[CLAP] Add CLAP to the library (#21370)

* add model like clip

* update

* text model ok

* clap text works

* some refactor

- `CLAPVision` to `CLAPAudio`
- refactor kwargs of audio modules

* more refactor

* more refactor

* more refactor

* correct fusion

* more refactor

* new modules

* add basic processor

* fixup

* remove whisper copioed from

* audio logits match

* add doc

* correct filters mel and add maxlength

* style

* few fixes

* forward passes

* fixup

* fixup

* some clean up

* remove mels form the dictionnary

* pad after the repeat

* update padding when dsmaller

* fix padding

* style

* use swin patch merging

* use copied from swin

* processor with any tokenizer

* more copied from

* some clean up

* more refactor

* fix mel when rand_trunc

* style

* remove unused imports

* update processing

* remove image processing tests

* add testing fiel

* fixmodeling issues

* replace with `is_longer`

* clap in serialization

* more refactor

* `make fixup`

* make fixup

* fix feature extractor

* update test feature extractor

* `make fixup`

* clean up config

* more clean up

* more cleanup

* update tests

* refactor tests and inits

* removeCLAP vision config

* remove CLAP from image procssing auto and dummy vision objects

* update inits

* style

* re order classes in modeling clap

* Use roberta tokenizer as the other weights are not open sourced

* small cleaup

* remove tokenization CLAP

* processor tokenizr is roberta

* update feature extraction doc

* remove vclap from model zero shot

* update f_min and f_max to frequency_xx

* some changes

- fix modeling keys
- add `is_longer` in the forward pass
- make fixup

* make fixup

* consistent behavior ebtween rand_crop and fusion

* add numpy resize and bilinear and documentation

* move resizing to image utils

* clean feature extraction

* import resize from correct file

* resize in image transforms

* update

* style

* style

* nit

* remove unused arguments form the feature extractor

* style

* few fixes + make fixup

* oops

* fix more tests

* add zero shot audio classification pipeline

* update zeroshot classification pipeline

* fixup

* fix copies

* all CI tests pass

* make fixup + fix docs

* fix docs

* fix docs

* update tests pip;eline

* update zero shot pipeline

* update feature extraction clap

* update tokenization auto

* use nested simplify

* update pipeline tests

* Apply suggestions from code review
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* split in two lines

* fixes

* refactor

* clean up

* add integration tests

* update config docstring

* style

* update processor

* fix processor test

* fix feat extractor tests

* update docs

* Apply suggestions from code review
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* fix readmes

* fix tips

* Update src/transformers/models/auto/configuration_auto.py

* update doc and remove todo -> properly explained

* fix idx and typo

* typoe

* cleanup config

* cleanup tests, styles and doc

* ignore docstyle on image transform

* add conversion script

* remove the `clap` indx in favor of `CLAP`

* update __init

* nits

* Update src/transformers/pipelines/__init__.py

* fix bug

* clarifiy config

* fix copy

* fix init

* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* fix model output

* fix comment

* make fixup

* make fixup

* rename to `Clap`

* replace to `Clap`

* replace to `Clap`

* repo consistency

* again repo-consistency

* make fixup

* Apply suggestions from code review
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* add config

* changes

* update conversion

* Apply suggestions from code review
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* remove unused function

* update based on code reviews

* style

* more comments

* cleanup

* clean up

* style

* apply suggestions

* Empty commit

* pipeline will be added in a different PR

* update calls to audio utils functions

* update pipeline init

* style

* style

* styling again

* use pad

* fix repo-consistency

* update utils and add doc for audio utils

* clean up resize by using torch. update inits accordingly

* style

* CLap's  tokenizer is RobertA

* add audio utils to internal toctreee

* update totctree

* style

* update documentation and normalize naming accross audio utils and feature extraction clap

* style

* clean up

* update doc and typos

* fix doctest

* update modelin code, got rid of a lot of reshaping

* style on added doc audio utils

* update modeling clap

* style

* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* docstringvariables with CLAP

* rename key

* update modeling CLAP

* update audio utils docstring

* update processing clap

* fix readmes

* fix toctree

* udpate configuration clap

* fix init

* make fixup

* fix

* fix

* update naming

* update

* update checkpoint path

* Apply suggestions from code review

* Major refactoring

* Update src/transformers/models/clap/configuration_clap.py

* merge

---------
Co-authored-by: younesbelkada <younesbelkada@gmail.com>
Co-authored-by: Younes Belkada <49240599+younesbelkada@users.noreply.github.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

