[Wav2Vec2] Rename model's feature extractor to feature encoder (#14959)

* rename classes

* clean up more namings

* remove bogus file

* Apply suggestions from code review

* Apply suggestions from code review

* replace more names

* more regex replace

* make style

* correct

* correct more

* make style

* finish

* correct more in wav2vec2

* make style

* improve freeze_extractor

* add aliases

* add tf aliases

examples/pytorch/audio-classification/run_audio_classification.py

@@ -17,6 +17,7 @@
 import logging
 import os
 import sys
+import warnings
 from dataclasses import dataclass, field
 from random import randint
 from typing import Optional
@@ -76,24 +77,24 @@ class DataTrainingArguments:
     eval_file: Optional[str] = field(
         default=None, metadata={"help": "A file containing the validation audio paths and labels."}
     )
-    train_split_name: Optional[str] = field(
+    train_split_name: str = field(
         default="train",
         metadata={
             "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
         },
     )
-    eval_split_name: Optional[str] = field(
+    eval_split_name: str = field(
         default="validation",
         metadata={
             "help": "The name of the training data set split to use (via the datasets library). Defaults to "
             "'validation'"
         },
     )
-    audio_column_name: Optional[str] = field(
+    audio_column_name: str = field(
         default="audio",
         metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
     )
-    label_column_name: Optional[str] = field(
+    label_column_name: str = field(
         default="label", metadata={"help": "The name of the dataset column containing the labels. Defaults to 'label'"}
     )
     max_train_samples: Optional[int] = field(
@@ -110,7 +111,7 @@ class DataTrainingArguments:
             "value if set."
         },
     )
-    max_length_seconds: Optional[float] = field(
+    max_length_seconds: float = field(
         default=20,
         metadata={"help": "Audio clips will be randomly cut to this length during training if the value is set."},
     )
@@ -136,11 +137,13 @@ class ModelArguments:
         default="main",
         metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
     )
-    feature_extractor_name: str = field(default=None, metadata={"help": "Name or path of preprocessor config."})
-    freeze_feature_extractor: Optional[bool] = field(
-        default=True, metadata={"help": "Whether to freeze the feature extractor layers of the model."}
+    feature_extractor_name: Optional[str] = field(
+        default=None, metadata={"help": "Name or path of preprocessor config."}
+    )
+    freeze_feature_encoder: bool = field(
+        default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
     )
-    attention_mask: Optional[bool] = field(
+    attention_mask: bool = field(
         default=True, metadata={"help": "Whether to generate an attention mask in the feature extractor."}
     )
     use_auth_token: bool = field(
@@ -150,6 +153,24 @@ class ModelArguments:
             "with private models)."
         },
     )
+    freeze_feature_extractor: Optional[bool] = field(
+        default=None, metadata={"help": "Whether to freeze the feature extractor layers of the model."}
+    )
+
+    def __post_init__(self):
+        if not self.freeze_feature_extractor and self.freeze_feature_encoder:
+            warnings.warn(
+                "The argument `--freeze_feature_extractor` is deprecated and "
+                "will be removed in a future version. Use `--freeze_feature_encoder`"
+                "instead. Setting `freeze_feature_encoder==True`.",
+                FutureWarning,
+            )
+        if self.freeze_feature_extractor and not self.freeze_feature_encoder:
+            raise ValueError(
+                "The argument `--freeze_feature_extractor` is deprecated and "
+                "should not be used in combination with `--freeze_feature_encoder`."
+                "Only make use of `--freeze_feature_encoder`."
+            )
 
 
 def main():
@@ -302,8 +323,8 @@ def main():
     )
 
     # freeze the convolutional waveform encoder
-    if model_args.freeze_feature_extractor:
-        model.freeze_feature_extractor()
+    if model_args.freeze_feature_encoder:
+        model.freeze_feature_encoder()
 
     if training_args.do_train:
         if data_args.max_train_samples is not None:

examples/pytorch/speech-recognition/README.md

@@ -78,7 +78,7 @@ python run_speech_recognition_ctc.py \
 	--eval_steps="100" \
 	--layerdrop="0.0" \
 	--save_total_limit="3" \
-	--freeze_feature_extractor \
+	--freeze_feature_encoder \
 	--gradient_checkpointing \
 	--chars_to_ignore , ? . ! - \; \: \" “ % ‘ ” � \
 	--fp16 \
@@ -113,7 +113,7 @@ python -m torch.distributed.launch \
 	--logging_steps="1" \
 	--layerdrop="0.0" \
 	--save_total_limit="3" \
-	--freeze_feature_extractor \
+	--freeze_feature_encoder \
 	--gradient_checkpointing \
 	--chars_to_ignore , ? . ! - \; \: \" “ % ‘ ” � \
 	--fp16 \
@@ -304,7 +304,7 @@ python run_speech_recognition_seq2seq.py \
 	--eval_steps="400" \
 	--logging_steps="10" \
 	--save_total_limit="1" \
-	--freeze_feature_extractor \
+	--freeze_feature_encoder \
 	--gradient_checkpointing \
 	--fp16 \
 	--group_by_length \
@@ -346,7 +346,7 @@ python -m torch.distributed.launch \
 	--eval_steps="400" \
 	--logging_steps="10" \
 	--save_total_limit="1" \
-	--freeze_feature_extractor \
+	--freeze_feature_encoder \
 	--gradient_checkpointing \
 	--fp16 \
 	--group_by_length \

examples/pytorch/speech-recognition/run_speech_recognition_ctc.py

@@ -78,29 +78,27 @@ class ModelArguments:
         default=None,
         metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
     )
-    freeze_feature_extractor: Optional[bool] = field(
-        default=True, metadata={"help": "Whether to freeze the feature extractor layers of the model."}
+    freeze_feature_encoder: bool = field(
+        default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
     )
-    attention_dropout: Optional[float] = field(
+    attention_dropout: float = field(
         default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
     )
-    activation_dropout: Optional[float] = field(
+    activation_dropout: float = field(
         default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
     )
-    feat_proj_dropout: Optional[float] = field(
-        default=0.0, metadata={"help": "The dropout ratio for the projected features."}
-    )
-    hidden_dropout: Optional[float] = field(
+    feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
+    hidden_dropout: float = field(
         default=0.0,
         metadata={
             "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
         },
     )
-    final_dropout: Optional[float] = field(
+    final_dropout: float = field(
         default=0.0,
         metadata={"help": "The dropout probability for the final projection layer."},
     )
-    mask_time_prob: Optional[float] = field(
+    mask_time_prob: float = field(
         default=0.05,
         metadata={
             "help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
@@ -108,22 +106,22 @@ class ModelArguments:
             "vectors will be masked along the time axis."
         },
     )
-    mask_time_length: Optional[int] = field(
+    mask_time_length: int = field(
         default=10,
         metadata={"help": "Length of vector span to mask along the time axis."},
     )
-    mask_feature_prob: Optional[float] = field(
+    mask_feature_prob: float = field(
         default=0.0,
         metadata={
             "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
             "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
         },
     )
-    mask_feature_length: Optional[int] = field(
+    mask_feature_length: int = field(
         default=10,
         metadata={"help": "Length of vector span to mask along the feature axis."},
     )
-    layerdrop: Optional[float] = field(default=0.0, metadata={"help": "The LayerDrop probability."})
+    layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
     ctc_loss_reduction: Optional[str] = field(
         default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
     )
@@ -142,26 +140,26 @@ class DataTrainingArguments:
     dataset_name: str = field(
         metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
     )
-    dataset_config_name: Optional[str] = field(
+    dataset_config_name: str = field(
         default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
     )
-    train_split_name: Optional[str] = field(
+    train_split_name: str = field(
         default="train+validation",
         metadata={
             "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
         },
     )
-    eval_split_name: Optional[str] = field(
+    eval_split_name: str = field(
         default="test",
         metadata={
             "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
         },
     )
-    audio_column_name: Optional[str] = field(
+    audio_column_name: str = field(
         default="audio",
         metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
     )
-    text_column_name: Optional[str] = field(
+    text_column_name: str = field(
         default="text",
         metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
     )
@@ -190,20 +188,20 @@ class DataTrainingArguments:
         default=None,
         metadata={"help": "A list of characters to remove from the transcripts."},
     )
-    eval_metrics: Optional[List[str]] = list_field(
+    eval_metrics: List[str] = list_field(
         default=["wer"],
         metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
     )
-    max_duration_in_seconds: Optional[float] = field(
+    max_duration_in_seconds: float = field(
         default=20.0,
         metadata={
             "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
         },
     )
-    min_duration_in_seconds: Optional[float] = field(
+    min_duration_in_seconds: float = field(
         default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
     )
-    preprocessing_only: Optional[bool] = field(
+    preprocessing_only: bool = field(
         default=False,
         metadata={
             "help": "Whether to only do data preprocessing and skip training. "
@@ -212,22 +210,22 @@ class DataTrainingArguments:
             "so that the cached datasets can consequently be loaded in distributed training"
         },
     )
-    use_auth_token: Optional[bool] = field(
+    use_auth_token: bool = field(
         default=False,
         metadata={
             "help": "If :obj:`True`, will use the token generated when running"
             ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
         },
     )
-    unk_token: Optional[str] = field(
+    unk_token: str = field(
         default="[UNK]",
         metadata={"help": "The unk token for the tokenizer"},
     )
-    pad_token: Optional[str] = field(
+    pad_token: str = field(
         default="[PAD]",
         metadata={"help": "The padding token for the tokenizer"},
     )
-    word_delimiter_token: Optional[str] = field(
+    word_delimiter_token: str = field(
         default="|",
         metadata={"help": "The word delimiter token for the tokenizer"},
     )
@@ -545,8 +543,8 @@ def main():
     )
 
