add `classifier_dropout` to classification heads (#12794)

* add classifier_dropout to Electra

* no type annotations yet
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* add classifier_dropout to Electra

* add classifier_dropout to Electra ForTokenClass.

* add classifier_dropout to bert

* add classifier_dropout to roberta

* add classifier_dropout to big_bird

* add classifier_dropout to mobilebert

* empty commit to trigger CI

* add classifier_dropout to reformer

* add classifier_dropout to ConvBERT

* add classifier_dropout to Albert

* add classifier_dropout to Albert
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

src/transformers/models/albert/modeling_albert.py

@@ -1088,7 +1088,12 @@ class AlbertForTokenClassification(AlbertPreTrainedModel):
         self.num_labels = config.num_labels
 
         self.albert = AlbertModel(config, add_pooling_layer=False)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        classifier_dropout_prob = (
+            config.classifier_dropout_prob
+            if config.classifier_dropout_prob is not None
+            else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout_prob)
         self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
 
         self.init_weights()

src/transformers/models/albert/modeling_tf_albert.py

@@ -1199,7 +1199,12 @@ class TFAlbertForTokenClassification(TFAlbertPreTrainedModel, TFTokenClassificat
         self.num_labels = config.num_labels
 
         self.albert = TFAlbertMainLayer(config, add_pooling_layer=False, name="albert")
-        self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
+        classifier_dropout_prob = (
+            config.classifier_dropout_prob
+            if config.classifier_dropout_prob is not None
+            else config.hidden_dropout_prob
+        )
+        self.dropout = tf.keras.layers.Dropout(rate=classifier_dropout_prob)
         self.classifier = tf.keras.layers.Dense(
             units=config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
         )

src/transformers/models/bert/configuration_bert.py

@@ -104,6 +104,8 @@ class BertConfig(PretrainedConfig):
         use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
             Whether or not the model should return the last key/values attentions (not used by all models). Only
             relevant if ``config.is_decoder=True``.
+        classifier_dropout (:obj:`float`, `optional`):
+            The dropout ratio for the classification head.
 
     Examples::
 
@@ -138,6 +140,7 @@ class BertConfig(PretrainedConfig):
         gradient_checkpointing=False,
         position_embedding_type="absolute",
         use_cache=True,
+        classifier_dropout=None,
         **kwargs
     ):
         super().__init__(pad_token_id=pad_token_id, **kwargs)
@@ -157,6 +160,7 @@ class BertConfig(PretrainedConfig):
         self.gradient_checkpointing = gradient_checkpointing
         self.position_embedding_type = position_embedding_type
         self.use_cache = use_cache
+        self.classifier_dropout = classifier_dropout
 
 
 class BertOnnxConfig(OnnxConfig):

src/transformers/models/bert/modeling_bert.py

@@ -1486,7 +1486,10 @@ class BertForSequenceClassification(BertPreTrainedModel):
         self.config = config
 
         self.bert = BertModel(config)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
         self.classifier = nn.Linear(config.hidden_size, config.num_labels)
 
         self.init_weights()
@@ -1677,7 +1680,10 @@ class BertForTokenClassification(BertPreTrainedModel):
         self.num_labels = config.num_labels
 
         self.bert = BertModel(config, add_pooling_layer=False)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
         self.classifier = nn.Linear(config.hidden_size, config.num_labels)
 
         self.init_weights()

src/transformers/models/bert/modeling_flax_bert.py

@@ -915,7 +915,12 @@ class FlaxBertForSequenceClassificationModule(nn.Module):
 
     def setup(self):
         self.bert = FlaxBertModule(config=self.config, dtype=self.dtype)
-        self.dropout = nn.Dropout(rate=self.config.hidden_dropout_prob)
+        classifier_dropout = (
+            self.config.classifier_dropout
+            if self.config.classifier_dropout is not None
+            else self.config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(rate=classifier_dropout)
         self.classifier = nn.Dense(
             self.config.num_labels,
             dtype=self.dtype,
@@ -1057,7 +1062,12 @@ class FlaxBertForTokenClassificationModule(nn.Module):
 
     def setup(self):
         self.bert = FlaxBertModule(config=self.config, dtype=self.dtype, add_pooling_layer=False)
-        self.dropout = nn.Dropout(rate=self.config.hidden_dropout_prob)
+        classifier_dropout = (
+            self.config.classifier_dropout
+            if self.config.classifier_dropout is not None
+            else self.config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(rate=classifier_dropout)
         self.classifier = nn.Dense(self.config.num_labels, dtype=self.dtype)
 