src/transformers/models/clap/__init__.py

+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
+
+
+_import_structure = {
+    "configuration_clap": [
+        "CLAP_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "ClapAudioConfig",
+        "ClapConfig",
+        "ClapTextConfig",
+    ],
+    "processing_clap": ["ClapProcessor"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_clap"] = [
+        "CLAP_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "ClapModel",
+        "ClapPreTrainedModel",
+        "ClapTextModel",
+        "ClapTextModelWithProjection",
+        "ClapAudioModel",
+        "ClapAudioModelWithProjection",
+    ]
+    _import_structure["feature_extraction_clap"] = ["ClapFeatureExtractor"]
+
+if TYPE_CHECKING:
+    from .configuration_clap import (
+        CLAP_PRETRAINED_MODEL_ARCHIVE_LIST,
+        ClapAudioConfig,
+        ClapConfig,
+        ClapTextConfig,
+    )
+    from .processing_clap import ClapProcessor
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .feature_extraction_clap import ClapFeatureExtractor
+        from .modeling_clap import (
+            CLAP_PRETRAINED_MODEL_ARCHIVE_LIST,
+            ClapAudioModel,
+            ClapAudioModelWithProjection,
+            ClapModel,
+            ClapPreTrainedModel,
+            ClapTextModel,
+            ClapTextModelWithProjection,
+        )
+
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/clap/configuration_clap.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" CLAP model configuration"""
+
+import copy
+import os
+from typing import Union
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+CLAP_PRETRAINED_MODEL_ARCHIVE_LIST = {
+    "laion/clap-htsat-fused": "https://huggingface.co/laion/clap-htsat-fused/resolve/main/config.json",
+    "laion/clap-htsat-unfused": "https://huggingface.co/laion/clap-htsat-unfused/resolve/main/config.json",
+}
+
+
+class ClapTextConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`ClapTextModel`]. It is used to instantiate a CLAP
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the CLAP
+    [calp-hsat-fused](https://huggingface.co/laion/clap-hsat-fused) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 30522):
+            Vocabulary size of the CLAP model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`ClapTextModel`].
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        hidden_act (`str` or `Callable`, *optional*, defaults to `"relu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"relu"`,
+            `"relu"`, `"silu"` and `"relu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (`int`, *optional*, defaults to 512):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        type_vocab_size (`int`, *optional*, defaults to 2):
+            The vocabulary size of the `token_type_ids` passed when calling [`ClapTextModel`].
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        position_embedding_type (`str`, *optional*, defaults to `"absolute"`):
+            Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`. For
+            positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to
+            [Self-Attention with Relative Position Representations (Shaw et al.)](https://arxiv.org/abs/1803.02155).
+            For more information on `"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).
+        is_decoder (`bool`, *optional*, defaults to `False`):
+            Whether the model is used as a decoder or not. If `False`, the model is used as an encoder.
+        use_cache (`bool`, *optional*, defaults to `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 (`float`, *optional*):
+            The dropout ratio for the classification head.
+        projection_hidden_act (`str`, *optional*, defaults to `"relu"`):
+            The non-linear activation function (function or string) in the projection layer. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        projection_dim (`int`, *optional*, defaults to 512)
+            Dimension of the projection head of the `ClapTextModelWithProjection`.
+
+    Examples:
+
+    ```python
+    >>> from transformers import ClapTextConfig, ClapTextModel
+
+    >>> # Initializing a CLAP text configuration
+    >>> configuration = ClapTextConfig()
+
+    >>> # Initializing a model (with random weights) from the configuration
+    >>> model = ClapTextModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "clap_text_model"
+
+    def __init__(
+        self,
+        vocab_size=50265,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=514,
+        type_vocab_size=1,
+        initializer_range=0.02,
+        initializer_factor=1.0,
+        layer_norm_eps=1e-12,
+        projection_dim=512,
+        pad_token_id=1,
+        bos_token_id=0,
+        eos_token_id=2,
+        position_embedding_type="absolute",
+        use_cache=True,
+        classifier_dropout=None,
+        projection_hidden_act="relu",
+        **kwargs,
+    ):
+        super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.initializer_range = initializer_range
+        self.initializer_factor = initializer_factor
+        self.layer_norm_eps = layer_norm_eps
+        self.position_embedding_type = position_embedding_type
+        self.use_cache = use_cache
+        self.classifier_dropout = classifier_dropout
+        self.projection_hidden_act = projection_hidden_act
+        self.projection_dim = projection_dim
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the text config dict if we are loading from ClapConfig
+        if config_dict.get("model_type") == "clap":
+            config_dict = config_dict["text_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class ClapAudioConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`ClapAudioModel`]. It is used to instantiate a
+    CLAP audio encoder according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the audio encoder of the CLAP
+    [laion/clap-htsat-fused](https://huggingface.co/laion/clap-htsat-fused) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        window_size (`int`, *optional*, defaults to 8):
+            Image size of the spectrogram
+        num_mel_bins (`int`, *optional*, defaults to 64):
+            Number of mel features used per frames. Should correspond to the value used in the `ClapProcessor` class.
+        spec_size (`int`, *optional*, defaults to 256):
+            Desired input size of the spectrogram that the model supports. It can be different from the output of the
+            `ClapFeatureExtractor`, in which case the input features will be resized. Corresponds to the `image_size`
+            of the audio models.
+        hidden_act (`str`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        patch_size (`int`, *optional*, defaults to 4):
+            Patch size for the audio spectrogram
+        patch_stride (`list`, *optional*, defaults to `[4, 4]`):
+            Patch stride for the audio spectrogram
+        num_classes (`int`, *optional*, defaults to 527):
+            Number of classes used for the head training
+        hidden_size (`int`, *optional*, defaults to 768):
+            Hidden size of the output of the audio encoder. Correspond to the dimension of the penultimate layer's
+            output,which is sent to the projection MLP layer.
+        projection_dim (`int`, *optional*, defaults to 512):
+            Hidden size of the projection layer.
+        depths (`list`, *optional*, defaults to `[2, 2, 6, 2]`):
+            Depths used for the Swin Layers of the audio model
+        num_attention_heads (`list`, *optional*, defaults to `[4, 8, 16, 32]`):
+            Number of attention heads used for the Swin Layers of the audio model
+        enable_fusion (`bool`, *optional*, defaults to `False`):
+            Whether or not to enable patch fusion. This is the main contribution of the authors, and should give the
+            best results.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probabilitiy for all fully connected layers in the encoder.
+        fusion_type (`[type]`, *optional*):
+            Fusion type used for the patch fusion.
+        patch_embed_input_channels (`int`, *optional*, defaults to 1):
+            Number of channels used for the input spectrogram
+        flatten_patch_embeds (`bool`, *optional*, defaults to `True`):
+            Whether or not to flatten the patch embeddings
+        patch_embeds_hidden_size (`int`, *optional*, defaults to 96):
+            Hidden size of the patch embeddings. It is used as the number of output channels.
+        enable_patch_layer_norm (`bool`, *optional*, defaults to `True`):
+            Whether or not to enable layer normalization for the patch embeddings
+        drop_path_rate (`float`, *optional*, defaults to 0.0):
+            Drop path rate for the patch fusion
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        qkv_bias (`bool`, *optional*, defaults to `True`):
+            Whether or not to add a bias to the query, key, value projections.
+        mlp_ratio (`float`, *optional*, defaults to 4.0):
+            Ratio of the mlp hidden dim to embedding dim.
+        aff_block_r (`int`, *optional*, defaults to 4):
+            downsize_ratio used in the AudioFF block
+        num_hidden_layers (`int`, *optional*, defaults to 4):
+            Number of hidden layers in the Transformer encoder.
+        projection_hidden_act (`str`, *optional*, defaults to `"relu"`):
+            The non-linear activation function (function or string) in the projection layer. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        layer_norm_eps (`[type]`, *optional*, defaults to `1e-5`):
+            The epsilon used by the layer normalization layers.
+        initializer_factor (`float`, *optional*, defaults to 1.0):
+            A factor for initializing all weight matrices (should be kept to 1, used internally for initialization
+            testing).
+
+    Example:
+
+    ```python
+    >>> from transformers import ClapAudioConfig, ClapAudioModel
+
+    >>> # Initializing a ClapAudioConfig with laion/clap-htsat-fused style configuration
+    >>> configuration = ClapAudioConfig()
+
+    >>> # Initializing a ClapAudioModel (with random weights) from the laion/clap-htsat-fused style configuration
+    >>> model = ClapAudioModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "clap_audio_model"
+
+    def __init__(
+        self,
+        window_size=8,
+        num_mel_bins=64,
+        spec_size=256,
+        hidden_act="gelu",
+        patch_size=4,
+        patch_stride=[4, 4],
+        num_classes=527,
+        hidden_size=768,
+        projection_dim=512,
+        depths=[2, 2, 6, 2],
+        num_attention_heads=[4, 8, 16, 32],
+        enable_fusion=False,
+        hidden_dropout_prob=0.1,
+        fusion_type=None,
+        patch_embed_input_channels=1,
+        flatten_patch_embeds=True,
+        patch_embeds_hidden_size=96,
+        enable_patch_layer_norm=True,
+        drop_path_rate=0.0,
+        attention_probs_dropout_prob=0.0,
+        qkv_bias=True,
+        mlp_ratio=4.0,
+        aff_block_r=4,
+        num_hidden_layers=4,
+        projection_hidden_act="relu",
+        layer_norm_eps=1e-5,
+        initializer_factor=1.0,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.window_size = window_size
+        self.num_mel_bins = num_mel_bins
+        self.spec_size = spec_size
+        self.patch_size = patch_size
+        self.patch_stride = patch_stride
+        self.num_classes = num_classes
+        self.hidden_size = hidden_size
+        self.depths = depths
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.window_size = window_size
+        self.enable_fusion = enable_fusion
+        self.fusion_type = fusion_type
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.projection_dim = projection_dim
+        self.flatten_patch_embeds = flatten_patch_embeds
+        self.patch_embeds_hidden_size = patch_embeds_hidden_size
+        self.enable_patch_layer_norm = enable_patch_layer_norm
+        self.drop_path_rate = drop_path_rate
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.qkv_bias = qkv_bias
+        self.mlp_ratio = mlp_ratio
+        self.patch_embed_input_channels = patch_embed_input_channels
+        self.aff_block_r = aff_block_r
+        self.layer_norm_eps = layer_norm_eps
+        self.initializer_factor = initializer_factor
+        self.projection_hidden_act = projection_hidden_act
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the audio config dict if we are loading from ClapConfig
+        if config_dict.get("model_type") == "clap":
+            config_dict = config_dict["audio_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class ClapConfig(PretrainedConfig):
+    r"""
+    [`ClapConfig`] is the configuration class to store the configuration of a [`ClapModel`]. It is used to instantiate
+    a CLAP model according to the specified arguments, defining the text model and audio model configs. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the CLAP
+    [laion/clap-htsat-fused](https://huggingface.co/laion/clap-htsat-fused) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        text_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`ClapTextConfig`].
+        audio_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`ClapAudioConfig`].
+        projection_dim (`int`, *optional*, defaults to 512):
+            Dimentionality of text and audio projection layers.
+        logit_scale_init_value (`float`, *optional*, defaults to 2.6592):
+            The inital value of the *logit_scale* paramter. Default is used as per the original CLAP implementation.
+        projection_hidden_act (`str`, *optional*, defaults to `"relu"`):
+            Activation function for the projection layers.
+        initializer_factor (`float`, *optional*, defaults to 1.0):
+            Factor to scale the initialization of the model weights.
+        kwargs (*optional*):
+            Dictionary of keyword arguments.
+
+    Example:
+
+    ```python
+    >>> from transformers import ClapConfig, ClapModel
+
+    >>> # Initializing a ClapConfig with laion-ai/base style configuration
+    >>> configuration = ClapConfig()
+
+    >>> # Initializing a ClapModel (with random weights) from the laion-ai/base style configuration
+    >>> model = ClapModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+
+    >>> # We can also initialize a ClapConfig from a ClapTextConfig and a ClapAudioConfig
+    >>> from transformers import ClapTextConfig, ClapAudioConfig
+
+    >>> # Initializing a ClapText and ClapAudioConfig configuration
+    >>> config_text = ClapTextConfig()
+    >>> config_audio = ClapAudioConfig()
+
+    >>> config = ClapConfig.from_text_audio_configs(config_text, config_audio)
+    ```"""
+
+    model_type = "clap"
+    is_composition = True
+
+    def __init__(
+        self,
+        text_config=None,
+        audio_config=None,
+        logit_scale_init_value=(1 / 0.07),
+        projection_dim=512,
+        projection_hidden_act="relu",
+        initializer_factor=1.0,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        if text_config is None:
+            text_config = {}
+            logger.info("text_config is None. Initializing the ClapTextConfig with default values.")
+
+        if audio_config is None:
+            audio_config = {}
+            logger.info("audio_config is None. initializing the ClapAudioConfig with default values.")
+
+        self.text_config = ClapTextConfig(**text_config)
+        self.audio_config = ClapAudioConfig(**audio_config)
+        self.text_config.projection_dim = projection_dim
+        self.audio_config.projection_dim = projection_dim
+
+        self.text_config.projection_hidden_act = projection_hidden_act
+        self.audio_config.projection_hidden_act = projection_hidden_act
+
+        self.projection_dim = projection_dim
+        self.projection_hidden_act = projection_hidden_act
+        self.hidden_size = self.text_config.hidden_size
+
+        self.logit_scale_init_value = logit_scale_init_value
+        self.initializer_factor = initializer_factor
+        self.num_hidden_layers = self.text_config.num_hidden_layers + len(self.audio_config.depths)
+
+    @classmethod
+    def from_text_audio_configs(cls, text_config: ClapTextConfig, audio_config: ClapAudioConfig, **kwargs):
+        r"""
+        Instantiate a [`ClapConfig`] (or a derived class) from clap text model configuration and clap audio model
+        configuration.
+
+        Returns:
+            [`ClapConfig`]: An instance of a configuration object
+        """
+
+        return cls(text_config=text_config.to_dict(), audio_config=audio_config.to_dict(), **kwargs)
+
+    def to_dict(self):
+        """
+        Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
+
+        Returns:
+            `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        output = copy.deepcopy(self.__dict__)
+        output["text_config"] = self.text_config.to_dict()
+        output["audio_config"] = self.audio_config.to_dict()
+        output["model_type"] = self.__class__.model_type
+        return output