     # freeze encoder
-    if model_args.freeze_feature_extractor:
-        model.freeze_feature_extractor()
+    if model_args.freeze_feature_encoder:
+        model.freeze_feature_encoder()
 
     # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
     # Thankfully, `datasets` takes care of automatically loading and resampling the audio,

examples/pytorch/speech-recognition/run_speech_recognition_seq2seq.py

@@ -91,8 +91,8 @@ class ModelArguments:
             "with private models)."
         },
     )
-    freeze_feature_extractor: Optional[bool] = field(
-        default=True, metadata={"help": "Whether to freeze the feature extractor layers of the model."}
+    freeze_feature_encoder: bool = field(
+        default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
     )
 
 
@@ -102,7 +102,7 @@ class DataTrainingArguments:
     Arguments pertaining to what data we are going to input our model for training and eval.
     """
 
-    dataset_name: Optional[str] = field(
+    dataset_name: str = field(
         default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
     )
     dataset_config_name: Optional[str] = field(
@@ -133,24 +133,24 @@ class DataTrainingArguments:
             "value if set."
         },
     )
-    audio_column_name: Optional[str] = field(
+    audio_column_name: str = field(
         default="audio",
         metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
     )
-    text_column_name: Optional[str] = field(
+    text_column_name: str = field(
         default="text",
         metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
     )
-    max_duration_in_seconds: Optional[float] = field(
+    max_duration_in_seconds: float = field(
         default=20.0,
         metadata={
             "help": "Truncate audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
         },
     )
-    min_duration_in_seconds: Optional[float] = field(
+    min_duration_in_seconds: float = field(
         default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
     )
-    preprocessing_only: Optional[bool] = field(
+    preprocessing_only: bool = field(
         default=False,
         metadata={
             "help": "Whether to only do data preprocessing and skip training. "
@@ -159,19 +159,19 @@ class DataTrainingArguments:
             "so that the cached datasets can consequently be loaded in distributed training"
         },
     )
-    train_split_name: Optional[str] = field(
+    train_split_name: str = field(
         default="train",
         metadata={
             "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
         },
     )
-    eval_split_name: Optional[str] = field(
+    eval_split_name: str = field(
         default="test",
         metadata={
             "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
         },
     )
-    do_lower_case: Optional[bool] = field(
+    do_lower_case: bool = field(
         default=True,
         metadata={"help": "Whether the target text should be lower cased."},
     )
@@ -335,8 +335,8 @@ def main():
     if model.config.decoder_start_token_id is None:
         raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")
 
-    if model_args.freeze_feature_extractor:
-        model.freeze_feature_extractor()
+    if model_args.freeze_feature_encoder:
+        model.freeze_feature_encoder()
 
     # 6. Resample speech dataset if necassary
     dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate

src/transformers/models/hubert/configuration_hubert.py

@@ -64,24 +64,24 @@ class HubertConfig(PretrainedConfig):
         layer_norm_eps (`float`, *optional*, defaults to 1e-12):
             The epsilon used by the layer normalization layers.
         feat_extract_norm (`str`, *optional*, defaults to `"group"`):
-            The norm to be applied to 1D convolutional layers in feature extractor. One of `"group"` for group
+            The norm to be applied to 1D convolutional layers in feature encoder. One of `"group"` for group
             normalization of only the first 1D convolutional layer or `"layer"` for layer normalization of all 1D
             convolutional layers.
         feat_proj_dropout (`float`, *optional*, defaults to 0.0):
-            The dropout probability for output of the feature extractor.
+            The dropout probability for output of the feature encoder.
         feat_proj_layer_norm (`bool`, *optional*, defaults to `True`):
-            Whether to apply LayerNorm to the output of the feature extractor.
+            Whether to apply LayerNorm to the output of the feature encoder.
         feat_extract_activation (`str, `optional`, defaults to `"gelu"`):
             The non-linear activation function (function or string) in the 1D convolutional layers of the feature
             extractor. If string, `"gelu"`, `"relu"`, `"selu"` and `"gelu_new"` are supported.
         conv_dim (`Tuple[int]`, *optional*, defaults to `(512, 512, 512, 512, 512, 512, 512)`):
             A tuple of integers defining the number of input and output channels of each 1D convolutional layer in the
-            feature extractor. The length of *conv_dim* defines the number of 1D convolutional layers.
+            feature encoder. The length of *conv_dim* defines the number of 1D convolutional layers.
         conv_stride (`Tuple[int]`, *optional*, defaults to `(5, 2, 2, 2, 2, 2, 2)`):
-            A tuple of integers defining the stride of each 1D convolutional layer in the feature extractor. The length
+            A tuple of integers defining the stride of each 1D convolutional layer in the feature encoder. The length
             of *conv_stride* defines the number of convolutional layers and has to match the the length of *conv_dim*.
         conv_kernel (`Tuple[int]`, *optional*, defaults to `(10, 3, 3, 3, 3, 3, 3)`):
-            A tuple of integers defining the kernel size of each 1D convolutional layer in the feature extractor. The
+            A tuple of integers defining the kernel size of each 1D convolutional layer in the feature encoder. The
             length of *conv_kernel* defines the number of convolutional layers and has to match the the length of
             *conv_dim*.
         conv_bias (`bool`, *optional*, defaults to `False`):
@@ -96,7 +96,7 @@ class HubertConfig(PretrainedConfig):
             True` corresponds to applying layer norm before the attention layer, whereas `do_stable_layer_norm is
             False` corresponds to applying layer norm after the attention layer.
         apply_spec_augment (`bool`, *optional*, defaults to `True`):
-            Whether to apply *SpecAugment* data augmentation to the outputs of the feature extractor. For reference see
+            Whether to apply *SpecAugment* data augmentation to the outputs of the feature encoder. For reference see
             [SpecAugment: A Simple Data Augmentation Method for Automatic Speech
             Recognition](https://arxiv.org/abs/1904.08779).
         mask_time_prob (`float`, *optional*, defaults to 0.05):

src/transformers/models/hubert/modeling_hubert.py

@@ -14,6 +14,7 @@
 # limitations under the License.
 """ PyTorch Hubert model."""
 