     def __call__(

src/transformers/models/bert/modeling_tf_bert.py

@@ -1386,7 +1386,10 @@ class TFBertForSequenceClassification(TFBertPreTrainedModel, TFSequenceClassific
         self.num_labels = config.num_labels
 
         self.bert = TFBertMainLayer(config, name="bert")
-        self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = tf.keras.layers.Dropout(rate=classifier_dropout)
         self.classifier = tf.keras.layers.Dense(
             units=config.num_labels,
             kernel_initializer=get_initializer(config.initializer_range),
@@ -1652,7 +1655,10 @@ class TFBertForTokenClassification(TFBertPreTrainedModel, TFTokenClassificationL
         self.num_labels = config.num_labels
 
         self.bert = TFBertMainLayer(config, add_pooling_layer=False, name="bert")
-        self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = tf.keras.layers.Dropout(rate=classifier_dropout)
         self.classifier = tf.keras.layers.Dense(
             units=config.num_labels,
             kernel_initializer=get_initializer(config.initializer_range),

src/transformers/models/big_bird/configuration_big_bird.py

@@ -84,6 +84,8 @@ class BigBirdConfig(PretrainedConfig):
             "block_sparse"`.
         gradient_checkpointing (:obj:`bool`, `optional`, defaults to :obj:`False`):
             If True, use gradient checkpointing to save memory at the expense of slower backward pass.
+        classifier_dropout (:obj:`float`, `optional`):
+            The dropout ratio for the classification head.
 
         Example::
 
@@ -126,6 +128,7 @@ class BigBirdConfig(PretrainedConfig):
         block_size=64,
         num_random_blocks=3,
         gradient_checkpointing=False,
+        classifier_dropout=None,
         **kwargs
     ):
         super().__init__(
@@ -157,3 +160,4 @@ class BigBirdConfig(PretrainedConfig):
         self.use_bias = use_bias
         self.block_size = block_size
         self.num_random_blocks = num_random_blocks
+        self.classifier_dropout = classifier_dropout

src/transformers/models/big_bird/modeling_big_bird.py

@@ -2605,7 +2605,10 @@ class BigBirdClassificationHead(nn.Module):
     def __init__(self, config):
         super().__init__()
         self.dense = nn.Linear(config.hidden_size, config.hidden_size)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
         self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
 
         self.config = config
@@ -2821,7 +2824,10 @@ class BigBirdForTokenClassification(BigBirdPreTrainedModel):
         self.num_labels = config.num_labels
 
         self.bert = BigBirdModel(config)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
         self.classifier = nn.Linear(config.hidden_size, config.num_labels)
 
         self.init_weights()

src/transformers/models/big_bird/modeling_flax_big_bird.py

@@ -1654,7 +1654,12 @@ class FlaxBigBirdClassificationHead(nn.Module):
 
     def setup(self):
         self.dense = nn.Dense(self.config.hidden_size, dtype=self.dtype)
-        self.dropout = nn.Dropout(self.config.hidden_dropout_prob)
+        classifier_dropout = (
+            self.config.classifier_dropout
+            if self.config.classifier_dropout is not None
+            else self.config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
         self.out_proj = nn.Dense(self.config.num_labels, dtype=self.dtype)
 
     def __call__(self, features, deterministic=True):
@@ -1831,7 +1836,12 @@ class FlaxBigBirdForTokenClassificationModule(nn.Module):
 
     def setup(self):
         self.bert = FlaxBigBirdModule(config=self.config, dtype=self.dtype, add_pooling_layer=False)
-        self.dropout = nn.Dropout(rate=self.config.hidden_dropout_prob)
+        classifier_dropout = (
+            self.config.classifier_dropout
+            if self.config.classifier_dropout is not None
+            else self.config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(rate=classifier_dropout)
         self.classifier = nn.Dense(self.config.num_labels, dtype=self.dtype)
 
     def __call__(

src/transformers/models/convbert/configuration_convbert.py

@@ -73,7 +73,8 @@ class ConvBertConfig(PretrainedConfig):
             The number of groups for grouped linear layers for ConvBert model
         conv_kernel_size (:obj:`int`, `optional`, defaults to 9):
             The size of the convolutional kernel.
-
+        classifier_dropout (:obj:`float`, `optional`):
+            The dropout ratio for the classification head.
 