src/transformers/models/clap/convert_clap_original_pytorch_to_hf.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import re
+
+import torch
+from CLAP import create_model
+
+from transformers import AutoFeatureExtractor, ClapConfig, ClapModel
+
+
+KEYS_TO_MODIFY_MAPPING = {
+    "text_branch": "text_model",
+    "audio_branch": "audio_model.audio_encoder",
+    "attn": "attention.self",
+    "self.proj": "output.dense",
+    "attention.self_mask": "attn_mask",
+    "mlp.fc1": "intermediate.dense",
+    "mlp.fc2": "output.dense",
+    "norm1": "layernorm_before",
+    "norm2": "layernorm_after",
+    "bn0": "batch_norm",
+}
+
+processor = AutoFeatureExtractor.from_pretrained("laion/clap-htsat-unfused", truncation="rand_trunc")
+
+
+def init_clap(checkpoint_path, enable_fusion=False):
+    model, model_cfg = create_model(
+        "HTSAT-tiny",
+        "roberta",
+        checkpoint_path,
+        precision="fp32",
+        device="cuda:0" if torch.cuda.is_available() else "cpu",
+        enable_fusion=enable_fusion,
+        fusion_type="aff_2d" if enable_fusion else None,
+    )
+    return model, model_cfg
+
+
+def rename_state_dict(state_dict):
+    model_state_dict = {}
+
+    sequential_layers_pattern = r".*sequential.(\d+).*"
+    text_projection_pattern = r".*_projection.(\d+).*"
+
+    for key, value in state_dict.items():
+        # check if any key needs to be modified
+        for key_to_modify, new_key in KEYS_TO_MODIFY_MAPPING.items():
+            if key_to_modify in key:
+                key = key.replace(key_to_modify, new_key)
+
+        if re.match(sequential_layers_pattern, key):
+            # replace sequential layers with list
+            sequential_layer = re.match(sequential_layers_pattern, key).group(1)
+
+            key = key.replace(f"sequential.{sequential_layer}.", f"layers.{int(sequential_layer)//3}.linear.")
+        elif re.match(text_projection_pattern, key):
+            projecton_layer = int(re.match(text_projection_pattern, key).group(1))
+
+            # Because in CLAP they use `nn.Sequential`...
+            transformers_projection_layer = 1 if projecton_layer == 0 else 2
+
+            key = key.replace(f"_projection.{projecton_layer}.", f"_projection.linear{transformers_projection_layer}.")
+
+        if "audio" and "qkv" in key:
+            # split qkv into query key and value
+            mixed_qkv = value
+            qkv_dim = mixed_qkv.size(0) // 3
+
+            query_layer = mixed_qkv[:qkv_dim]
+            key_layer = mixed_qkv[qkv_dim : qkv_dim * 2]
+            value_layer = mixed_qkv[qkv_dim * 2 :]
+
+            model_state_dict[key.replace("qkv", "query")] = query_layer
+            model_state_dict[key.replace("qkv", "key")] = key_layer
+            model_state_dict[key.replace("qkv", "value")] = value_layer
+        else:
+            model_state_dict[key] = value
+
+    return model_state_dict
+
+
+def convert_clap_checkpoint(checkpoint_path, pytorch_dump_folder_path, config_path, enable_fusion=False):
+    clap_model, clap_model_cfg = init_clap(checkpoint_path, enable_fusion=enable_fusion)
+
+    clap_model.eval()
+    state_dict = clap_model.state_dict()
+    state_dict = rename_state_dict(state_dict)
+
+    transformers_config = ClapConfig()
+    transformers_config.audio_config.enable_fusion = enable_fusion
+    model = ClapModel(transformers_config)
+
+    # ignore the spectrogram embedding layer
+    model.load_state_dict(state_dict, strict=False)
+
+    model.save_pretrained(pytorch_dump_folder_path)
+    transformers_config.save_pretrained(pytorch_dump_folder_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model.")
+    parser.add_argument("--checkpoint_path", default=None, type=str, help="Path to fairseq checkpoint")
+    parser.add_argument("--config_path", default=None, type=str, help="Path to hf config.json of model to convert")
+    parser.add_argument("--enable_fusion", action="store_true", help="Whether to enable fusion or not")
+    args = parser.parse_args()
+
+    convert_clap_checkpoint(args.checkpoint_path, args.pytorch_dump_folder_path, args.config_path, args.enable_fusion)

