Revert "Attempt five!"

This reverts commit 3302207.

Author: Rocketknight1

src/transformers/models/albert/modeling_tf_albert.py

@@ -146,7 +146,7 @@ def __init__(self, config: AlbertConfig, **kwargs):
         self.LayerNorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
         self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
 
-    def build(self, input_shape=None):
+    def build(self, input_shape: tf.TensorShape):
         with tf.name_scope("word_embeddings"):
             self.weight = self.add_weight(
                 name="weight",
@@ -168,12 +168,7 @@ def build(self, input_shape=None):
                 initializer=get_initializer(self.initializer_range),
             )
 
-        if self.built:
-            return
-        self.built = True
-        if getattr(self, "LayerNorm", None) is not None:
-            with tf.name_scope(self.LayerNorm.name):
-                self.LayerNorm.build([None, None, self.config.embedding_size])
+        super().build(input_shape)
 
     # Copied from transformers.models.bert.modeling_tf_bert.TFBertEmbeddings.call
     def call(
@@ -251,7 +246,6 @@ def __init__(self, config: AlbertConfig, **kwargs):
         # Two different dropout probabilities; see https://github.com/google-research/albert/blob/master/modeling.py#L971-L993
         self.attention_dropout = tf.keras.layers.Dropout(rate=config.attention_probs_dropout_prob)
         self.output_dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
-        self.config = config
 
     def transpose_for_scores(self, tensor: tf.Tensor, batch_size: int) -> tf.Tensor:
         # Reshape from [batch_size, seq_length, all_head_size] to [batch_size, seq_length, num_attention_heads, attention_head_size]
@@ -313,26 +307,6 @@ def call(
 
         return outputs
 
-    def build(self, input_shape=None):
-        if self.built:
-            return
-        self.built = True
-        if getattr(self, "query", None) is not None:
-            with tf.name_scope(self.query.name):
-                self.query.build(self.config.hidden_size)
-        if getattr(self, "key", None) is not None:
-            with tf.name_scope(self.key.name):
-                self.key.build(self.config.hidden_size)
-        if getattr(self, "value", None) is not None:
-            with tf.name_scope(self.value.name):
-                self.value.build(self.config.hidden_size)
-        if getattr(self, "dense", None) is not None:
-            with tf.name_scope(self.dense.name):
-                self.dense.build(self.config.hidden_size)
-        if getattr(self, "LayerNorm", None) is not None:
-            with tf.name_scope(self.LayerNorm.name):
-                self.LayerNorm.build([None, None, self.config.hidden_size])
-
 
 class TFAlbertLayer(tf.keras.layers.Layer):
     def __init__(self, config: AlbertConfig, **kwargs):
@@ -355,7 +329,6 @@ def __init__(self, config: AlbertConfig, **kwargs):
             epsilon=config.layer_norm_eps, name="full_layer_layer_norm"
         )
         self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
-        self.config = config
 
     def call(
         self,
@@ -383,23 +356,6 @@ def call(
 
         return outputs
 
-    def build(self, input_shape=None):
-        if self.built:
-            return
-        self.built = True
-        if getattr(self, "attention", None) is not None:
-            with tf.name_scope(self.attention.name):
-                self.attention.build(None)
-        if getattr(self, "ffn", None) is not None:
-            with tf.name_scope(self.ffn.name):
-                self.ffn.build(self.config.hidden_size)
-        if getattr(self, "ffn_output", None) is not None:
-            with tf.name_scope(self.ffn_output.name):
-                self.ffn_output.build(self.config.intermediate_size)
-        if getattr(self, "full_layer_layer_norm", None) is not None:
-            with tf.name_scope(self.full_layer_layer_norm.name):
-                self.full_layer_layer_norm.build([None, None, self.config.hidden_size])
-
 
 class TFAlbertLayerGroup(tf.keras.layers.Layer):
     def __init__(self, config: AlbertConfig, **kwargs):
@@ -443,15 +399,6 @@ def call(
 
         return tuple(v for v in [hidden_states, layer_hidden_states, layer_attentions] if v is not None)
 
