"
-                ),
+                table=table_config_id_to_table_config_with_zero_init[
+                    id(feature_config.table)
+                ],
             )
             for feature_config in feature_configs
         }

keras_rs/src/layers/embedding/tensorflow/config_conversion.py

@@ -53,7 +53,7 @@
 # KerasRS to TensorFlow
 
 
-def translate_keras_rs_configuration(
+def keras_to_tf_tpu_configuration(
     feature_configs: types.Nested[FeatureConfig],
     table_stacking: str | Sequence[str] | Sequence[Sequence[str]],
     num_replicas_in_sync: int,
@@ -66,14 +66,15 @@ def translate_keras_rs_configuration(
     Args:
       feature_configs: The nested Keras RS feature configs.
       table_stacking: The Keras RS table stacking.
+      num_replicas_in_sync: The number of replicas in sync from the strategy.
 
     Returns:
       A tuple containing the TensorFlow TPU feature configs and the TensorFlow
       TPU sparse core embedding config.
     """
-    tables: dict[TableConfig, tf.tpu.experimental.embedding.TableConfig] = {}
+    tables: dict[int, tf.tpu.experimental.embedding.TableConfig] = {}
     feature_configs = keras.tree.map_structure(
-        lambda f: translate_keras_rs_feature_config(
+        lambda f: keras_to_tf_tpu_feature_config(
             f, tables, num_replicas_in_sync
         ),
         feature_configs,
@@ -108,9 +109,9 @@ def translate_keras_rs_configuration(
     return feature_configs, sparse_core_embedding_config
 
 
-def translate_keras_rs_feature_config(
+def keras_to_tf_tpu_feature_config(
     feature_config: FeatureConfig,
-    tables: dict[TableConfig, tf.tpu.experimental.embedding.TableConfig],
+    tables: dict[int, tf.tpu.experimental.embedding.TableConfig],
     num_replicas_in_sync: int,
 ) -> tf.tpu.experimental.embedding.FeatureConfig:
     """Translates a Keras RS feature config to a TensorFlow TPU feature config.
@@ -120,7 +121,8 @@ def translate_keras_rs_feature_config(
 
     Args:
       feature_config: The Keras RS feature config to translate.
-      tables: A mapping of KerasRS table configs to TF TPU table configs.
+      tables: A mapping of KerasRS table config ids to TF TPU table configs.
+      num_replicas_in_sync: The number of replicas in sync from the strategy.
 
     Returns:
       The TensorFlow TPU feature config.
@@ -131,10 +133,10 @@ def translate_keras_rs_feature_config(
             f"but got {num_replicas_in_sync}."
         )
 
-    table = tables.get(feature_config.table, None)
+    table = tables.get(id(feature_config.table), None)
     if table is None:
-        table = translate_keras_rs_table_config(feature_config.table)
-        tables[feature_config.table] = table
+        table = keras_to_tf_tpu_table_config(feature_config.table)
+        tables[id(feature_config.table)] = table
 
     if len(feature_config.output_shape) < 2:
         raise ValueError(
@@ -168,7 +170,7 @@ def translate_keras_rs_feature_config(
     )
 
 
-def translate_keras_rs_table_config(
+def keras_to_tf_tpu_table_config(
     table_config: TableConfig,
 ) -> tf.tpu.experimental.embedding.TableConfig:
     initializer = table_config.initializer
@@ -179,13 +181,13 @@ def translate_keras_rs_table_config(
         vocabulary_size=table_config.vocabulary_size,
         dim=table_config.embedding_dim,
         initializer=initializer,
-        optimizer=translate_optimizer(table_config.optimizer),
+        optimizer=to_tf_tpu_optimizer(table_config.optimizer),
         combiner=table_config.combiner,
         name=table_config.name,
     )
 
 
-def translate_keras_optimizer(
+def keras_to_tf_tpu_optimizer(
     optimizer: keras.optimizers.Optimizer,
 ) -> TfTpuOptimizer:
     """Translates a Keras optimizer to a TensorFlow TPU `_Optimizer`.
@@ -258,7 +260,7 @@ def translate_keras_optimizer(
     return optimizer_mapping.tpu_optimizer_class(**tpu_optimizer_kwargs)
 
 
-def translate_optimizer(
+def to_tf_tpu_optimizer(
     optimizer: str | keras.optimizers.Optimizer | TfTpuOptimizer | None,
 ) -> TfTpuOptimizer:
     """Translates a Keras optimizer into a TensorFlow TPU `_Optimizer`.
@@ -299,7 +301,7 @@ def translate_optimizer(
                 "'sgd', 'adagrad', 'adam', or 'ftrl'"
             )
     elif isinstance(optimizer, keras.optimizers.Optimizer):
-        return translate_keras_optimizer(optimizer)
+        return keras_to_tf_tpu_optimizer(optimizer)
     else:
         raise ValueError(
             f"Unknown optimizer type {type(optimizer)}. Please pass an "
@@ -312,7 +314,7 @@ def translate_optimizer(
 # TensorFlow to TensorFlow
 
 
-def clone_tf_feature_configs(
+def clone_tf_tpu_feature_configs(
     feature_configs: types.Nested[tf.tpu.experimental.embedding.FeatureConfig],
 ) -> types.Nested[tf.tpu.experimental.embedding.FeatureConfig]:
     """Clones and resolves TensorFlow TPU feature configs.
@@ -327,7 +329,7 @@ def clone_tf_feature_configs(
     """
     table_configs_dict = {}
 
-    def clone_and_resolve_tf_feature_config(
+    def clone_and_resolve_tf_tpu_feature_config(
         fc: tf.tpu.experimental.embedding.FeatureConfig,
     ) -> tf.tpu.experimental.embedding.FeatureConfig:
         if fc.table not in table_configs_dict:
@@ -336,7 +338,7 @@ def clone_and_resolve_tf_feature_config(
                     vocabulary_size=fc.table.vocabulary_size,
                     dim=fc.table.dim,
                     initializer=fc.table.initializer,
-                    optimizer=translate_optimizer(fc.table.optimizer),
+                    optimizer=to_tf_tpu_optimizer(fc.table.optimizer),
                     combiner=fc.table.combiner,
                     name=fc.table.name,
                     quantization_config=fc.table.quantization_config,
@@ -352,5 +354,5 @@ def clone_and_resolve_tf_feature_config(
         )
 
     return keras.tree.map_structure(
-        clone_and_resolve_tf_feature_config, feature_configs
+        clone_and_resolve_tf_tpu_feature_config, feature_configs
     )