src/transformers/models/clap/feature_extraction_clap.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Feature extractor class for CLAP."""
+
+
+import copy
+from typing import Any, Dict, List, Optional, Union
+
+import numpy as np
+import torch
+
+from ...audio_utils import fram_wave, get_mel_filter_banks, power_to_db, stft
+from ...feature_extraction_sequence_utils import SequenceFeatureExtractor
+from ...feature_extraction_utils import BatchFeature
+from ...utils import TensorType, logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class ClapFeatureExtractor(SequenceFeatureExtractor):
+    r"""
+    Constructs a CLAP feature extractor.
+
+    This feature extractor inherits from [`~feature_extraction_sequence_utils.SequenceFeatureExtractor`] which contains
+    most of the main methods. Users should refer to this superclass for more information regarding those methods.
+
+    This class extracts mel-filter bank features from raw speech using a custom numpy implementation of the *Short Time
+    Fourier Transform* (STFT) which should match pytorch's `torch.stft` equivalent.
+
+    Args:
+        feature_size (`int`, defaults to 64):
+            The feature dimension of the extracted Mel spectrograms. This corresponds to the number of mel filters
+            (`n_mels`).
+        sampling_rate (`int`, defaults to 48_000):
+            The sampling rate at which the audio files should be digitalized expressed in hertz (Hz). This only serves
+            to warn users if the audio fed to the feature extractor does not have the same sampling rate.
+        hop_length (`int`, defaults to 480):
+            Length of the overlaping windows for the STFT used to obtain the Mel Spectrogram. The audio will be split
+            in smaller `frames` with a step of `hop_length` between each frame.
+        max_length_s (`int`, defaults to 10):
+            The maximum input lenght of the model in seconds. This is used to pad the audio.
+        fft_window_size (`int`, defaults to 1024):
+            Size of the window (in samples) on which the Fourier transform is applied. This controls the frequency
+            resolution of the spectrogram. 400 means that the fourrier transform is computed on windows of 400 samples.
+        padding_value (`float`, *optional*, defaults to 0.0):
+            Padding value used to pad the audio. Should correspond to silences.
+        return_attention_mask (`bool`, *optional*, defaults to `False`):
+            Whether or not the model should return the attention masks coresponding to the input.
+        frequency_min (`float`, *optional*, default to 0):
+            The lowest frequency of interest. The STFT will not be computed for values below this.
+        frequency_max (`float`, *optional*, default to 14_000):
+            The highest frequency of interest. The STFT will not be computed for values above this.
+        top_db (`float`, *optional*):
+            The highest decibel value used to convert the mel spectrogram to the log scale. For more details see the
+            `audio_utils.power_to_db` function
+        truncation (`str`, *optional*, default to `"fusions"`):
+            Truncation pattern for long audio inputs. Two patterns are available:
+                - `fusion` will use `_random_mel_fusion`, which stacks 3 random crops from the mel spectrogram and a
+                  downsampled version of the entire mel spectrogram.
+            If `config.fusion` is set to True, shorter audios also need to to return 4 mels, which will just be a copy
+            of the original mel obtained from the padded audio.
+                - `rand_trunc` will select a random crop of the mel spectrogram.
+        padding (`str`, *optional*, defaults to `"repeatpad"`):
+               Padding pattern for shorter audio inputs. Three patterns were originally implemented:
+                - `repeatpad`: the audio is repeated, and then padded to fit the `max_length`.
+                - `repeat`: the audio is repeated and then cut to fit the `max_length`
+                - `pad`: the audio is padded.
+    """
+
+    model_input_names = ["input_features", "is_longer"]
+
+    def __init__(
+        self,
+        feature_size=64,
+        sampling_rate=48_000,
+        hop_length=480,
+        max_length_s=10,
+        fft_window_size=1024,
+        padding_value=0.0,
+        return_attention_mask=False,  # pad inputs to max length with silence token (zero) and no attention mask
+        frequency_min: float = 0,
+        frequency_max: float = 14_000,
+        top_db: int = None,
+        truncation: str = "fusion",
+        padding: str = "repeatpad",
+        **kwargs,
+    ):
+        super().__init__(
+            feature_size=feature_size,
+            sampling_rate=sampling_rate,
+            padding_value=padding_value,
+            return_attention_mask=return_attention_mask,
+            **kwargs,
+        )
+        self.top_db = top_db
+        self.truncation = truncation
+        self.padding = padding
+        self.fft_window_size = fft_window_size
+        self.nb_frequency_bins = (fft_window_size >> 1) + 1
+        self.hop_length = hop_length
+        self.max_length_s = max_length_s
+        self.nb_max_samples = max_length_s * sampling_rate
+        self.sampling_rate = sampling_rate
+        self.frequency_min = frequency_min
+        self.frequency_max = frequency_max
+        self.mel_filters = get_mel_filter_banks(
+            nb_frequency_bins=self.nb_frequency_bins,
+            nb_mel_filters=feature_size,
+            frequency_min=frequency_min,
+            frequency_max=frequency_max,
+            sample_rate=sampling_rate,
+            norm=None,
+            mel_scale="htk",
+        )
+        self.mel_filters_slaney = get_mel_filter_banks(
+            nb_frequency_bins=self.nb_frequency_bins,
+            nb_mel_filters=feature_size,
+            frequency_min=frequency_min,
+            frequency_max=frequency_max,
+            sample_rate=sampling_rate,
+            norm="slaney",
+            mel_scale="slaney",
+        )
+
+    def to_dict(self) -> Dict[str, Any]:
+        """
+        Serializes this instance to a Python dictionary.
+
+        Returns:
+            `Dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance, excpet for the
+            mel filter banks, which do not need to be saved or printed as they are too long.
+        """
+        output = copy.deepcopy(self.__dict__)
+        output["feature_extractor_type"] = self.__class__.__name__
+        if "mel_filters" in output:
+            del output["mel_filters"]
+        if "mel_filters_slaney" in output:
+            del output["mel_filters_slaney"]
+        return output
+
+    def _np_extract_fbank_features(self, waveform: np.array, mel_filters: Optional[np.array] = None) -> np.ndarray:
+        """
+        Compute the log-Mel spectrogram of the provided `waveform` using the `hanning` window. In CLAP, two different
+        filter banks are used depending on the truncation pattern:
+            - `self.mel_filters`: they correspond to the defaults parameters of `torchaduio` which can be obtained from
+              calling `torchaudio.transforms.MelSpectrogram().mel_scale.fb`. These filters are used when `truncation`
+              is set to `"fusion"`.
+            - `self.mel_filteres_slaney` : they correspond to the defaults parameters of `torchlibrosa` which used
+              `librosa.filters.mel` when computing the mel spectrogram. These filters were only used in the original
+              implementation when the truncation mode is not `"fusion"`.
+        """
+        window = np.hanning(self.fft_window_size + 1)[:-1]
+        frames = fram_wave(waveform, self.hop_length, self.fft_window_size)
+        spectrogram = stft(frames, window, fft_window_size=self.fft_window_size)
+
+        magnitudes = np.abs(spectrogram) ** 2
+        mel_spectrogram = np.matmul(mel_filters.T, magnitudes)
+        log_mel_spectrogram = power_to_db(mel_spectrogram).T
+        log_mel_spectrogram = np.asarray(log_mel_spectrogram, np.float32)
+        return log_mel_spectrogram
+
+    def _random_mel_fusion(self, mel, total_frames, chunk_frames):
+        ranges = np.array_split(list(range(0, total_frames - chunk_frames + 1)), 3)
+        if len(ranges[1]) == 0:
+            # if the audio is too short, we just use the first chunk
+            ranges[1] = [0]
+        if len(ranges[2]) == 0:
+            # if the audio is too short, we just use the first chunk
+            ranges[2] = [0]
+        # randomly choose index for each part
+        idx_front = np.random.choice(ranges[0])
+        idx_middle = np.random.choice(ranges[1])
+        idx_back = np.random.choice(ranges[2])
+
+        mel_chunk_front = mel[idx_front : idx_front + chunk_frames, :]
+        mel_chunk_middle = mel[idx_middle : idx_middle + chunk_frames, :]
+        mel_chunk_back = mel[idx_back : idx_back + chunk_frames, :]
+
+        mel = torch.tensor(mel[None, None, :])
+        mel_shrink = torch.nn.functional.interpolate(
+            mel, size=[chunk_frames, 64], mode="bilinear", align_corners=False, antialias=False
+        )
+        mel_shrink = mel_shrink[0][0].numpy()
+        mel_fusion = np.stack([mel_chunk_front, mel_chunk_middle, mel_chunk_back, mel_shrink], axis=0)
+        return mel_fusion
+
+    def _get_input_mel(self, waveform: np.array, max_length, truncation, padding) -> np.array:
+        """
+        Extracts the mel spectrogram and prepares it for the mode based on the `truncation` and `padding` arguments.
+        Four different path are possible:
+            - `truncation="fusion"` and the length of the waveform is greater than the max length: the mel spectrogram
+              will be computed on the entire audio. 3 random crops and a dowsampled version of the full mel spectrogram
+              are then stacked together. They will later be used for `feature_fusion`.
+            - `truncation="rand_trunc"` and the length of the waveform is smaller than the max length: the audio is
+              padded based on `padding`.
+            - `truncation="fusion"` and the length of the waveform is smaller than the max length: the audio is padded
+              based on `padding`, and is repeated `4` times.
+            - `truncation="rand_trunc"` and the length of the waveform is greater than the max length: the mel
+              spectrogram will be computed on a random crop of the waveform.
+
+        """
+        if waveform.shape[0] > max_length:
+            if truncation == "rand_trunc":
+                longer = True
+                # random crop to max_length (for compatibility) -> this should be handled by self.pad
+                overflow = len(waveform) - max_length
+                idx = np.random.randint(0, overflow + 1)
+                waveform = waveform[idx : idx + max_length]
+                input_mel = self._np_extract_fbank_features(waveform, self.mel_filters_slaney)[None, :]
+            elif truncation == "fusion":
+                mel = self._np_extract_fbank_features(waveform, self.mel_filters)
+                chunk_frames = max_length // self.hop_length + 1  # the +1 related to how the spectrogram is computed
+                total_frames = mel.shape[0]
+                if chunk_frames == total_frames:
+                    # there is a corner case where the audio length is larger than max_length but smaller than max_length+hop_length.
+                    # In this case, we just use the whole audio.
+                    input_mel = np.stack([mel, mel, mel, mel], axis=0)
+                    longer = False
+                else:
+                    input_mel = self._random_mel_fusion(mel, total_frames, chunk_frames)
+                    longer = True
+            else:
+                raise NotImplementedError(f"data_truncating {truncation} not implemented")
+
+        else:
+            longer = False
+            # only use repeat as a new possible value for padding. you repeat the audio before applying the usual max_length padding
+            if waveform.shape[0] < max_length:
+                if padding == "repeat":
+                    n_repeat = int(max_length / len(waveform))
+                    waveform = np.stack(np.tile(waveform, n_repeat + 1))[:max_length]
+                if padding == "repeatpad":
+                    n_repeat = int(max_length / len(waveform))
+                    waveform = np.stack(np.tile(waveform, n_repeat))
+                waveform = np.pad(waveform, (0, max_length - waveform.shape[0]), mode="constant", constant_values=0)
+
+            if truncation == "fusion":
+                input_mel = self._np_extract_fbank_features(waveform, self.mel_filters)
+                input_mel = np.stack([input_mel, input_mel, input_mel, input_mel], axis=0)
+            else:
+                input_mel = self._np_extract_fbank_features(waveform, self.mel_filters_slaney)[None, :]
+
+        return input_mel, longer
+
+    def __call__(
+        self,
+        raw_speech: Union[np.ndarray, List[float], List[np.ndarray], List[List[float]]],
+        truncation: str = None,
+        padding: Optional[str] = None,
+        max_length: Optional[int] = None,
+        sampling_rate: Optional[int] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        **kwargs,
+    ) -> BatchFeature:
+        """
+        Main method to featurize and prepare for the model one or several sequence(s).
+
+        Args:
+            raw_speech (`np.ndarray`, `List[float]`, `List[np.ndarray]`, `List[List[float]]`):
+                The sequence or batch of sequences to be padded. Each sequence can be a numpy array, a list of float
+                values, a list of numpy arrays or a list of list of float values.
+            truncation (`str`, *optional*):
+                Truncation pattern for long audio inputs. Two patterns are available:
+                    - `fusion` will use `_random_mel_fusion`, which stacks 3 random crops from the mel spectrogram and
+                      a downsampled version of the entire mel spectrogram.
+                If `config.fusion` is set to True, shorter audios also need to to return 4 mels, which will just be a
+                copy of the original mel obtained from the padded audio.
+                    - `rand_trunc` will select a random crop of the mel spectrogram.
+            padding (`str`, *optional*):
+               Padding pattern for shorter audio inputs. Three patterns were originally implemented:
+                    - `repeatpad`: the audio is repeated, and then padded to fit the `max_length`.
+                    - `repeat`: the audio is repeated and then cut to fit the `max_length`
+                    - `pad`: the audio is padded.
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors instead of list of python integers. Acceptable values are:
+
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.np.array` objects.
+                - `'np'`: Return Numpy `np.ndarray` objects.
+            sampling_rate (`int`, *optional*):
+                The sampling rate at which the `raw_speech` input was sampled. It is strongly recommended to pass
+                `sampling_rate` at the forward call to prevent silent errors and allow automatic speech recognition
+                pipeline.
+        """
+        truncation = truncation if truncation is not None else self.truncation
+        padding = padding if padding else self.padding
+
+        if sampling_rate is not None:
+            if sampling_rate != self.sampling_rate:
+                raise ValueError(
+                    f"The model corresponding to this feature extractor: {self.__class__.__name__} was trained using a"
+                    f" sampling rate of {self.sampling_rate}. Please make sure that the provided `raw_speech` input"
+                    f" was sampled with {self.sampling_rate} and not {sampling_rate}."
+                )
+        else:
+            logger.warning(
+                "It is strongly recommended to pass the `sampling_rate` argument to this function. "
+                "Failing to do so can result in silent errors that might be hard to debug."
+            )
+
+        is_batched = bool(
+            isinstance(raw_speech, (list, tuple))
+            and (isinstance(raw_speech[0], np.ndarray) or isinstance(raw_speech[0], (tuple, list)))
+        )
+
+        if is_batched:
+            raw_speech = [np.asarray(speech, dtype=np.float64) for speech in raw_speech]
+        elif not is_batched and not isinstance(raw_speech, np.ndarray):
+            raw_speech = np.asarray(raw_speech, dtype=np.float64)
+        elif isinstance(raw_speech, np.ndarray) and raw_speech.dtype is np.dtype(np.float64):
+            raw_speech = raw_speech.astype(np.float64)
+
+        # always return batch
+        if not is_batched:
+            raw_speech = [np.asarray(raw_speech)]
+
+        # convert to mel spectrogram, truncate and pad if needed.
+        padded_inputs = [
+            self._get_input_mel(waveform, max_length if max_length else self.nb_max_samples, truncation, padding)
+            for waveform in raw_speech
+        ]
+
+        input_mel = []
+        is_longer = []
+        for mel, longer in padded_inputs:
+            input_mel.append(mel)
+            is_longer.append(longer)
+
+        if truncation == "fusion" and sum(is_longer) == 0:
+            # if no audio is longer than 10s, then randomly select one audio to be longer
+            rand_idx = np.random.randint(0, len(input_mel))
+            is_longer[rand_idx] = True
+
+        if isinstance(input_mel[0], List):
+            input_mel = [np.asarray(feature, dtype=np.float64) for feature in input_mel]
+
+        input_features = {"input_features": input_mel, "is_longer": is_longer}
+        input_features = BatchFeature(input_features)
+
+        if return_tensors is not None:
+            input_features = input_features.convert_to_tensors(return_tensors)
+
+        return input_features