-    def build(self, input_shape=None):
-        if self.built:
-            return
-        self.built = True
-        if getattr(self, "albert_layers", None) is not None:
-            for layer in self.albert_layers:
-                with tf.name_scope(layer.name):
-                    layer.build(None)
-
 
 class TFAlbertTransformer(tf.keras.layers.Layer):
     def __init__(self, config: AlbertConfig, **kwargs):
@@ -469,7 +416,6 @@ def __init__(self, config: AlbertConfig, **kwargs):
         self.albert_layer_groups = [
             TFAlbertLayerGroup(config, name=f"albert_layer_groups_._{i}") for i in range(config.num_hidden_groups)
         ]
-        self.config = config
 
     def call(
         self,
@@ -511,18 +457,6 @@ def call(
             last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
         )
 
-    def build(self, input_shape=None):
-        if self.built:
-            return
-        self.built = True
-        if getattr(self, "embedding_hidden_mapping_in", None) is not None:
-            with tf.name_scope(self.embedding_hidden_mapping_in.name):
-                self.embedding_hidden_mapping_in.build(self.config.embedding_size)
-        if getattr(self, "albert_layer_groups", None) is not None:
-            for layer in self.albert_layer_groups:
-                with tf.name_scope(layer.name):
-                    layer.build(None)
-
 
 class TFAlbertPreTrainedModel(TFPreTrainedModel):
     """
@@ -554,21 +488,13 @@ def __init__(self, config: AlbertConfig, input_embeddings: tf.keras.layers.Layer
         # an output-only bias for each token.
         self.decoder = input_embeddings
 
-    def build(self, input_shape=None):
+    def build(self, input_shape: tf.TensorShape):
         self.bias = self.add_weight(shape=(self.config.vocab_size,), initializer="zeros", trainable=True, name="bias")
         self.decoder_bias = self.add_weight(
             shape=(self.config.vocab_size,), initializer="zeros", trainable=True, name="decoder/bias"
         )
 
-        if self.built:
-            return
-        self.built = True
-        if getattr(self, "dense", None) is not None:
-            with tf.name_scope(self.dense.name):
-                self.dense.build(self.config.hidden_size)
-        if getattr(self, "LayerNorm", None) is not None:
-            with tf.name_scope(self.LayerNorm.name):
-                self.LayerNorm.build([None, None, self.config.embedding_size])
+        super().build(input_shape)
 
     def get_output_embeddings(self) -> tf.keras.layers.Layer:
         return self.decoder
@@ -724,20 +650,6 @@ def call(
             attentions=encoder_outputs.attentions,
         )
 
-    def build(self, input_shape=None):
-        if self.built:
-            return
-        self.built = True
-        if getattr(self, "embeddings", None) is not None:
-            with tf.name_scope(self.embeddings.name):
-                self.embeddings.build(None)
-        if getattr(self, "encoder", None) is not None:
-            with tf.name_scope(self.encoder.name):
-                self.encoder.build(None)
-        if getattr(self, "pooler", None) is not None:
-            with tf.name_scope(self.pooler.name):
-                self.pooler.build(None)  # TODO Matt might be wrong
-
 
 @dataclass
 class TFAlbertForPreTrainingOutput(ModelOutput):
@@ -913,14 +825,6 @@ def call(
 
         return outputs
 
-    def build(self, input_shape=None):
-        if self.built:
-            return
-        self.built = True
-        if getattr(self, "albert", None) is not None:
-            with tf.name_scope(self.albert.name):
-                self.albert.build(None)
-
 
 @add_start_docstrings(
     """
@@ -1017,20 +921,6 @@ def call(
             attentions=outputs.attentions,
         )
 
-    def build(self, input_shape=None):
-        if self.built:
-            return
-        self.built = True
-        if getattr(self, "albert", None) is not None:
-            with tf.name_scope(self.albert.name):
-                self.albert.build(None)
-        if getattr(self, "predictions", None) is not None:
-            with tf.name_scope(self.predictions.name):
-                self.predictions.build(None)
-        if getattr(self, "sop_classifier", None) is not None:
-            with tf.name_scope(self.sop_classifier.name):
-                self.sop_classifier.build(None)
-
 
 class TFAlbertSOPHead(tf.keras.layers.Layer):
     def __init__(self, config: AlbertConfig, **kwargs):
@@ -1042,22 +932,13 @@ def __init__(self, config: AlbertConfig, **kwargs):
             kernel_initializer=get_initializer(config.initializer_range),
             name="classifier",
         )
-        self.config = config
 
     def call(self, pooled_output: tf.Tensor, training: bool) -> tf.Tensor:
         dropout_pooled_output = self.dropout(inputs=pooled_output, training=training)
         logits = self.classifier(inputs=dropout_pooled_output)
 
         return logits
 
-    def build(self, input_shape=None):
-        if self.built:
-            return
-        self.built = True
-        if getattr(self, "classifier", None) is not None:
-            with tf.name_scope(self.classifier.name):
-                self.classifier.build(self.config.hidden_size)
-
 
