Add a convenience method for building in your own name scope

src/transformers/modeling_tf_utils.py

@@ -1122,6 +1122,10 @@ def dummy_inputs(self) -> Dict[str, tf.Tensor]:
                     )
         return dummies
 
+    def build_in_name_scope(self):
+        with tf.name_scope(self.name):
+            self.build(input_shape=None)
+
     @property
     def framework(self) -> str:
         """
@@ -1869,7 +1873,7 @@ def set_input_embeddings(self, value):
             main_layer.set_input_embeddings(value)
         except AttributeError:
             logger.info("Building the model")
-            self.build()
+            self.build_in_name_scope()
             main_layer.set_input_embeddings(value)
 
     def get_output_embeddings(self) -> Union[None, tf.keras.layers.Layer]:
@@ -1886,7 +1890,7 @@ def get_output_embeddings(self) -> Union[None, tf.keras.layers.Layer]:
                 return lm_head.get_output_embeddings()
             except AttributeError:
                 logger.info("Building the model")
-                self.build()
+                self.build_in_name_scope()
 
                 return lm_head().get_output_embeddings()
 
@@ -1906,7 +1910,7 @@ def set_output_embeddings(self, value):
                 lm_head.set_output_embeddings(value)
             except AttributeError:
                 logger.info("Building the model")
-                self.build()
+                self.build_in_name_scope()
                 lm_head.set_output_embeddings(value)
 
     def get_output_layer_with_bias(self) -> Union[None, tf.keras.layers.Layer]:
@@ -1944,7 +1948,7 @@ def get_bias(self) -> Union[None, Dict[str, tf.Variable]]:
             try:
                 return lm_head.get_bias()
             except AttributeError:
-                self.build()
+                self.build_in_name_scope()
 
                 return lm_head.get_bias()
         return None
@@ -1962,7 +1966,7 @@ def set_bias(self, value):
             try:
                 lm_head.set_bias(value)
             except AttributeError:
-                self.build()
+                self.build_in_name_scope()
                 lm_head.set_bias(value)
 
     def get_lm_head(self) -> tf.keras.layers.Layer:
@@ -2049,7 +2053,7 @@ def _get_word_embedding_weight(model, embedding_layer):
         # The reason why the attributes don't exist might be
         # because the model is not built, so retry getting
         # the argument after building the model
-        model.build()
+        model.build_in_name_scope()
 
         embeds = getattr(embedding_layer, "weight", None)
         if embeds is not None:
@@ -2908,9 +2912,9 @@ def from_pretrained(
         # we might need to extend the variable scope for composite models
         if load_weight_prefix is not None:
             with tf.compat.v1.variable_scope(load_weight_prefix):
-                model.build()  # build the network with dummy inputs
+                model.build_in_name_scope()  # build the network with dummy inputs
         else:
-            model.build()  # build the network with dummy inputs
+            model.build_in_name_scope()  # build the network with dummy inputs
 
         if safetensors_from_pt:
             from .modeling_tf_pytorch_utils import load_pytorch_state_dict_in_tf2_model