src/transformers/models/clap/modeling_clap.py

+# coding=utf-8
+# Copyright 2023 The LAION-AI Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch CLAP model."""
+import collections
+import math
+from dataclasses import dataclass
+from typing import Any, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ...activations import ACT2FN
+from ...modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPooling,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+)
+from ...modeling_utils import PreTrainedModel
+from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, meshgrid, prune_linear_layer
+from ...utils import (
+    ModelOutput,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_clap import ClapAudioConfig, ClapConfig, ClapTextConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "laion/clap-htsat-fused"
+
+CLAP_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "laion/clap-htsat-fused",
+    "laion/clap-htsat-unfused",
+    # See all clap models at https://huggingface.co/models?filter=clap
+]
+
+
+# Adapted from: https://github.com/LAION-AI/CLAP/blob/6ad05a971ba0622f6acee8c41993e0d02bbed639/src/open_clip/utils.py#L191
+def interpolate(hidden_states, ratio):
+    """
+    Interpolate data in time domain. This is used to compensate the resolution reduction in downsampling of a CNN.
+
+    Args:
+        hidden_states (`torch.FloatTensor` of shape (batch_size, time_length, classes_num)):
+            Input hidden states
+        ratio (`int`):
+            The ratio of the length of the output to the length of the input.
+    """
+    (batch_size, time_length, classes_num) = hidden_states.shape
+    upsampled = hidden_states[:, :, None, :].repeat(1, 1, ratio, 1)
+    upsampled = upsampled.reshape(batch_size, time_length * ratio, classes_num)
+    return upsampled
+
+
+# Adapted from https://github.com/LAION-AI/CLAP/blob/6ad05a971ba0622f6acee8c41993e0d02bbed639/src/open_clip/htsat.py#L249
+def window_partition(hidden_states, window_size):
+    """
+    Returns the resized hidden states. The output shape should be `(batch_size * num_windows, window_size, window_size,
+    num_channels)`
+
+    Args:
+        hidden_states (`torch.FloatTensor` of shape `(batch_size, height, width, num_channels)`):
+            Input hidden states
+        window_size (`int`):
+            Window size
+    """
+    batch_size, height, width, num_channels = hidden_states.shape
+
+    hidden_states = hidden_states.view(
+        batch_size, height // window_size, window_size, width // window_size, window_size, num_channels
+    )
+    windows = hidden_states.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, num_channels)
+    return windows
+
+
+# Adapted from https://github.com/LAION-AI/CLAP/blob/6ad05a971ba0622f6acee8c41993e0d02bbed639/src/open_clip/htsat.py#L263
+def window_reverse(windows, window_size, height, width):
+    """
+    Args:
+        windows (`torch.FloatTensor` of shape `(num_windows * batch_size, window_size, window_size, num_channels)`):
+            Input windows
+        window_size (`int`):
+            Window size
+        height (`int`):
+            Height of the resized audio
+        width (`int`):
+            Width of the resized audio
+    """
+    batch_size = int(windows.shape[0] / (height * width / window_size / window_size))
+
+    hidden_states = windows.view(batch_size, height // window_size, width // window_size, window_size, window_size, -1)
+    hidden_states = hidden_states.permute(0, 1, 3, 2, 4, 5).contiguous().view(batch_size, height, width, -1)
+    return hidden_states
+
+
+# Copied from transformers.models.roberta.modeling_roberta.create_position_ids_from_input_ids
+def create_position_ids_from_input_ids(input_ids, padding_idx, past_key_values_length=0):
+    """
+    Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols
+    are ignored. This is modified from fairseq's `utils.make_positions`.
+
+    Args:
+        x: torch.Tensor x:
+
+    Returns: torch.Tensor
+    """
+    # The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
+    mask = input_ids.ne(padding_idx).int()
+    incremental_indices = (torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length) * mask
+    return incremental_indices.long() + padding_idx
+
+
+# contrastive loss function, adapted from
+# https://sachinruk.github.io/blog/pytorch/pytorch%20lightning/loss%20function/gpu/2021/03/07/CLIP.html#CLIP-loss-function
+def contrastive_loss(logits: torch.Tensor) -> torch.Tensor:
+    labels = torch.arange(len(logits), device=logits.device)
+    return nn.functional.cross_entropy(logits, labels)
+
+
+@dataclass
+# Copied from transformers.models.clip.modeling_clip.CLIPTextModelOutput with CLIP->Clap
+class ClapTextModelOutput(ModelOutput):
+    """
+    Base class for text model's outputs that also contains a pooling of the last hidden states.
+
+    Args:
+        text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
+            The text embeddings obtained by applying the projection layer to the pooler_output.
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    text_embeds: Optional[torch.FloatTensor] = None
+    last_hidden_state: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class ClapAudioModelOutput(ModelOutput):
+    """
+    ClapAudio model output to mimic the output of the original implementation.
+
+    Args:
+        audio_embeds (`torch.FloatTensor` of shape `(batch_size, hidden_size)`):
+            The Audio embeddings obtained by applying the projection layer to the pooler_output.
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+    """
+
+    audio_embeds: Optional[torch.FloatTensor] = None
+    last_hidden_state: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+# Copied from transformers.models.clip.modeling_clip.CLIPOutput with CLIP->Clap, vision->audio, Vision->Audio, image->audio
+class ClapOutput(ModelOutput):
+    """
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`):
+            Contrastive loss for audio-text similarity.
+        logits_per_audio:(`torch.FloatTensor` of shape `(audio_batch_size, text_batch_size)`):
+            The scaled dot product scores between `audio_embeds` and `text_embeds`. This represents the audio-text
+            similarity scores.
+        logits_per_text:(`torch.FloatTensor` of shape `(text_batch_size, audio_batch_size)`):
+            The scaled dot product scores between `text_embeds` and `audio_embeds`. This represents the text-audio
+            similarity scores.
+        text_embeds(`torch.FloatTensor` of shape `(batch_size, output_dim`):
+            The text embeddings obtained by applying the projection layer to the pooled output of [`ClapTextModel`].
+        audio_embeds(`torch.FloatTensor` of shape `(batch_size, output_dim`):
+            The audio embeddings obtained by applying the projection layer to the pooled output of [`ClapAudioModel`].
+        text_model_output(`BaseModelOutputWithPooling`):
+            The output of the [`ClapTextModel`].
+        audio_model_output(`BaseModelOutputWithPooling`):
+            The output of the [`ClapAudioModel`].
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits_per_audio: torch.FloatTensor = None
+    logits_per_text: torch.FloatTensor = None
+    text_embeds: torch.FloatTensor = None
+    audio_embeds: torch.FloatTensor = None
+    text_model_output: BaseModelOutputWithPooling = None
+    audio_model_output: BaseModelOutputWithPooling = None
+
+    def to_tuple(self) -> Tuple[Any]:
+        return tuple(
+            self[k] if k not in ["text_model_output", "audio_model_output"] else getattr(self, k).to_tuple()
+            for k in self.keys()
+        )
+
+
+# Adapted from transformers.models.swin.modeling_swin.SwinDropPath
+class ClapDropPath(nn.Module):
+    """
+    Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks). This is a slightly
+    refactored version of the `SwinDropPath` implementation.
+    """
+
+    def __init__(self, drop_prob=None):
+        super().__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, hidden_states):
+        if self.drop_prob == 0.0 or not self.training:
+            return hidden_states
+
+        keep_prob = 1 - self.drop_prob
+        # work with diff dim tensors, not just 2D ConvNets
+        shape = (hidden_states.shape[0],) + (1,) * (hidden_states.ndim - 1)
+
+        random_tensor = keep_prob + torch.rand(shape, dtype=hidden_states.dtype, device=hidden_states.device)
+        random_tensor.floor_()  # binarize
+        output = hidden_states.div(keep_prob) * random_tensor
+        return output
+
+
+# Adapted from https://github.com/LAION-AI/CLAP/blob/6ad05a971ba0622f6acee8c41993e0d02bbed639/src/open_clip/feature_fusion.py#L133
+class ClapAudioAFFBlock(nn.Module):
+    r"""
+    ATTENTIONAL FEATURE FUSION Block from CLAP, since in CLAP we are always in 2D mode, it is not needed to implement
+    the 1D version.
+    """
+
+    def __init__(self, config: ClapAudioConfig):
+        super().__init__()
+        channels = config.patch_embeds_hidden_size
+        downsize_ratio = config.aff_block_r
+        inter_channels = int(channels // downsize_ratio)
+
+        self.local_att = nn.Sequential(
+            nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+            nn.BatchNorm2d(inter_channels),
+            nn.ReLU(inplace=True),
+            nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+            nn.BatchNorm2d(channels),
+        )
+        self.global_att = nn.Sequential(
+            nn.AdaptiveAvgPool2d(1),
+            nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+            nn.BatchNorm2d(inter_channels),
+            nn.ReLU(inplace=True),
+            nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+            nn.BatchNorm2d(channels),
+        )
+
+        self.sigmoid = nn.Sigmoid()
+
+    def forward(self, hidden_states, residual):
+        attention_input = hidden_states + residual
+
+        fused_layer_output = self.local_att(attention_input) + self.global_att(attention_input)
+        fused_layer_output = self.sigmoid(fused_layer_output)
+
+        output = 2 * hidden_states * fused_layer_output + 2 * residual * (1 - fused_layer_output)
+        return output
+
+
+class ClapAudioPatchEmbed(nn.Module):
+    """
+    This module converts the hidden states reshaped as an image to patch embeddings ready to be passed to the
+    Transformer block.
+    """
+
+    def __init__(self, config: ClapAudioConfig):
+        super().__init__()
+        img_size = (config.spec_size, config.spec_size) if isinstance(config.spec_size, int) else config.spec_size
+        patch_size = (
+            (config.patch_size, config.patch_size) if isinstance(config.patch_size, int) else config.patch_size
+        )
+        patch_stride = (
+            (config.patch_stride, config.patch_stride) if isinstance(config.patch_stride, int) else config.patch_stride
+        )
+
+        self.img_size = img_size
+        self.patch_stride = patch_stride
+
+        self.grid_size = (img_size[0] // patch_stride[0], img_size[1] // patch_stride[1])
+        self.num_patches = self.grid_size[0] * self.grid_size[1]
+
+        self.flatten = config.flatten_patch_embeds
+        self.enable_fusion = config.enable_fusion
+
+        padding = ((patch_size[0] - patch_stride[0]) // 2, (patch_size[1] - patch_stride[1]) // 2)
+
+        scale_factor = 4 if (self.enable_fusion) and (config.fusion_type == "channel_map") else 1
+
+        self.proj = nn.Conv2d(
+            config.patch_embed_input_channels * scale_factor,