+import warnings
 from typing import Optional, Tuple, Union
 
 import numpy as np
@@ -284,8 +285,8 @@ class HubertSamePadLayer(nn.Module):
         return hidden_states
 
 
-# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeatureExtractor with Wav2Vec2->Hubert
-class HubertFeatureExtractor(nn.Module):
+# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeatureEncoder with Wav2Vec2->Hubert
+class HubertFeatureEncoder(nn.Module):
     """Construct the features from raw audio waveform"""
 
     def __init__(self, config):
@@ -336,6 +337,17 @@ class HubertFeatureExtractor(nn.Module):
         return hidden_states
 
 
+class HubertFeatureExtractor(HubertFeatureEncoder):
+    def __init__(self, config):
+        super().__init__(config)
+        warnings.warn(
+            f"The class `{self.__class__.__name__}` has been depreciated "
+            "and will be removed in Transformers v5. "
+            f"Use `{self.__class__.__bases__[0].__name__}` instead.",
+            FutureWarning,
+        )
+
+
 class HubertFeatureProjection(nn.Module):
     def __init__(self, config):
         super().__init__()
@@ -902,7 +914,7 @@ class HubertModel(HubertPreTrainedModel):
     def __init__(self, config: HubertConfig):
         super().__init__(config)
         self.config = config
-        self.feature_extractor = HubertFeatureExtractor(config)
+        self.feature_extractor = HubertFeatureEncoder(config)
         self.feature_projection = HubertFeatureProjection(config)
 
         if config.mask_time_prob > 0.0 or config.mask_feature_prob > 0.0:
@@ -1063,8 +1075,20 @@ class HubertForCTC(HubertPreTrainedModel):
 
     def freeze_feature_extractor(self):
         """
-        Calling this function will disable the gradient computation for the feature extractor so that its parameter
-        will not be updated during training.
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
+        """
+        warnings.warn(
+            "The method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5."
+            "Please use the equivalent `freeze_feature_encoder` method instead.",
+            FutureWarning,
+        )
+        self.freeze_feature_encoder()
+
+    def freeze_feature_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
         """
         self.hubert.feature_extractor._freeze_parameters()
 
@@ -1172,8 +1196,20 @@ class HubertForSequenceClassification(HubertPreTrainedModel):
 
     def freeze_feature_extractor(self):
         """
-        Calling this function will disable the gradient computation for the feature extractor so that its parameters
-        will not be updated during training.
+        Calling this function will disable the gradient computation for the feature encoder so that its parameters will
+        not be updated during training.
+        """
+        warnings.warn(
+            "The method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5."
+            "Please use the equivalent `freeze_feature_encoder` method instead.",
+            FutureWarning,
+        )
+        self.freeze_feature_encoder()
+
+    def freeze_feature_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
         """
         self.hubert.feature_extractor._freeze_parameters()
 

src/transformers/models/hubert/modeling_tf_hubert.py

@@ -659,7 +659,7 @@ class TFHubertSamePadLayer(tf.keras.layers.Layer):
         return hidden_states
 
 
-class TFHubertFeatureExtractor(tf.keras.layers.Layer):
+class TFHubertFeatureEncoder(tf.keras.layers.Layer):
     def __init__(self, config: HubertConfig, **kwargs: Any) -> None:
         super().__init__(**kwargs)
 
@@ -686,6 +686,17 @@ class TFHubertFeatureExtractor(tf.keras.layers.Layer):
         return hidden_states
 
 
+class TFHubertFeatureExtractor(TFHubertFeatureEncoder):
+    def __init__(self, config, **kwargs):
+        super().__init__(config, **kwargs)
+        warnings.warn(
+            f"The class `{self.__class__.__name__}` has been depreciated "
+            "and will be removed in Transformers v5. "
+            f"Use `{self.__class__.__bases__[0].__name__}` instead.",
+            FutureWarning,
+        )
+
+
 class TFHubertFeatureProjection(tf.keras.layers.Layer):
     def __init__(self, config: HubertConfig, **kwargs):
         super().__init__(**kwargs)
@@ -1116,7 +1127,7 @@ class TFHubertMainLayer(tf.keras.layers.Layer):
     def __init__(self, config: HubertConfig, **kwargs):
         super().__init__(**kwargs)
         self.config = config
-        self.feature_extractor = TFHubertFeatureExtractor(config, name="feature_extractor")
+        self.feature_extractor = TFHubertFeatureEncoder(config, name="feature_extractor")
         self.feature_projection = TFHubertFeatureProjection(config, name="feature_projection")
 
         if config.do_stable_layer_norm:
@@ -1490,8 +1501,20 @@ class TFHubertForCTC(TFHubertPreTrainedModel):
 
     def freeze_feature_extractor(self):
         """
-        Calling this function will disable the gradient computation for the feature extractor so that its parameter
-        will not be updated during training.
+        Calling this function will disable the gradient computation for the feature encoder so that its parameters will
+        not be updated during training.
+        """
+        warnings.warn(
+            "The method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5."
+            "Please use the equivalent `freeze_feature_encoder` method instead.",
+            FutureWarning,
+        )
+        self.freeze_feature_encoder()
+
+    def freeze_feature_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
         """
         self.hubert.feature_extractor.trainable = False
 

src/transformers/models/sew/configuration_sew.py

@@ -65,22 +65,22 @@ class SEWConfig(PretrainedConfig):
         layer_norm_eps (`float`, *optional*, defaults to 1e-12):
             The epsilon used by the layer normalization layers.
         feat_extract_norm (`str`, *optional*, defaults to `"group"`):
-            The norm to be applied to 1D convolutional layers in feature extractor. One of `"group"` for group
+            The norm to be applied to 1D convolutional layers in feature encoder. One of `"group"` for group
             normalization of only the first 1D convolutional layer or `"layer"` for layer normalization of all 1D
             convolutional layers.
         feat_proj_dropout (`float`, *optional*, defaults to 0.0):
-            The dropout probability for output of the feature extractor.
+            The dropout probability for output of the feature encoder.
         feat_extract_activation (`str, `optional`, defaults to `"gelu"`):
             The non-linear activation function (function or string) in the 1D convolutional layers of the feature
             extractor. If string, `"gelu"`, `"relu"`, `"selu"` and `"gelu_new"` are supported.
         conv_dim (`Tuple[int]`, *optional*, defaults to `(64, 128, 128, 128, 128, 256, 256, 256, 256, 512, 512, 512, 512)`):
             A tuple of integers defining the number of input and output channels of each 1D convolutional layer in the
-            feature extractor. The length of *conv_dim* defines the number of 1D convolutional layers.
+            feature encoder. The length of *conv_dim* defines the number of 1D convolutional layers.
         conv_stride (`Tuple[int]`, *optional*, defaults to `(5, 2, 1, 2, 1, 2, 1, 2, 1, 2, 1, 2, 1)`):
-            A tuple of integers defining the stride of each 1D convolutional layer in the feature extractor. The length
+            A tuple of integers defining the stride of each 1D convolutional layer in the feature encoder. The length
             of *conv_stride* defines the number of convolutional layers and has to match the the length of *conv_dim*.
         conv_kernel (`Tuple[int]`, *optional*, defaults to `(10, 3, 1, 3, 1, 3, 1, 3, 1, 2, 1, 2, 1)`):
-            A tuple of integers defining the kernel size of each 1D convolutional layer in the feature extractor. The
+            A tuple of integers defining the kernel size of each 1D convolutional layer in the feature encoder. The
             length of *conv_kernel* defines the number of convolutional layers and has to match the the length of
             *conv_dim*.
         conv_bias (`bool`, *optional*, defaults to `False`):
@@ -91,7 +91,7 @@ class SEWConfig(PretrainedConfig):
         num_conv_pos_embedding_groups (`int`, *optional*, defaults to 16):
             Number of groups of 1D convolutional positional embeddings layer.
         apply_spec_augment (`bool`, *optional*, defaults to `True`):
-            Whether to apply *SpecAugment* data augmentation to the outputs of the feature extractor. For reference see
+            Whether to apply *SpecAugment* data augmentation to the outputs of the feature encoder. For reference see
             [SpecAugment: A Simple Data Augmentation Method for Automatic Speech
             Recognition](https://arxiv.org/abs/1904.08779).
         mask_time_prob (`float`, *optional*, defaults to 0.05):

src/transformers/models/sew/modeling_sew.py

@@ -15,6 +15,7 @@
 """ PyTorch SEW model."""
 
