Simplify the implementation

Author: james77777778

keras_nlp/api/models/__init__.py

@@ -106,7 +106,6 @@
 )
 from keras_nlp.src.models.falcon.falcon_preprocessor import FalconPreprocessor
 from keras_nlp.src.models.falcon.falcon_tokenizer import FalconTokenizer
-from keras_nlp.src.models.feature_pyramid_backbone import FeaturePyramidBackbone
 from keras_nlp.src.models.gemma.gemma_backbone import GemmaBackbone
 from keras_nlp.src.models.gemma.gemma_causal_lm import GemmaCausalLM
 from keras_nlp.src.models.gemma.gemma_causal_lm_preprocessor import (

keras_nlp/src/models/backbone.py

@@ -108,6 +108,39 @@ def token_embedding(self):
     def token_embedding(self, value):
         self._token_embedding = value
 
+    @property
+    def pyramid_outputs(self):
+        """A dict for feature pyramid outputs.
+
+        The key is a string represents the name of the feature output and the
+        value is a `keras.KerasTensor`. A typical feature pyramid has multiple
+        levels corresponding to scales such as `["P2", "P3", "P4", "P5"]`. Scale
+        `Pn` represents a feature map `2^n` times smaller in width and height
+        than the inputs.
+        """
+        return getattr(self, "_pyramid_outputs", {})
+
+    @pyramid_outputs.setter
+    def pyramid_outputs(self, value):
+        if not isinstance(value, dict):
+            raise TypeError(
+                "`pyramid_outputs` must be a dictionary. "
+                f"Received: value={value} of type {type(value)}"
+            )
+        for k, v in value.items():
+            if not isinstance(k, str):
+                raise TypeError(
+                    "The key of `pyramid_outputs` must be a string. "
+                    f"Received: key={k} of type {type(k)}"
+                )
+            if not isinstance(v, keras.KerasTensor):
+                raise TypeError(
+                    "The value of `pyramid_outputs` must be a "
+                    "`keras.KerasTensor`. "
+                    f"Received: value={v} of type {type(v)}"
+                )
+        self._pyramid_outputs = value
+
     def quantize(self, mode, **kwargs):
         assert_quantization_support()
         return super().quantize(mode, **kwargs)

keras_nlp/src/models/feature_pyramid_backbone.py

@@ -15,13 +15,12 @@
 from keras import layers
 
 from keras_nlp.src.api_export import keras_nlp_export
-from keras_nlp.src.models.feature_pyramid_backbone import FeaturePyramidBackbone
-from keras_nlp.src.utils.keras_utils import get_tensor_name
+from keras_nlp.src.models.backbone import Backbone
 from keras_nlp.src.utils.keras_utils import standardize_data_format
 
 
 @keras_nlp_export("keras_nlp.models.ResNetBackbone")
-class ResNetBackbone(FeaturePyramidBackbone):
+class ResNetBackbone(Backbone):
     """ResNet and ResNetV2 core network with hyperparameters.
 
     This class implements a ResNet backbone as described in [Deep Residual
@@ -181,7 +180,7 @@ def __init__(
                 dtype=dtype,
                 name=f"{version}_stack{stack_index}",
             )
-            pyramid_outputs[f"P{stack_index + 2}"] = get_tensor_name(x)
+            pyramid_outputs[f"P{stack_index + 2}"] = x
 
         if use_pre_activation:
             x = layers.BatchNormalization(

keras_nlp/src/models/resnet/resnet_backbone.py

@@ -128,13 +128,13 @@ def __init__(
         outputs = {}
         for k in output_keys:
             try:
-                name = self.pyramid_outputs[k]
+                output = self.pyramid_outputs[k]
             except KeyError:
                 raise KeyError(
                     f"'{k}' not in self.pyramid_outputs. The available keys "
                     f"are: {list(self.pyramid_outputs.keys())}"
                 )
-            outputs[k] = self.get_layer(name).output
+            outputs[k] = output
 
         super(ResNetBackbone, self).__init__(
             inputs=self.inputs, outputs=outputs, dtype=dtype, **kwargs