+            config.patch_embeds_hidden_size,
+            kernel_size=patch_size,
+            stride=patch_stride,
+            padding=padding,
+        )
+
+        self.norm = nn.LayerNorm(config.patch_embeds_hidden_size) if config.enable_patch_layer_norm else nn.Identity()
+        if self.enable_fusion:
+            self.fusion_model = ClapAudioAFFBlock(config)
+            self.mel_conv2d = nn.Conv2d(
+                config.patch_embed_input_channels,
+                config.patch_embeds_hidden_size,
+                kernel_size=(patch_size[0], patch_size[1] * 3),
+                stride=(patch_stride[0], patch_stride[1] * 3),
+                padding=padding,
+            )
+
+    def forward(self, hidden_states, is_longer_idx=None):
+        if self.enable_fusion:
+            # retrieve the last mel as we have transposed the input
+            global_hidden_states = hidden_states[:, 0:1, :, :]
+
+            # global processing
+            batch_size, num_channels, height, width = global_hidden_states.shape
+
+            if height != self.img_size[0] or width != self.img_size[1]:
+                raise ValueError(
+                    f"Input audio size ({height}*{width}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+                )
+
+            global_hidden_states = self.proj(global_hidden_states)
+            output_width = global_hidden_states.size(-1)
+            if len(is_longer_idx) > 0:
+                # local processing
+                local_hidden_states = hidden_states[is_longer_idx, 1:, :, :].contiguous()
+                batch_size, num_channels, height, width = local_hidden_states.shape
+                local_hidden_states = local_hidden_states.view(batch_size * num_channels, 1, height, width)
+
+                local_hidden_states = self.mel_conv2d(local_hidden_states)
+
+                _, features, height, width = local_hidden_states.shape
+                local_hidden_states = local_hidden_states.view(batch_size, num_channels, features, height, width)
+                local_hidden_states = local_hidden_states.permute((0, 2, 3, 1, 4)).contiguous().flatten(3)
+
+                local_width = local_hidden_states.size(-1)
+                local_hidden_states = torch.nn.functional.pad(
+                    local_hidden_states, (0, output_width - local_width), "constant", 0
+                )
+
+                global_hidden_states[is_longer_idx] = self.fusion_model(
+                    global_hidden_states[is_longer_idx], local_hidden_states
+                )
+            hidden_states = global_hidden_states
+        else:
+            _, _, height, width = hidden_states.shape
+            if height != self.img_size[0] or width != self.img_size[1]:
+                raise ValueError(
+                    f"Input audio size ({height}*{width}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+                )
+            hidden_states = self.proj(hidden_states)
+
+        if self.flatten:
+            hidden_states = hidden_states.flatten(2).transpose(1, 2)
+        hidden_states = self.norm(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.swin.modeling_swin.SwinSelfAttention with Swin->ClapAudio
+class ClapAudioSelfAttention(nn.Module):
+    def __init__(self, config, dim, num_heads, window_size):
+        super().__init__()
+        if dim % num_heads != 0:
+            raise ValueError(
+                f"The hidden size ({dim}) is not a multiple of the number of attention heads ({num_heads})"
+            )
+
+        self.num_attention_heads = num_heads
+        self.attention_head_size = int(dim / num_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+        self.window_size = (
+            window_size if isinstance(window_size, collections.abc.Iterable) else (window_size, window_size)
+        )
+
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * self.window_size[0] - 1) * (2 * self.window_size[1] - 1), num_heads)
+        )
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        coords = torch.stack(meshgrid([coords_h, coords_w], indexing="ij"))
+        coords_flatten = torch.flatten(coords, 1)
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()
+        relative_coords[:, :, 0] += self.window_size[0] - 1
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.query = nn.Linear(self.all_head_size, self.all_head_size, bias=config.qkv_bias)
+        self.key = nn.Linear(self.all_head_size, self.all_head_size, bias=config.qkv_bias)
+        self.value = nn.Linear(self.all_head_size, self.all_head_size, bias=config.qkv_bias)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        batch_size, dim, num_channels = hidden_states.shape
+        mixed_query_layer = self.query(hidden_states)
+
+        key_layer = self.transpose_for_scores(self.key(hidden_states))
+        value_layer = self.transpose_for_scores(self.value(hidden_states))
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+
+        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)]
+        relative_position_bias = relative_position_bias.view(
+            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1
+        )
+
+        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()
+        attention_scores = attention_scores + relative_position_bias.unsqueeze(0)
+
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in ClapAudioModel forward() function)
+            mask_shape = attention_mask.shape[0]
+            attention_scores = attention_scores.view(
+                batch_size // mask_shape, mask_shape, self.num_attention_heads, dim, dim
+            )
+            attention_scores = attention_scores + attention_mask.unsqueeze(1).unsqueeze(0)
+            attention_scores = attention_scores.view(-1, self.num_attention_heads, dim, dim)
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        return outputs
+
+
+# Copied from transformers.models.swin.modeling_swin.SwinSelfOutput with Swin->ClapAudio
+class ClapAudioSelfOutput(nn.Module):
+    def __init__(self, config, dim):
+        super().__init__()
+        self.dense = nn.Linear(dim, dim)
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+
+        return hidden_states
+
+
+# Copied from transformers.models.swin.modeling_swin.SwinAttention with Swin->ClapAudio
+class ClapAudioAttention(nn.Module):
+    def __init__(self, config, dim, num_heads, window_size):
+        super().__init__()
+        self.self = ClapAudioSelfAttention(config, dim, num_heads, window_size)
+        self.output = ClapAudioSelfOutput(config, dim)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        self_outputs = self.self(hidden_states, attention_mask, head_mask, output_attentions)
+        attention_output = self.output(self_outputs[0], hidden_states)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+# Copied from transformers.models.swin.modeling_swin.SwinIntermediate with Swin->ClapAudio
+class ClapAudioIntermediate(nn.Module):
+    def __init__(self, config, dim):
+        super().__init__()
+        self.dense = nn.Linear(dim, int(config.mlp_ratio * dim))
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.swin.modeling_swin.SwinOutput with Swin->ClapAudio
+class ClapAudioOutput(nn.Module):
+    def __init__(self, config, dim):
+        super().__init__()
+        self.dense = nn.Linear(int(config.mlp_ratio * dim), dim)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.swin.modeling_swin.SwinLayer with SwinDropPath->ClapDropPath, Swin->ClapAudio
+class ClapAudioLayer(nn.Module):
+    def __init__(self, config, dim, input_resolution, num_heads, shift_size=0):
+        super().__init__()
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.shift_size = shift_size
+        self.window_size = config.window_size
+        self.input_resolution = input_resolution
+        self.layernorm_before = nn.LayerNorm(dim, eps=config.layer_norm_eps)
+        self.attention = ClapAudioAttention(config, dim, num_heads, window_size=self.window_size)
+        self.drop_path = ClapDropPath(config.drop_path_rate) if config.drop_path_rate > 0.0 else nn.Identity()
+        self.layernorm_after = nn.LayerNorm(dim, eps=config.layer_norm_eps)
+        self.intermediate = ClapAudioIntermediate(config, dim)
+        self.output = ClapAudioOutput(config, dim)
+
+    def set_shift_and_window_size(self, input_resolution):
+        if min(input_resolution) <= self.window_size:
+            # if window size is larger than input resolution, we don't partition windows
+            self.shift_size = 0
+            self.window_size = min(input_resolution)
+
+    def get_attn_mask(self, height, width, dtype):
+        if self.shift_size > 0:
+            # calculate attention mask for SW-MSA
+            img_mask = torch.zeros((1, height, width, 1), dtype=dtype)
+            height_slices = (
+                slice(0, -self.window_size),
+                slice(-self.window_size, -self.shift_size),
+                slice(-self.shift_size, None),
+            )
+            width_slices = (
+                slice(0, -self.window_size),
+                slice(-self.window_size, -self.shift_size),
+                slice(-self.shift_size, None),
+            )
+            count = 0
+            for height_slice in height_slices:
+                for width_slice in width_slices:
+                    img_mask[:, height_slice, width_slice, :] = count
+                    count += 1
+
+            mask_windows = window_partition(img_mask, self.window_size)
+            mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+            attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+            attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+        else:
+            attn_mask = None
+        return attn_mask
+
+    def maybe_pad(self, hidden_states, height, width):
+        pad_right = (self.window_size - width % self.window_size) % self.window_size
+        pad_bottom = (self.window_size - height % self.window_size) % self.window_size
+        pad_values = (0, 0, 0, pad_right, 0, pad_bottom)
+        hidden_states = nn.functional.pad(hidden_states, pad_values)
+        return hidden_states, pad_values
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        input_dimensions: Tuple[int, int],
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = False,
+        always_partition: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        if not always_partition:
+            self.set_shift_and_window_size(input_dimensions)
+        else:
+            pass
+        height, width = input_dimensions
+        batch_size, _, channels = hidden_states.size()
+        shortcut = hidden_states
+
+        hidden_states = self.layernorm_before(hidden_states)
+
+        hidden_states = hidden_states.view(batch_size, height, width, channels)
+
+        # pad hidden_states to multiples of window size
+        hidden_states, pad_values = self.maybe_pad(hidden_states, height, width)
+
+        _, height_pad, width_pad, _ = hidden_states.shape
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_hidden_states = torch.roll(hidden_states, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+        else:
+            shifted_hidden_states = hidden_states
+
+        # partition windows
+        hidden_states_windows = window_partition(shifted_hidden_states, self.window_size)
+        hidden_states_windows = hidden_states_windows.view(-1, self.window_size * self.window_size, channels)
+        attn_mask = self.get_attn_mask(height_pad, width_pad, dtype=hidden_states.dtype)