 @add_start_docstrings("""Albert Model with a `language modeling` head on top.""", ALBERT_START_DOCSTRING)
 class TFAlbertForMaskedLM(TFAlbertPreTrainedModel, TFMaskedLanguageModelingLoss):
@@ -1154,17 +1035,6 @@ def call(
             attentions=outputs.attentions,
         )
 
-    def build(self, input_shape=None):
-        if self.built:
-            return
-        self.built = True
-        if getattr(self, "albert", None) is not None:
-            with tf.name_scope(self.albert.name):
-                self.albert.build(None)
-        if getattr(self, "predictions", None) is not None:
-            with tf.name_scope(self.predictions.name):
-                self.predictions.build(None)
-
 
 @add_start_docstrings(
     """
@@ -1188,7 +1058,6 @@ def __init__(self, config: AlbertConfig, *inputs, **kwargs):
         self.classifier = tf.keras.layers.Dense(
             units=config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
         )
-        self.config = config
 
     @unpack_inputs
     @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@@ -1248,17 +1117,6 @@ def call(
             attentions=outputs.attentions,
         )
 
-    def build(self, input_shape=None):
-        if self.built:
-            return
-        self.built = True
-        if getattr(self, "albert", None) is not None:
-            with tf.name_scope(self.albert.name):
-                self.albert.build(None)
-        if getattr(self, "classifier", None) is not None:
-            with tf.name_scope(self.classifier.name):
-                self.classifier.build(self.config.hidden_size)
-
 
 @add_start_docstrings(
     """
@@ -1287,7 +1145,6 @@ def __init__(self, config: AlbertConfig, *inputs, **kwargs):
         self.classifier = tf.keras.layers.Dense(
             units=config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
         )
-        self.config = config
 
     @unpack_inputs
     @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@@ -1343,17 +1200,6 @@ def call(
             attentions=outputs.attentions,
         )
 
-    def build(self, input_shape=None):
-        if self.built:
-            return
-        self.built = True
-        if getattr(self, "albert", None) is not None:
-            with tf.name_scope(self.albert.name):
-                self.albert.build(None)
-        if getattr(self, "classifier", None) is not None:
-            with tf.name_scope(self.classifier.name):
-                self.classifier.build(self.config.hidden_size)
-
 
 @add_start_docstrings(
     """
@@ -1375,7 +1221,6 @@ def __init__(self, config: AlbertConfig, *inputs, **kwargs):
         self.qa_outputs = tf.keras.layers.Dense(
             units=config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="qa_outputs"
         )
-        self.config = config
 
     @unpack_inputs
     @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@@ -1450,17 +1295,6 @@ def call(
             attentions=outputs.attentions,
         )
 
-    def build(self, input_shape=None):
-        if self.built:
-            return
-        self.built = True
-        if getattr(self, "albert", None) is not None:
-            with tf.name_scope(self.albert.name):
-                self.albert.build(None)
-        if getattr(self, "qa_outputs", None) is not None:
-            with tf.name_scope(self.qa_outputs.name):
-                self.qa_outputs.build(self.config.hidden_size)
-
 
 @add_start_docstrings(
     """
@@ -1482,7 +1316,6 @@ def __init__(self, config: AlbertConfig, *inputs, **kwargs):
         self.classifier = tf.keras.layers.Dense(
             units=1, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
         )
-        self.config = config
 
     @unpack_inputs
     @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
@@ -1561,14 +1394,3 @@ def call(
             hidden_states=outputs.hidden_states,
             attentions=outputs.attentions,
         )
-
-    def build(self, input_shape=None):
-        if self.built:
-            return
-        self.built = True
-        if getattr(self, "albert", None) is not None:
-            with tf.name_scope(self.albert.name):
-                self.albert.build(None)
-        if getattr(self, "classifier", None) is not None:
-            with tf.name_scope(self.classifier.name):
-                self.classifier.build(self.config.hidden_size)