 import math
+import warnings
 from typing import Optional, Tuple, Union
 
 import numpy as np
@@ -301,8 +302,8 @@ class SEWUpsampling(nn.Module):
         return hidden_states
 
 
-# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeatureExtractor with Wav2Vec2->SEW
-class SEWFeatureExtractor(nn.Module):
+# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeatureEncoder with Wav2Vec2->SEW
+class SEWFeatureEncoder(nn.Module):
     """Construct the features from raw audio waveform"""
 
     def __init__(self, config):
@@ -353,6 +354,17 @@ class SEWFeatureExtractor(nn.Module):
         return hidden_states
 
 
+class SEWFeatureExtractor(SEWFeatureEncoder):
+    def __init__(self, config):
+        super().__init__(config)
+        warnings.warn(
+            f"The class `{self.__class__.__name__}` has been depreciated "
+            "and will be removed in Transformers v5. "
+            f"Use `{self.__class__.__bases__[0].__name__}` instead.",
+            FutureWarning,
+        )
+
+
 # Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->SEW
 class SEWAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
@@ -712,7 +724,7 @@ class SEWPreTrainedModel(PreTrainedModel):
             module.bias.data.zero_()
 
     def _set_gradient_checkpointing(self, module, value=False):
-        if isinstance(module, (SEWEncoder, SEWFeatureExtractor)):
+        if isinstance(module, (SEWEncoder, SEWFeatureEncoder)):
             module.gradient_checkpointing = value
 
     def _get_feat_extract_output_lengths(self, input_lengths: Union[torch.LongTensor, int]):
@@ -797,7 +809,7 @@ class SEWModel(SEWPreTrainedModel):
     def __init__(self, config: SEWConfig):
         super().__init__(config)
         self.config = config
-        self.feature_extractor = SEWFeatureExtractor(config)
+        self.feature_extractor = SEWFeatureEncoder(config)
         self.layer_norm = nn.LayerNorm(config.conv_dim[-1], eps=config.layer_norm_eps)
 
         self.project_features = config.conv_dim[-1] != config.hidden_size
@@ -943,8 +955,20 @@ class SEWForCTC(SEWPreTrainedModel):
 
     def freeze_feature_extractor(self):
         """
-        Calling this function will disable the gradient computation for the feature extractor so that its parameter
-        will not be updated during training.
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
+        """
+        warnings.warn(
+            "The method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5."
+            "Please use the equivalent `freeze_feature_encoder` method instead.",
+            FutureWarning,
+        )
+        self.freeze_feature_encoder()
+
+    def freeze_feature_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
         """
         self.sew.feature_extractor._freeze_parameters()
 
@@ -1052,8 +1076,20 @@ class SEWForSequenceClassification(SEWPreTrainedModel):
 
     def freeze_feature_extractor(self):
         """
-        Calling this function will disable the gradient computation for the feature extractor so that its parameters
-        will not be updated during training.
+        Calling this function will disable the gradient computation for the feature encoder so that its parameters will
+        not be updated during training.
+        """
+        warnings.warn(
+            "The method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5."
+            "Please use the equivalent `freeze_feature_encoder` method instead.",
+            FutureWarning,
+        )
+        self.freeze_feature_encoder()
+
+    def freeze_feature_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
         """
         self.sew.feature_extractor._freeze_parameters()
 

src/transformers/models/sew_d/configuration_sew_d.py

@@ -81,24 +81,24 @@ class SEWDConfig(PretrainedConfig):
         layer_norm_eps (`float`, *optional*, defaults to 1e-7):
             The epsilon used by the layer normalization layers in the transformer encoder.
         feature_layer_norm_eps (`float`, *optional*, defaults to 1e-5):
-            The epsilon used by the layer normalization after the feature extractor.
+            The epsilon used by the layer normalization after the feature encoder.
         feat_extract_norm (`str`, *optional*, defaults to `"group"`):
-            The norm to be applied to 1D convolutional layers in feature extractor. One of `"group"` for group
+            The norm to be applied to 1D convolutional layers in feature encoder. One of `"group"` for group
             normalization of only the first 1D convolutional layer or `"layer"` for layer normalization of all 1D
             convolutional layers.
         feat_proj_dropout (`float`, *optional*, defaults to 0.0):
-            The dropout probability for output of the feature extractor.
+            The dropout probability for output of the feature encoder.
         feat_extract_activation (`str, `optional`, defaults to `"gelu"`):
             The non-linear activation function (function or string) in the 1D convolutional layers of the feature
             extractor. If string, `"gelu"`, `"relu"`, `"selu"` and `"gelu_new"` are supported.
         conv_dim (`Tuple[int]`, *optional*, defaults to `(64, 128, 128, 128, 128, 256, 256, 256, 256, 512, 512, 512, 512)`):
             A tuple of integers defining the number of input and output channels of each 1D convolutional layer in the
-            feature extractor. The length of *conv_dim* defines the number of 1D convolutional layers.
+            feature encoder. The length of *conv_dim* defines the number of 1D convolutional layers.
         conv_stride (`Tuple[int]`, *optional*, defaults to `(5, 2, 1, 2, 1, 2, 1, 2, 1, 2, 1, 2, 1)`):
-            A tuple of integers defining the stride of each 1D convolutional layer in the feature extractor. The length
+            A tuple of integers defining the stride of each 1D convolutional layer in the feature encoder. The length
             of *conv_stride* defines the number of convolutional layers and has to match the the length of *conv_dim*.
         conv_kernel (`Tuple[int]`, *optional*, defaults to `(10, 3, 1, 3, 1, 3, 1, 3, 1, 2, 1, 2, 1)`):
-            A tuple of integers defining the kernel size of each 1D convolutional layer in the feature extractor. The
+            A tuple of integers defining the kernel size of each 1D convolutional layer in the feature encoder. The
             length of *conv_kernel* defines the number of convolutional layers and has to match the the length of
             *conv_dim*.
         conv_bias (`bool`, *optional*, defaults to `False`):
@@ -109,7 +109,7 @@ class SEWDConfig(PretrainedConfig):
         num_conv_pos_embedding_groups (`int`, *optional*, defaults to 16):
             Number of groups of 1D convolutional positional embeddings layer.
         apply_spec_augment (`bool`, *optional*, defaults to `True`):
-            Whether to apply *SpecAugment* data augmentation to the outputs of the feature extractor. For reference see
+            Whether to apply *SpecAugment* data augmentation to the outputs of the feature encoder. For reference see
             [SpecAugment: A Simple Data Augmentation Method for Automatic Speech
             Recognition](https://arxiv.org/abs/1904.08779).
         mask_time_prob (`float`, *optional*, defaults to 0.05):

src/transformers/models/sew_d/modeling_sew_d.py

@@ -15,6 +15,7 @@
 """ PyTorch SEW model."""
 