     Example::
         >>> from transformers import ConvBertModel, ConvBertConfig
@@ -108,6 +109,7 @@ class ConvBertConfig(PretrainedConfig):
         head_ratio=2,
         conv_kernel_size=9,
         num_groups=1,
+        classifier_dropout=None,
         **kwargs,
     ):
         super().__init__(
@@ -134,3 +136,4 @@ class ConvBertConfig(PretrainedConfig):
         self.head_ratio = head_ratio
         self.conv_kernel_size = conv_kernel_size
         self.num_groups = num_groups
+        self.classifier_dropout = classifier_dropout

src/transformers/models/convbert/modeling_convbert.py

@@ -936,7 +936,10 @@ class ConvBertClassificationHead(nn.Module):
     def __init__(self, config):
         super().__init__()
         self.dense = nn.Linear(config.hidden_size, config.hidden_size)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
         self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
 
         self.config = config
@@ -1152,7 +1155,10 @@ class ConvBertForTokenClassification(ConvBertPreTrainedModel):
         self.num_labels = config.num_labels
 
         self.convbert = ConvBertModel(config)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
         self.classifier = nn.Linear(config.hidden_size, config.num_labels)
 
         self.init_weights()

src/transformers/models/convbert/modeling_tf_convbert.py

@@ -970,7 +970,10 @@ class TFConvBertClassificationHead(tf.keras.layers.Layer):
         self.dense = tf.keras.layers.Dense(
             config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
         )
-        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = tf.keras.layers.Dropout(classifier_dropout)
         self.out_proj = tf.keras.layers.Dense(
             config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="out_proj"
         )
@@ -1240,7 +1243,10 @@ class TFConvBertForTokenClassification(TFConvBertPreTrainedModel, TFTokenClassif
 
         self.num_labels = config.num_labels
         self.convbert = TFConvBertMainLayer(config, name="convbert")
-        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = tf.keras.layers.Dropout(classifier_dropout)
         self.classifier = tf.keras.layers.Dense(
             config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
         )

src/transformers/models/electra/configuration_electra.py

@@ -104,6 +104,8 @@ class ElectraConfig(PretrainedConfig):
             <https://arxiv.org/abs/1803.02155>`__. For more information on :obj:`"relative_key_query"`, please refer to
             `Method 4` in `Improve Transformer Models with Better Relative Position Embeddings (Huang et al.)
             <https://arxiv.org/abs/2009.13658>`__.
+        classifier_dropout (:obj:`float`, `optional`):
+            The dropout ratio for the classification head.
 
     Examples::
 
@@ -141,6 +143,7 @@ class ElectraConfig(PretrainedConfig):
         summary_last_dropout=0.1,
         pad_token_id=0,
         position_embedding_type="absolute",
+        classifier_dropout=None,
         **kwargs
     ):
         super().__init__(pad_token_id=pad_token_id, **kwargs)
@@ -164,3 +167,4 @@ class ElectraConfig(PretrainedConfig):
         self.summary_activation = summary_activation
         self.summary_last_dropout = summary_last_dropout
         self.position_embedding_type = position_embedding_type
+        self.classifier_dropout = classifier_dropout

src/transformers/models/electra/modeling_electra.py

@@ -900,7 +900,10 @@ class ElectraClassificationHead(nn.Module):
     def __init__(self, config):
         super().__init__()
         self.dense = nn.Linear(config.hidden_size, config.hidden_size)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
         self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
 
     def forward(self, features, **kwargs):
@@ -1200,7 +1203,10 @@ class ElectraForTokenClassification(ElectraPreTrainedModel):
         super().__init__(config)
 
         self.electra = ElectraModel(config)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
         self.classifier = nn.Linear(config.hidden_size, config.num_labels)
         self.init_weights()
 

src/transformers/models/electra/modeling_flax_electra.py

@@ -783,7 +783,12 @@ class FlaxElectraForTokenClassificationModule(nn.Module):
 
     def setup(self):
         self.electra = FlaxElectraModule(config=self.config, dtype=self.dtype)
-        self.dropout = nn.Dropout(self.config.hidden_dropout_prob)
+        classifier_dropout = (
+            self.config.classifier_dropout
+            if self.config.classifier_dropout is not None
+            else self.config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
         self.classifier = nn.Dense(self.config.num_labels)
 
     def __call__(
@@ -1069,7 +1074,12 @@ class FlaxElectraClassificationHead(nn.Module):
 
     def setup(self):
         self.dense = nn.Dense(self.config.hidden_size, dtype=self.dtype)
-        self.dropout = nn.Dropout(self.config.hidden_dropout_prob)
+        classifier_dropout = (
+            self.config.classifier_dropout
+            if self.config.classifier_dropout is not None
+            else self.config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
         self.out_proj = nn.Dense(self.config.num_labels, dtype=self.dtype)
 