+        if attn_mask is not None:
+            attn_mask = attn_mask.to(hidden_states_windows.device)
+
+        attention_outputs = self.attention(
+            hidden_states_windows, attn_mask, head_mask, output_attentions=output_attentions
+        )
+
+        attention_output = attention_outputs[0]
+
+        attention_windows = attention_output.view(-1, self.window_size, self.window_size, channels)
+        shifted_windows = window_reverse(attention_windows, self.window_size, height_pad, width_pad)
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            attention_windows = torch.roll(shifted_windows, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            attention_windows = shifted_windows
+
+        was_padded = pad_values[3] > 0 or pad_values[5] > 0
+        if was_padded:
+            attention_windows = attention_windows[:, :height, :width, :].contiguous()
+
+        attention_windows = attention_windows.view(batch_size, height * width, channels)
+
+        hidden_states = shortcut + self.drop_path(attention_windows)
+
+        layer_output = self.layernorm_after(hidden_states)
+        layer_output = self.intermediate(layer_output)
+        layer_output = hidden_states + self.output(layer_output)
+
+        layer_outputs = (layer_output, attention_outputs[1]) if output_attentions else (layer_output,)
+        return layer_outputs
+
+
+# Copied from transformers.models.swin.modeling_swin.SwinStage with Swin->ClapAudio
+class ClapAudioStage(nn.Module):
+    def __init__(self, config, dim, input_resolution, depth, num_heads, drop_path, downsample):
+        super().__init__()
+        self.config = config
+        self.dim = dim
+        self.blocks = nn.ModuleList(
+            [
+                ClapAudioLayer(
+                    config=config,
+                    dim=dim,
+                    input_resolution=input_resolution,
+                    num_heads=num_heads,
+                    shift_size=0 if (i % 2 == 0) else config.window_size // 2,
+                )
+                for i in range(depth)
+            ]
+        )
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(input_resolution, dim=dim, norm_layer=nn.LayerNorm)
+        else:
+            self.downsample = None
+
+        self.pointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        input_dimensions: Tuple[int, int],
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = False,
+        always_partition: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        height, width = input_dimensions
+        for i, layer_module in enumerate(self.blocks):
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+
+            layer_outputs = layer_module(
+                hidden_states, input_dimensions, layer_head_mask, output_attentions, always_partition
+            )
+
+            hidden_states = layer_outputs[0]
+
+        hidden_states_before_downsampling = hidden_states
+        if self.downsample is not None:
+            height_downsampled, width_downsampled = (height + 1) // 2, (width + 1) // 2
+            output_dimensions = (height, width, height_downsampled, width_downsampled)
+            hidden_states = self.downsample(hidden_states_before_downsampling, input_dimensions)
+        else:
+            output_dimensions = (height, width, height, width)
+
+        stage_outputs = (hidden_states, hidden_states_before_downsampling, output_dimensions)
+
+        if output_attentions:
+            stage_outputs += layer_outputs[1:]
+        return stage_outputs
+
+
+# Copied from transformers.models.swin.modeling_swin.SwinPatchMerging with Swin->ClapAudio
+class ClapAudioPatchMerging(nn.Module):
+    """
+    Patch Merging Layer.
+
+    Args:
+        input_resolution (`Tuple[int]`):
+            Resolution of input feature.
+        dim (`int`):
+            Number of input channels.
+        norm_layer (`nn.Module`, *optional*, defaults to `nn.LayerNorm`):
+            Normalization layer class.
+    """
+
+    def __init__(self, input_resolution: Tuple[int], dim: int, norm_layer: nn.Module = nn.LayerNorm) -> None:
+        super().__init__()
+        self.input_resolution = input_resolution
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(4 * dim)
+
+    def maybe_pad(self, input_feature, height, width):
+        should_pad = (height % 2 == 1) or (width % 2 == 1)
+        if should_pad:
+            pad_values = (0, 0, 0, width % 2, 0, height % 2)
+            input_feature = nn.functional.pad(input_feature, pad_values)
+
+        return input_feature
+
+    def forward(self, input_feature: torch.Tensor, input_dimensions: Tuple[int, int]) -> torch.Tensor:
+        height, width = input_dimensions
+        # `dim` is height * width
+        batch_size, dim, num_channels = input_feature.shape
+
+        input_feature = input_feature.view(batch_size, height, width, num_channels)
+        # pad input to be disible by width and height, if needed
+        input_feature = self.maybe_pad(input_feature, height, width)
+        # [batch_size, height/2, width/2, num_channels]
+        input_feature_0 = input_feature[:, 0::2, 0::2, :]
+        # [batch_size, height/2, width/2, num_channels]
+        input_feature_1 = input_feature[:, 1::2, 0::2, :]
+        # [batch_size, height/2, width/2, num_channels]
+        input_feature_2 = input_feature[:, 0::2, 1::2, :]
+        # [batch_size, height/2, width/2, num_channels]
+        input_feature_3 = input_feature[:, 1::2, 1::2, :]
+        # batch_size height/2 width/2 4*num_channels
+        input_feature = torch.cat([input_feature_0, input_feature_1, input_feature_2, input_feature_3], -1)
+        input_feature = input_feature.view(batch_size, -1, 4 * num_channels)  # batch_size height/2*width/2 4*C
+
+        input_feature = self.norm(input_feature)
+        input_feature = self.reduction(input_feature)
+
+        return input_feature
+
+
+class ClapAudioEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.num_layers = len(config.depths)
+
+        self.config = config
+        self.patch_embed = ClapAudioPatchEmbed(config)
+        self.enable_fusion = config.enable_fusion
+        grid_size = self.patch_embed.grid_size
+        self.patch_stride = self.patch_embed.patch_stride
+        self.spec_size = config.spec_size
+        self.freq_ratio = self.spec_size // config.num_mel_bins
+
+        self.num_features = int(config.patch_embeds_hidden_size * 2 ** (self.num_layers - 1))
+        self.freq_ratio = config.spec_size // config.num_mel_bins
+
+        drop_path_rate = [x.item() for x in torch.linspace(0, config.drop_path_rate, sum(config.depths))]
+
+        self.input_resolutions = [(grid_size[0] // (2**i), grid_size[1] // (2**i)) for i in range(self.num_layers)]
+
+        self.layers = nn.ModuleList(
+            [
+                ClapAudioStage(
+                    config=config,
+                    dim=int(config.patch_embeds_hidden_size * 2**i_layer),
+                    input_resolution=self.input_resolutions[i_layer],
+                    depth=config.depths[i_layer],
+                    num_heads=config.num_attention_heads[i_layer],
+                    drop_path=drop_path_rate[sum(config.depths[:i_layer]) : sum(config.depths[: i_layer + 1])],
+                    downsample=ClapAudioPatchMerging if (i_layer < self.num_layers - 1) else None,
+                )
+                for i_layer in range(self.num_layers)
+            ]
+        )
+
+        self.gradient_checkpointing = False
+
+        self.batch_norm = nn.BatchNorm2d(config.num_mel_bins)
+        self.norm = nn.LayerNorm(self.num_features)
+        self.depths = config.depths
+        self.avgpool = nn.AdaptiveAvgPool1d(1)
+
+    def reshape_mel2img(self, normalized_input_features):
+        """
+        The input is 4 normalized log mel spectrograms. It is reshape to the common shape of images. Each channel
+        should represent 1 of the 4 crops of the spectrogram. For more details, refer to the [`ClapFeatureExtractor`].
+        """
+        _, _, time_length, freq_length = normalized_input_features.shape
+
+        spec_width = int(self.spec_size * self.freq_ratio)
+        spec_heigth = self.spec_size // self.freq_ratio
+
+        if time_length > spec_width or freq_length > spec_heigth:
+            raise ValueError("the wav size should be less than or equal to the swin input size")
+
+        # to avoid bicubic zero error
+        if time_length < spec_width:
+            normalized_input_features = nn.functional.interpolate(
+                normalized_input_features, (spec_width, freq_length), mode="bicubic", align_corners=True
+            )
+        if freq_length < spec_heigth:
+            normalized_input_features = nn.functional.interpolate(
+                normalized_input_features, (time_length, spec_heigth), mode="bicubic", align_corners=True
+            )
+
+        batch, channels, time, freq = normalized_input_features.shape
+
+        # batch_size, channels, spec_width, spec_heigth --> batch_size, channels, spec_heigth * freq_ratio, spec_width // freq_ratio
+        normalized_input_features = normalized_input_features.reshape(
+            batch, channels * self.freq_ratio, time // self.freq_ratio, freq
+        )
+        normalized_input_features = normalized_input_features.permute(0, 1, 3, 2).contiguous()
+        normalized_input_features = normalized_input_features.reshape(
+            batch, channels, freq * self.freq_ratio, time // self.freq_ratio
+        )
+
+        return normalized_input_features
+
+    def forward(
+        self,
+        input_features,
+        head_mask: Optional[torch.FloatTensor] = None,
+        is_longer: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = False,
+        output_hidden_states: Optional[bool] = False,
+        output_hidden_states_before_downsampling: Optional[bool] = False,
+        always_partition: Optional[bool] = False,
+        return_dict: Optional[bool] = True,
+    ) -> Union[Tuple, ClapAudioModelOutput]:
+        input_features = input_features.transpose(1, 3)
+        normalized_input_features = self.batch_norm(input_features)
+        normalized_input_features = normalized_input_features.transpose(1, 3)
+
+        is_longer_list_idx = None
+        if self.enable_fusion:
+            is_longer_list = is_longer.to(input_features.device)
+            is_longer_list_idx = torch.where(is_longer_list == 1)[0]
+
+        hidden_states = self.reshape_mel2img(normalized_input_features)
+
+        frames_num = hidden_states.shape[2]
+
+        hidden_states = self.patch_embed(hidden_states, is_longer_list_idx)
+
+        all_hidden_states = () if output_hidden_states else None
+        all_reshaped_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+
+        input_dimensions = self.input_resolutions[0]
+
+        if output_hidden_states:
+            batch_size, _, hidden_size = hidden_states.shape
+            # rearrange batch_size (height width) channels -> batch_size channel height width
+            reshaped_hidden_state = hidden_states.view(batch_size, *input_dimensions, hidden_size)
+            reshaped_hidden_state = reshaped_hidden_state.permute(0, 3, 1, 2)
+            all_hidden_states += (hidden_states,)
+            all_reshaped_hidden_states += (reshaped_hidden_state,)
+
+        for i, layer_module in enumerate(self.layers):
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+
+            input_dimensions = self.input_resolutions[i]