 import math
+import warnings
 from collections.abc import Sequence
 from typing import Optional, Tuple, Union
 
@@ -387,8 +388,8 @@ class SEWDUpsampling(nn.Module):
         return hidden_states
 
 
-# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeatureExtractor with Wav2Vec2->SEWD
-class SEWDFeatureExtractor(nn.Module):
+# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeatureEncoder with Wav2Vec2->SEWD
+class SEWDFeatureEncoder(nn.Module):
     """Construct the features from raw audio waveform"""
 
     def __init__(self, config):
@@ -439,6 +440,17 @@ class SEWDFeatureExtractor(nn.Module):
         return hidden_states
 
 
+class SEWDFeatureExtractor(SEWDFeatureEncoder):
+    def __init__(self, config):
+        super().__init__(config)
+        warnings.warn(
+            f"The class `{self.__class__.__name__}` has been depreciated "
+            "and will be removed in Transformers v5. "
+            f"Use `{self.__class__.__bases__[0].__name__}` instead.",
+            FutureWarning,
+        )
+
+
 # Copied from transformers.models.deberta.modeling_deberta.ContextPooler
 class ContextPooler(nn.Module):
     def __init__(self, config):
@@ -1333,7 +1345,7 @@ class SEWDModel(SEWDPreTrainedModel):
     def __init__(self, config: SEWDConfig):
         super().__init__(config)
         self.config = config
-        self.feature_extractor = SEWDFeatureExtractor(config)
+        self.feature_extractor = SEWDFeatureEncoder(config)
         self.layer_norm = nn.LayerNorm(config.conv_dim[-1], eps=config.feature_layer_norm_eps)
 
         self.project_features = config.conv_dim[-1] != config.hidden_size
@@ -1479,8 +1491,20 @@ class SEWDForCTC(SEWDPreTrainedModel):
 
     def freeze_feature_extractor(self):
         """
-        Calling this function will disable the gradient computation for the feature extractor so that its parameter
-        will not be updated during training.
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
+        """
+        warnings.warn(
+            "The method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5."
+            "Please use the equivalent `freeze_feature_encoder` method instead.",
+            FutureWarning,
+        )
+        self.freeze_feature_encoder()
+
+    def freeze_feature_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
         """
         self.sew_d.feature_extractor._freeze_parameters()
 
@@ -1588,8 +1612,20 @@ class SEWDForSequenceClassification(SEWDPreTrainedModel):
 
     def freeze_feature_extractor(self):
         """
-        Calling this function will disable the gradient computation for the feature extractor so that its parameters
-        will not be updated during training.
+        Calling this function will disable the gradient computation for the feature encoder so that its parameters will
+        not be updated during training.
+        """
+        warnings.warn(
+            "The method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5."
+            "Please use the equivalent `freeze_feature_encoder` method instead.",
+            FutureWarning,
+        )
+        self.freeze_feature_encoder()
+
+    def freeze_feature_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
         """
         self.sew_d.feature_extractor._freeze_parameters()
 

src/transformers/models/speech_encoder_decoder/modeling_speech_encoder_decoder.py

@@ -265,12 +265,12 @@ class SpeechEncoderDecoderModel(PreTrainedModel):
     def set_output_embeddings(self, new_embeddings):
         return self.decoder.set_output_embeddings(new_embeddings)
 
-    def freeze_feature_extractor(self):
+    def freeze_feature_encoder(self):
         """
-        Calling this function will disable the gradient computation for the feature extractor of the speech encoder so
+        Calling this function will disable the gradient computation for the feature encoder of the speech encoder so
         that its parameters will not be updated during training.
         """
-        self.encoder.freeze_feature_extractor()
+        self.encoder.freeze_feature_encoder()
 
     @classmethod
     def from_pretrained(cls, *args, **kwargs):

src/transformers/models/unispeech/configuration_unispeech.py

@@ -65,24 +65,24 @@ class UniSpeechConfig(PretrainedConfig):
         layer_norm_eps (`float`, *optional*, defaults to 1e-12):
             The epsilon used by the layer normalization layers.
         feat_extract_norm (`str`, *optional*, defaults to `"group"`):
-            The norm to be applied to 1D convolutional layers in feature extractor. One of `"group"` for group
+            The norm to be applied to 1D convolutional layers in feature encoder. One of `"group"` for group
             normalization of only the first 1D convolutional layer or `"layer"` for layer normalization of all 1D
             convolutional layers.
         feat_proj_dropout (`float`, *optional*, defaults to 0.0):
-            The dropout probability for output of the feature extractor.
+            The dropout probability for output of the feature encoder.
         feat_extract_activation (`str, `optional`, defaults to `"gelu"`):
             The non-linear activation function (function or string) in the 1D convolutional layers of the feature
             extractor. If string, `"gelu"`, `"relu"`, `"selu"` and `"gelu_new"` are supported.
         feat_quantizer_dropout (`float`, *optional*, defaults to 0.0):
-            The dropout probabilitiy for quantized feature extractor states.
+            The dropout probabilitiy for quantized feature encoder states.
         conv_dim (`Tuple[int]`, *optional*, defaults to `(512, 512, 512, 512, 512, 512, 512)`):
             A tuple of integers defining the number of input and output channels of each 1D convolutional layer in the
-            feature extractor. The length of *conv_dim* defines the number of 1D convolutional layers.
+            feature encoder. The length of *conv_dim* defines the number of 1D convolutional layers.
         conv_stride (`Tuple[int]`, *optional*, defaults to `(5, 2, 2, 2, 2, 2, 2)`):
-            A tuple of integers defining the stride of each 1D convolutional layer in the feature extractor. The length
+            A tuple of integers defining the stride of each 1D convolutional layer in the feature encoder. The length
             of *conv_stride* defines the number of convolutional layers and has to match the the length of *conv_dim*.
         conv_kernel (`Tuple[int]`, *optional*, defaults to `(10, 3, 3, 3, 3, 3, 3)`):
-            A tuple of integers defining the kernel size of each 1D convolutional layer in the feature extractor. The
+            A tuple of integers defining the kernel size of each 1D convolutional layer in the feature encoder. The
             length of *conv_kernel* defines the number of convolutional layers and has to match the the length of
             *conv_dim*.
         conv_bias (`bool`, *optional*, defaults to `False`):
@@ -97,7 +97,7 @@ class UniSpeechConfig(PretrainedConfig):
             True` corresponds to applying layer norm before the attention layer, whereas `do_stable_layer_norm is
             False` corresponds to applying layer norm after the attention layer.
         apply_spec_augment (`bool`, *optional*, defaults to `True`):
-            Whether to apply *SpecAugment* data augmentation to the outputs of the feature extractor. For reference see
+            Whether to apply *SpecAugment* data augmentation to the outputs of the feature encoder. For reference see
             [SpecAugment: A Simple Data Augmentation Method for Automatic Speech
             Recognition](https://arxiv.org/abs/1904.08779).
         mask_time_prob (`float`, *optional*, defaults to 0.05):
@@ -132,7 +132,7 @@ class UniSpeechConfig(PretrainedConfig):
         contrastive_logits_temperature (`float`, *optional*, defaults to 0.1):
             The temperature *kappa* in the contrastive loss.
         feat_quantizer_dropout (`float`, *optional*, defaults to 0.0):
-            The dropout probabilitiy for the output of the feature extractor that's used by the quantizer.
+            The dropout probabilitiy for the output of the feature encoder that's used by the quantizer.
         num_negatives (`int`, *optional*, defaults to 100):
             Number of negative samples for the contrastive loss.
         codevector_dim (`int`, *optional*, defaults to 256):

src/transformers/models/unispeech/modeling_unispeech.py

@@ -15,6 +15,7 @@
 """ PyTorch UniSpeech model."""
 