     def __call__(self, hidden_states, deterministic: bool = True):

src/transformers/models/electra/modeling_tf_electra.py

@@ -1039,7 +1039,12 @@ class TFElectraClassificationHead(tf.keras.layers.Layer):
         self.dense = tf.keras.layers.Dense(
             config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
         )
-        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+        classifier_dropout = (
+            config.classifhidden_dropout_probier_dropout
+            if config.classifier_dropout is not None
+            else config.hidden_dropout_prob
+        )
+        self.dropout = tf.keras.layers.Dropout(classifier_dropout)
         self.out_proj = tf.keras.layers.Dense(
             config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="out_proj"
         )
@@ -1309,7 +1314,10 @@ class TFElectraForTokenClassification(TFElectraPreTrainedModel, TFTokenClassific
         super().__init__(config, **kwargs)
 
         self.electra = TFElectraMainLayer(config, name="electra")
-        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = tf.keras.layers.Dropout(classifier_dropout)
         self.classifier = tf.keras.layers.Dense(
             config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
         )

src/transformers/models/mobilebert/configuration_mobilebert.py

@@ -84,6 +84,8 @@ class MobileBertConfig(PretrainedConfig):
             Number of FFNs in a block.
         normalization_type (:obj:`str`, `optional`, defaults to :obj:`"no_norm"`):
             The normalization type in MobileBERT.
+        classifier_dropout (:obj:`float`, `optional`):
+            The dropout ratio for the classification head.
 
     Examples::
 
@@ -128,6 +130,7 @@ class MobileBertConfig(PretrainedConfig):
         num_feedforward_networks=4,
         normalization_type="no_norm",
         classifier_activation=True,
+        classifier_dropout=None,
         **kwargs
     ):
         super().__init__(pad_token_id=pad_token_id, **kwargs)
@@ -158,3 +161,5 @@ class MobileBertConfig(PretrainedConfig):
             self.true_hidden_size = intra_bottleneck_size
         else:
             self.true_hidden_size = hidden_size
+
+        self.classifier_dropout = classifier_dropout

src/transformers/models/mobilebert/modeling_mobilebert.py

@@ -1212,7 +1212,10 @@ class MobileBertForSequenceClassification(MobileBertPreTrainedModel):
         self.config = config
 
         self.mobilebert = MobileBertModel(config)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
         self.classifier = nn.Linear(config.hidden_size, config.num_labels)
 
         self.init_weights()
@@ -1510,7 +1513,10 @@ class MobileBertForTokenClassification(MobileBertPreTrainedModel):
         self.num_labels = config.num_labels
 
         self.mobilebert = MobileBertModel(config, add_pooling_layer=False)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
         self.classifier = nn.Linear(config.hidden_size, config.num_labels)
 
         self.init_weights()

src/transformers/models/mobilebert/modeling_tf_mobilebert.py

@@ -1339,7 +1339,10 @@ class TFMobileBertForSequenceClassification(TFMobileBertPreTrainedModel, TFSeque
         self.num_labels = config.num_labels
 
         self.mobilebert = TFMobileBertMainLayer(config, name="mobilebert")
-        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = tf.keras.layers.Dropout(classifier_dropout)
         self.classifier = tf.keras.layers.Dense(
             config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
         )
@@ -1730,7 +1733,10 @@ class TFMobileBertForTokenClassification(TFMobileBertPreTrainedModel, TFTokenCla
         self.num_labels = config.num_labels
 
         self.mobilebert = TFMobileBertMainLayer(config, add_pooling_layer=False, name="mobilebert")
-        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = tf.keras.layers.Dropout(classifier_dropout)
         self.classifier = tf.keras.layers.Dense(
             config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
         )

src/transformers/models/reformer/configuration_reformer.py

@@ -140,6 +140,8 @@ class ReformerConfig(PretrainedConfig):
             Whether to tie input and output embeddings.
         use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
             Whether or not the model should return the last key/values attentions (not used by all models).
+        classifier_dropout (:obj:`float`, `optional`):
+            The dropout ratio for the classification head.
 
     Examples::
 
@@ -191,6 +193,7 @@ class ReformerConfig(PretrainedConfig):
         vocab_size=320,
         tie_word_embeddings=False,
         use_cache=True,
+        classifier_dropout=None,
         **kwargs
     ):
         super().__init__(
@@ -230,3 +233,4 @@ class ReformerConfig(PretrainedConfig):
         self.chunk_size_lm_head = chunk_size_lm_head
         self.attn_layers = attn_layers
         self.use_cache = use_cache
+        self.classifier_dropout = classifier_dropout