 import math
+import warnings
 from dataclasses import dataclass
 from typing import Optional, Tuple, Union
 
@@ -351,8 +352,8 @@ class UniSpeechSamePadLayer(nn.Module):
         return hidden_states
 
 
-# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeatureExtractor with Wav2Vec2->UniSpeech
-class UniSpeechFeatureExtractor(nn.Module):
+# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeatureEncoder with Wav2Vec2->UniSpeech
+class UniSpeechFeatureEncoder(nn.Module):
     """Construct the features from raw audio waveform"""
 
     def __init__(self, config):
@@ -406,6 +407,17 @@ class UniSpeechFeatureExtractor(nn.Module):
         return hidden_states
 
 
+class UniSpeechFeatureExtractor(UniSpeechFeatureEncoder):
+    def __init__(self, config):
+        super().__init__(config)
+        warnings.warn(
+            f"The class `{self.__class__.__name__}` has been depreciated "
+            "and will be removed in Transformers v5. "
+            f"Use `{self.__class__.__bases__[0].__name__}` instead.",
+            FutureWarning,
+        )
+
+
 # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeatureProjection with Wav2Vec2->UniSpeech
 class UniSpeechFeatureProjection(nn.Module):
     def __init__(self, config):
@@ -980,7 +992,7 @@ class UniSpeechPreTrainedModel(PreTrainedModel):
         return attention_mask
 
     def _set_gradient_checkpointing(self, module, value=False):
-        if isinstance(module, (UniSpeechEncoder, UniSpeechEncoderStableLayerNorm, UniSpeechFeatureExtractor)):
+        if isinstance(module, (UniSpeechEncoder, UniSpeechEncoderStableLayerNorm, UniSpeechFeatureEncoder)):
             module.gradient_checkpointing = value
 
 
@@ -1049,7 +1061,7 @@ class UniSpeechModel(UniSpeechPreTrainedModel):
     def __init__(self, config: UniSpeechConfig):
         super().__init__(config)
         self.config = config
-        self.feature_extractor = UniSpeechFeatureExtractor(config)
+        self.feature_extractor = UniSpeechFeatureEncoder(config)
         self.feature_projection = UniSpeechFeatureProjection(config)
 
         if config.mask_time_prob > 0.0 or config.mask_feature_prob > 0.0:
@@ -1193,8 +1205,20 @@ class UniSpeechForPreTraining(UniSpeechPreTrainedModel):
 
     def freeze_feature_extractor(self):
         """
-        Calling this function will disable the gradient computation for the feature extractor so that its parameters
-        will not be updated during training.
+        Calling this function will disable the gradient computation for the feature encoder so that its parameters will
+        not be updated during training.
+        """
+        warnings.warn(
+            "The method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5."
+            "Please use the equivalent `freeze_feature_encoder` method instead.",
+            FutureWarning,
+        )
+        self.freeze_feature_encoder()
+
+    def freeze_feature_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
         """
         self.unispeech.feature_extractor._freeze_parameters()
 
@@ -1358,8 +1382,20 @@ class UniSpeechForCTC(UniSpeechPreTrainedModel):
 
     def freeze_feature_extractor(self):
         """
-        Calling this function will disable the gradient computation for the feature extractor so that its parameter
-        will not be updated during training.
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
+        """
+        warnings.warn(
+            "The method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5."
+            "Please use the equivalent `freeze_feature_encoder` method instead.",
+            FutureWarning,
+        )
+        self.freeze_feature_encoder()
+
+    def freeze_feature_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
         """
         self.unispeech.feature_extractor._freeze_parameters()
 
@@ -1467,8 +1503,20 @@ class UniSpeechForSequenceClassification(UniSpeechPreTrainedModel):
 
     def freeze_feature_extractor(self):
         """
-        Calling this function will disable the gradient computation for the feature extractor so that its parameters
-        will not be updated during training.
+        Calling this function will disable the gradient computation for the feature encoder so that its parameters will
+        not be updated during training.
+        """
+        warnings.warn(
+            "The method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5."
+            "Please use the equivalent `freeze_feature_encoder` method instead.",
+            FutureWarning,
+        )
+        self.freeze_feature_encoder()
+
+    def freeze_feature_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
         """
         self.unispeech.feature_extractor._freeze_parameters()
 

src/transformers/models/unispeech_sat/configuration_unispeech_sat.py

@@ -65,24 +65,24 @@ class UniSpeechSatConfig(PretrainedConfig):
         layer_norm_eps (`float`, *optional*, defaults to 1e-12):
             The epsilon used by the layer normalization layers.
         feat_extract_norm (`str`, *optional*, defaults to `"group"`):
-            The norm to be applied to 1D convolutional layers in feature extractor. One of `"group"` for group
+            The norm to be applied to 1D convolutional layers in feature encoder. One of `"group"` for group
             normalization of only the first 1D convolutional layer or `"layer"` for layer normalization of all 1D
             convolutional layers.
         feat_proj_dropout (`float`, *optional*, defaults to 0.0):
-            The dropout probability for output of the feature extractor.
+            The dropout probability for output of the feature encoder.
         feat_extract_activation (`str, `optional`, defaults to `"gelu"`):
             The non-linear activation function (function or string) in the 1D convolutional layers of the feature
             extractor. If string, `"gelu"`, `"relu"`, `"selu"` and `"gelu_new"` are supported.
         feat_quantizer_dropout (`float`, *optional*, defaults to 0.0):
-            The dropout probabilitiy for quantized feature extractor states.
+            The dropout probabilitiy for quantized feature encoder states.
         conv_dim (`Tuple[int]`, *optional*, defaults to `(512, 512, 512, 512, 512, 512, 512)`):
             A tuple of integers defining the number of input and output channels of each 1D convolutional layer in the
-            feature extractor. The length of *conv_dim* defines the number of 1D convolutional layers.
+            feature encoder. The length of *conv_dim* defines the number of 1D convolutional layers.
         conv_stride (`Tuple[int]`, *optional*, defaults to `(5, 2, 2, 2, 2, 2, 2)`):
-            A tuple of integers defining the stride of each 1D convolutional layer in the feature extractor. The length
+            A tuple of integers defining the stride of each 1D convolutional layer in the feature encoder. The length
             of *conv_stride* defines the number of convolutional layers and has to match the the length of *conv_dim*.
         conv_kernel (`Tuple[int]`, *optional*, defaults to `(10, 3, 3, 3, 3, 3, 3)`):
-            A tuple of integers defining the kernel size of each 1D convolutional layer in the feature extractor. The
+            A tuple of integers defining the kernel size of each 1D convolutional layer in the feature encoder. The
             length of *conv_kernel* defines the number of convolutional layers and has to match the the length of
             *conv_dim*.
         conv_bias (`bool`, *optional*, defaults to `False`):
@@ -97,7 +97,7 @@ class UniSpeechSatConfig(PretrainedConfig):
             True` corresponds to applying layer norm before the attention layer, whereas `do_stable_layer_norm is
             False` corresponds to applying layer norm after the attention layer.
         apply_spec_augment (`bool`, *optional*, defaults to `True`):
-            Whether to apply *SpecAugment* data augmentation to the outputs of the feature extractor. For reference see
+            Whether to apply *SpecAugment* data augmentation to the outputs of the feature encoder. For reference see
             [SpecAugment: A Simple Data Augmentation Method for Automatic Speech
             Recognition](https://arxiv.org/abs/1904.08779).
         mask_time_prob (`float`, *optional*, defaults to 0.05):
@@ -132,7 +132,7 @@ class UniSpeechSatConfig(PretrainedConfig):
         contrastive_logits_temperature (`float`, *optional*, defaults to 0.1):
             The temperature *kappa* in the contrastive loss.
         feat_quantizer_dropout (`float`, *optional*, defaults to 0.0):
-            The dropout probabilitiy for the output of the feature extractor that's used by the quantizer.
+            The dropout probabilitiy for the output of the feature encoder that's used by the quantizer.
         num_negatives (`int`, *optional*, defaults to 100):
             Number of negative samples for the contrastive loss.
         codevector_dim (`int`, *optional*, defaults to 256):

src/transformers/models/unispeech_sat/convert_unispeech_sat_original_pytorch_checkpoint_to_pytorch.py

@@ -20,12 +20,7 @@ import argparse
 import fairseq
 import torch
 
-from transformers import (  # UniSpeechSatCTCTokenizer,; UniSpeechSatFeatureExtractor,; UniSpeechSatProcessor,
-    UniSpeechSatConfig,
-    UniSpeechSatForCTC,
-    UniSpeechSatForPreTraining,
-    logging,
-)
+from transformers import UniSpeechSatConfig, UniSpeechSatForCTC, UniSpeechSatForPreTraining, logging
 
 
 logging.set_verbosity_info()

src/transformers/models/unispeech_sat/modeling_unispeech_sat.py

@@ -15,6 +15,7 @@
 """ PyTorch UniSpeechSat model."""
 
 import math
+import warnings
 from dataclasses import dataclass
 from typing import Optional, Tuple, Union
 
@@ -385,8 +386,8 @@ class UniSpeechSatSamePadLayer(nn.Module):
         return hidden_states
 
 
-# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeatureExtractor with Wav2Vec2->UniSpeechSat
-class UniSpeechSatFeatureExtractor(nn.Module):
+# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeatureEncoder with Wav2Vec2->UniSpeechSat
+class UniSpeechSatFeatureEncoder(nn.Module):
     """Construct the features from raw audio waveform"""
 
     def __init__(self, config):
@@ -440,6 +441,17 @@ class UniSpeechSatFeatureExtractor(nn.Module):
         return hidden_states
 
 
+class UniSpeechSatFeatureExtractor(UniSpeechSatFeatureEncoder):
+    def __init__(self, config):
+        super().__init__(config)
+        warnings.warn(
+            f"The class `{self.__class__.__name__}` has been depreciated "
+            "and will be removed in Transformers v5. "
+            f"Use `{self.__class__.__bases__[0].__name__}` instead.",
+            FutureWarning,
+        )
+
+
 # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2FeatureProjection with Wav2Vec2->UniSpeechSat
 class UniSpeechSatFeatureProjection(nn.Module):
     def __init__(self, config):
@@ -1014,7 +1026,7 @@ class UniSpeechSatPreTrainedModel(PreTrainedModel):
         return attention_mask
 
     def _set_gradient_checkpointing(self, module, value=False):
-        if isinstance(module, (UniSpeechSatEncoder, UniSpeechSatEncoderStableLayerNorm, UniSpeechSatFeatureExtractor)):
+        if isinstance(module, (UniSpeechSatEncoder, UniSpeechSatEncoderStableLayerNorm, UniSpeechSatFeatureEncoder)):
             module.gradient_checkpointing = value
 
 
@@ -1084,7 +1096,7 @@ class UniSpeechSatModel(UniSpeechSatPreTrainedModel):
     def __init__(self, config: UniSpeechSatConfig):
         super().__init__(config)
         self.config = config
-        self.feature_extractor = UniSpeechSatFeatureExtractor(config)
+        self.feature_extractor = UniSpeechSatFeatureEncoder(config)
         self.feature_projection = UniSpeechSatFeatureProjection(config)
 
         self.masked_spec_embed = nn.Parameter(torch.FloatTensor(config.hidden_size).uniform_())
@@ -1232,10 +1244,22 @@ class UniSpeechSatForPreTraining(UniSpeechSatPreTrainedModel):
 
     def freeze_feature_extractor(self):
         """
-        Calling this function will disable the gradient computation for the feature extractor so that its parameters
-        will not be updated during training.
+        Calling this function will disable the gradient computation for the feature encoder so that its parameters will
+        not be updated during training.
         """
-        self.unispeech_sat.feature_extractor._freeze_parameters()
+        warnings.warn(
+            "The method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5."
+            "Please use the equivalent `freeze_feature_encoder` method instead.",
+            FutureWarning,
+        )
+        self.freeze_feature_encoder()
+
+    def freeze_feature_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
+        """
+        self.wav2vec2.feature_extractor._freeze_parameters()
 
     @staticmethod
     def compute_contrastive_logits(
@@ -1274,12 +1298,12 @@ class UniSpeechSatForPreTraining(UniSpeechSatPreTrainedModel):
 
         ```python
         >>> import torch
-        >>> from transformers import UniSpeechSatFeatureExtractor, UniSpeechSatForPreTraining
+        >>> from transformers import UniSpeechSatFeatureEncoder, UniSpeechSatForPreTraining
         >>> from transformers.models.unispeech_sat.modeling_unispeech_sat import _compute_mask_indices
         >>> from datasets import load_dataset
         >>> import soundfile as sf
 
-        >>> feature_extractor = UniSpeechSatFeatureExtractor.from_pretrained("patrickvonplaten/unispeech_sat-base")
+        >>> feature_extractor = UniSpeechSatFeatureEncoder.from_pretrained("patrickvonplaten/unispeech_sat-base")
         >>> model = UniSpeechSatForPreTraining.from_pretrained("patrickvonplaten/unispeech_sat-base")
 
 
@@ -1383,8 +1407,20 @@ class UniSpeechSatForCTC(UniSpeechSatPreTrainedModel):
 
     def freeze_feature_extractor(self):
         """
-        Calling this function will disable the gradient computation for the feature extractor so that its parameter
-        will not be updated during training.
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
+        """
+        warnings.warn(
+            "The method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5."
+            "Please use the equivalent `freeze_feature_encoder` method instead.",
+            FutureWarning,
+        )
+        self.freeze_feature_encoder()
+
+    def freeze_feature_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
         """
         self.unispeech_sat.feature_extractor._freeze_parameters()
 
@@ -1492,8 +1528,20 @@ class UniSpeechSatForSequenceClassification(UniSpeechSatPreTrainedModel):
 
     def freeze_feature_extractor(self):
         """
-        Calling this function will disable the gradient computation for the feature extractor so that its parameters
-        will not be updated during training.
+        Calling this function will disable the gradient computation for the feature encoder so that its parameters will
+        not be updated during training.
+        """
+        warnings.warn(
+            "The method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5."
+            "Please use the equivalent `freeze_feature_encoder` method instead.",
+            FutureWarning,
+        )
+        self.freeze_feature_encoder()
+
+    def freeze_feature_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
         """
         self.unispeech_sat.feature_extractor._freeze_parameters()
 
@@ -1596,8 +1644,20 @@ class UniSpeechSatForAudioFrameClassification(UniSpeechSatPreTrainedModel):
 
     def freeze_feature_extractor(self):
         """
-        Calling this function will disable the gradient computation for the feature extractor so that its parameters
-        will not be updated during training.
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
+        """
+        warnings.warn(
+            "The method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5."
+            "Please use the equivalent `freeze_feature_encoder` method instead.",
+            FutureWarning,
+        )
+        self.freeze_feature_encoder()
+
+    def freeze_feature_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
         """
         self.unispeech_sat.feature_extractor._freeze_parameters()
 
@@ -1745,8 +1805,20 @@ class UniSpeechSatForXVector(UniSpeechSatPreTrainedModel):
 
     def freeze_feature_extractor(self):
         """
-        Calling this function will disable the gradient computation for the feature extractor so that its parameters
-        will not be updated during training.
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
+        """
+        warnings.warn(
+            "The method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5."
+            "Please use the equivalent `freeze_feature_encoder` method instead.",
+            FutureWarning,
+        )
+        self.freeze_feature_encoder()
+
+    def freeze_feature_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
         """
         self.unispeech_sat.feature_extractor._freeze_parameters()
 

src/transformers/models/wav2vec2/configuration_wav2vec2.py

@@ -65,24 +65,24 @@ class Wav2Vec2Config(PretrainedConfig):
         layer_norm_eps (`float`, *optional*, defaults to 1e-12):
             The epsilon used by the layer normalization layers.
         feat_extract_norm (`str`, *optional*, defaults to `"group"`):
-            The norm to be applied to 1D convolutional layers in feature extractor. One of `"group"` for group
+            The norm to be applied to 1D convolutional layers in feature encoder. One of `"group"` for group
             normalization of only the first 1D convolutional layer or `"layer"` for layer normalization of all 1D
             convolutional layers.
         feat_proj_dropout (`float`, *optional*, defaults to 0.0):
-            The dropout probability for output of the feature extractor.
+            The dropout probability for output of the feature encoder.
         feat_extract_activation (`str, `optional`, defaults to `"gelu"`):
             The non-linear activation function (function or string) in the 1D convolutional layers of the feature
             extractor. If string, `"gelu"`, `"relu"`, `"selu"` and `"gelu_new"` are supported.
         feat_quantizer_dropout (`float`, *optional*, defaults to 0.0):
-            The dropout probabilitiy for quantized feature extractor states.
+            The dropout probabilitiy for quantized feature encoder states.
         conv_dim (`Tuple[int]`, *optional*, defaults to `(512, 512, 512, 512, 512, 512, 512)`):
             A tuple of integers defining the number of input and output channels of each 1D convolutional layer in the
-            feature extractor. The length of *conv_dim* defines the number of 1D convolutional layers.
+            feature encoder. The length of *conv_dim* defines the number of 1D convolutional layers.
         conv_stride (`Tuple[int]`, *optional*, defaults to `(5, 2, 2, 2, 2, 2, 2)`):
-            A tuple of integers defining the stride of each 1D convolutional layer in the feature extractor. The length
+            A tuple of integers defining the stride of each 1D convolutional layer in the feature encoder. The length
             of *conv_stride* defines the number of convolutional layers and has to match the length of *conv_dim*.
         conv_kernel (`Tuple[int]`, *optional*, defaults to `(10, 3, 3, 3, 3, 3, 3)`):
-            A tuple of integers defining the kernel size of each 1D convolutional layer in the feature extractor. The
+            A tuple of integers defining the kernel size of each 1D convolutional layer in the feature encoder. The
             length of *conv_kernel* defines the number of convolutional layers and has to match the length of
             *conv_dim*.
         conv_bias (`bool`, *optional*, defaults to `False`):
@@ -97,7 +97,7 @@ class Wav2Vec2Config(PretrainedConfig):
             True` corresponds to applying layer norm before the attention layer, whereas `do_stable_layer_norm is
             False` corresponds to applying layer norm after the attention layer.
         apply_spec_augment (`bool`, *optional*, defaults to `True`):
-            Whether to apply *SpecAugment* data augmentation to the outputs of the feature extractor. For reference see
+            Whether to apply *SpecAugment* data augmentation to the outputs of the feature encoder. For reference see
             [SpecAugment: A Simple Data Augmentation Method for Automatic Speech
             Recognition](https://arxiv.org/abs/1904.08779).
         mask_time_prob (`float`, *optional*, defaults to 0.05):
@@ -132,7 +132,7 @@ class Wav2Vec2Config(PretrainedConfig):
         contrastive_logits_temperature (`float`, *optional*, defaults to 0.1):
             The temperature *kappa* in the contrastive loss.
         feat_quantizer_dropout (`float`, *optional*, defaults to 0.0):
-            The dropout probabilitiy for the output of the feature extractor that's used by the quantizer.
+            The dropout probabilitiy for the output of the feature encoder that's used by the quantizer.
         num_negatives (`int`, *optional*, defaults to 100):
             Number of negative samples for the contrastive loss.
         codevector_dim (`int`, *optional*, defaults to 256):

src/transformers/models/wav2vec2/modeling_flax_wav2vec2.py

@@ -395,7 +395,7 @@ class FlaxConvLayersCollection(nn.Module):
         return hidden_states
 
 
-class FlaxWav2Vec2FeatureExtractor(nn.Module):
+class FlaxWav2Vec2FeatureEncoder(nn.Module):
     """Construct the features from raw audio waveform"""
 
     config: Wav2Vec2Config
@@ -849,7 +849,7 @@ class FlaxWav2Vec2Module(nn.Module):
     dtype: jnp.dtype = jnp.float32
 
     def setup(self):
-        self.feature_extractor = FlaxWav2Vec2FeatureExtractor(self.config, dtype=self.dtype)
+        self.feature_extractor = FlaxWav2Vec2FeatureEncoder(self.config, dtype=self.dtype)
         self.feature_projection = FlaxWav2Vec2FeatureProjection(self.config, dtype=self.dtype)
         self.masked_spec_embed = self.param(
             "masked_spec_embed", jax.nn.initializers.uniform(), (self.config.hidden_size,)

src/transformers/models/wav2vec2/modeling_tf_wav2vec2.py

@@ -655,7 +655,7 @@ class TFWav2Vec2SamePadLayer(tf.keras.layers.Layer):
         return hidden_states
 
 
-class TFWav2Vec2FeatureExtractor(tf.keras.layers.Layer):
+class TFWav2Vec2FeatureEncoder(tf.keras.layers.Layer):
     def __init__(self, config: Wav2Vec2Config, **kwargs: Any) -> None:
         super().__init__(**kwargs)
 
@@ -682,6 +682,17 @@ class TFWav2Vec2FeatureExtractor(tf.keras.layers.Layer):
         return hidden_states
 
 
+class TFWav2Vec2FeatureExtractor(TFWav2Vec2FeatureEncoder):
+    def __init__(self, config, **kwargs):
+        super().__init__(config, **kwargs)
+        warnings.warn(
+            f"The class `{self.__class__.__name__}` has been depreciated "
+            "and will be removed in Transformers v5. "
+            f"Use `{self.__class__.__bases__[0].__name__}` instead.",
+            FutureWarning,
+        )
+
+
 class TFWav2Vec2FeatureProjection(tf.keras.layers.Layer):
     def __init__(self, config: Wav2Vec2Config, **kwargs):
         super().__init__(**kwargs)
@@ -1107,7 +1118,7 @@ class TFWav2Vec2MainLayer(tf.keras.layers.Layer):
     def __init__(self, config: Wav2Vec2Config, **kwargs):
         super().__init__(**kwargs)
         self.config = config
-        self.feature_extractor = TFWav2Vec2FeatureExtractor(config, name="feature_extractor")
+        self.feature_extractor = TFWav2Vec2FeatureEncoder(config, name="feature_extractor")
         self.feature_projection = TFWav2Vec2FeatureProjection(config, name="feature_projection")
 
         if config.do_stable_layer_norm:
@@ -1481,8 +1492,20 @@ class TFWav2Vec2ForCTC(TFWav2Vec2PreTrainedModel):
 
     def freeze_feature_extractor(self):
         """
-        Calling this function will disable the gradient computation for the feature extractor so that its parameter
-        will not be updated during training.
+        Calling this function will disable the gradient computation for the feature encoder so that its parameters will
+        not be updated during training.
+        """
+        warnings.warn(
+            "The method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5."
+            "Please use the equivalent `freeze_feature_encoder` method instead.",
+            FutureWarning,
+        )
+        self.freeze_feature_encoder()
+
+    def freeze_feature_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the feature encoder so that its parameter will
+        not be updated during training.
         """
         self.wav2vec2.feature_extractor.trainable = False