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module), hidden_states, input_dimensions, layer_head_mask
+                )
+            else:
+                layer_outputs = layer_module(
+                    hidden_states, input_dimensions, layer_head_mask, output_attentions, always_partition
+                )
+
+            hidden_states = layer_outputs[0]
+
+            hidden_states_before_downsampling = layer_outputs[1]
+            output_dimensions = layer_outputs[2]
+
+            input_dimensions = (output_dimensions[-2], output_dimensions[-1])
+
+            if output_hidden_states and output_hidden_states_before_downsampling:
+                batch_size, _, hidden_size = hidden_states_before_downsampling.shape
+                # rearrange batch_size (height width) channels -> batch_size channel height width
+                # here we use the original (not downsampled) height and width
+                reshaped_hidden_state = hidden_states_before_downsampling.view(
+                    batch_size, *(output_dimensions[0], output_dimensions[1]), hidden_size
+                )
+                reshaped_hidden_state = reshaped_hidden_state.permute(0, 3, 1, 2)
+                all_hidden_states += (hidden_states_before_downsampling,)
+                all_reshaped_hidden_states += (reshaped_hidden_state,)
+            elif output_hidden_states and not output_hidden_states_before_downsampling:
+                batch_size, _, hidden_size = hidden_states.shape
+                # rearrange batch_size (height width) channels -> batch_size channel height width
+                reshaped_hidden_state = hidden_states.view(batch_size, *input_dimensions, hidden_size)
+                reshaped_hidden_state = reshaped_hidden_state.permute(0, 3, 1, 2)
+                all_hidden_states += (hidden_states,)
+                all_reshaped_hidden_states += (reshaped_hidden_state,)
+
+            if output_attentions:
+                all_self_attentions += layer_outputs[3:]
+
+        last_hidden_state = self.norm(hidden_states)
+
+        batch_size, _, n_channels = last_hidden_state.shape
+
+        freq_shape = frames_num // (2 ** (len(self.depths) - 1)) // self.patch_stride[0]
+        temporal_shape = frames_num // (2 ** (len(self.depths) - 1)) // self.patch_stride[1]
+
+        last_hidden_state = (
+            last_hidden_state.permute(0, 2, 1).contiguous().reshape(batch_size, n_channels, freq_shape, temporal_shape)
+        )
+
+        batch_size, n_channels, n_frequencies, n_temp = last_hidden_state.shape
+        # group 2D CNN
+        c_freq_bin = n_frequencies // self.freq_ratio
+        last_hidden_state = last_hidden_state.reshape(
+            batch_size, n_channels, n_frequencies // c_freq_bin, c_freq_bin, n_temp
+        )
+        last_hidden_state = (
+            last_hidden_state.permute(0, 1, 3, 2, 4).contiguous().reshape(batch_size, n_channels, c_freq_bin, -1)
+        )
+        latent_output = self.avgpool(torch.flatten(last_hidden_state, 2))
+        latent_output = torch.flatten(latent_output, 1)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    last_hidden_state,
+                    latent_output,
+                    all_reshaped_hidden_states,
+                    all_self_attentions,
+                ]
+                if v is not None
+            )
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=latent_output,
+            hidden_states=all_reshaped_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+
+CLAP_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`ClapConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+CLAP_TEXT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+CLAP_AUDIO_INPUTS_DOCSTRING = r"""
+    Args:
+        input_features (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Input audio features. This should be returnes by the [`ClapFeatureExtractor`] class that you can also
+            retrieve from [`AutoFeatureExtractor`]. See [`ClapFeatureExtractor.__call__`] for details.
+        is_longer (`torch.FloatTensor`, of shape `(batch_size, 1)`, *optional*):
+            Whether the audio clip is longer than `max_length`. If `True`, a feature fusion will be enabled to enhance
+            the features.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+CLAP_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        input_features (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Input audio features. This should be returnes by the [`ClapFeatureExtractor`] class that you can also
+            retrieve from [`AutoFeatureExtractor`]. See [`ClapFeatureExtractor.__call__`] for details.
+        return_loss (`bool`, *optional*):
+            Whether or not to return the contrastive loss.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+class ClapProjectionLayer(nn.Module):
+    def __init__(self, config: Union[ClapAudioConfig, ClapTextConfig]):
+        super().__init__()
+        self.config = config
+        hidden_size = config.hidden_size
+        projection_dim = config.projection_dim
+
+        self.linear1 = nn.Linear(hidden_size, projection_dim)
+        self.activation = ACT2FN[config.projection_hidden_act]
+        self.linear2 = nn.Linear(projection_dim, projection_dim)
+
+    def forward(self, hidden_states):
+        hidden_states = self.linear1(hidden_states)
+        hidden_states = self.activation(hidden_states)
+        hidden_states = self.linear2(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaEmbeddings with Roberta->ClapText, persistent=False->persistent=True
+class ClapTextEmbeddings(nn.Module):
+    """
+    Same as BertEmbeddings with a tiny tweak for positional embeddings indexing.
+    """
+
+    # Copied from transformers.models.bert.modeling_bert.BertEmbeddings.__init__
+    def __init__(self, config):
+        super().__init__()
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
+        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.hidden_size)
+        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+        self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
+        self.register_buffer(
+            "token_type_ids", torch.zeros(self.position_ids.size(), dtype=torch.long), persistent=True
+        )
+
+        # End copy
+        self.padding_idx = config.pad_token_id
+        self.position_embeddings = nn.Embedding(
+            config.max_position_embeddings, config.hidden_size, padding_idx=self.padding_idx
+        )
+
+    def forward(
+        self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0
+    ):
+        if position_ids is None:
+            if input_ids is not None:
+                # Create the position ids from the input token ids. Any padded tokens remain padded.
+                position_ids = create_position_ids_from_input_ids(input_ids, self.padding_idx, past_key_values_length)
+            else:
+                position_ids = self.create_position_ids_from_inputs_embeds(inputs_embeds)
+
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
+        # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
+        # issue #5664
+        if token_type_ids is None:
+            if hasattr(self, "token_type_ids"):
+                buffered_token_type_ids = self.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(input_shape[0], seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        embeddings = inputs_embeds + token_type_embeddings
+        if self.position_embedding_type == "absolute":
+            position_embeddings = self.position_embeddings(position_ids)
+            embeddings += position_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+    def create_position_ids_from_inputs_embeds(self, inputs_embeds):
+        """
+        We are provided embeddings directly. We cannot infer which are padded so just generate sequential position ids.
+
+        Args:
+            inputs_embeds: torch.Tensor
+
+        Returns: torch.Tensor
+        """
+        input_shape = inputs_embeds.size()[:-1]
+        sequence_length = input_shape[1]
+
+        position_ids = torch.arange(
+            self.padding_idx + 1, sequence_length + self.padding_idx + 1, dtype=torch.long, device=inputs_embeds.device
+        )
+        return position_ids.unsqueeze(0).expand(input_shape)
+
+
+# Copied from transformers.models.bert.modeling_bert.BertSelfAttention with Bert->ClapText
+class ClapTextSelfAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
+                f"heads ({config.num_attention_heads})"
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = position_embedding_type or getattr(
+            config, "position_embedding_type", "absolute"
+        )
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
+
+        self.is_decoder = config.is_decoder
+
+    def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        mixed_query_layer = self.query(hidden_states)
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_layer = past_key_value[0]
+            value_layer = past_key_value[1]
+            attention_mask = encoder_attention_mask
+        elif is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
+            attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
+        else:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        use_cache = past_key_value is not None
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_layer, value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            query_length, key_length = query_layer.shape[2], key_layer.shape[2]
+            if use_cache:
+                position_ids_l = torch.tensor(key_length - 1, dtype=torch.long, device=hidden_states.device).view(
+                    -1, 1
+                )
+            else:
+                position_ids_l = torch.arange(query_length, dtype=torch.long, device=hidden_states.device).view(-1, 1)
+            position_ids_r = torch.arange(key_length, dtype=torch.long, device=hidden_states.device).view(1, -1)
+            distance = position_ids_l - position_ids_r
+
+            positional_embedding = self.distance_embedding(distance + self.max_position_embeddings - 1)
+            positional_embedding = positional_embedding.to(dtype=query_layer.dtype)  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                relative_position_scores_key = torch.einsum("bhrd,lrd->bhlr", key_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores_query + relative_position_scores_key
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in ClapTextModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        if self.is_decoder:
+            outputs = outputs + (past_key_value,)
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertSelfOutput
+class ClapTextSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertAttention with Bert->ClapText
+class ClapTextAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        self.self = ClapTextSelfAttention(config, position_embedding_type=position_embedding_type)
+        self.output = ClapTextSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        self_outputs = self.self(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+        )
+        attention_output = self.output(self_outputs[0], hidden_states)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertIntermediate
+class ClapTextIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertOutput
+class ClapTextOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertLayer with Bert->ClapText
+class ClapTextLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = ClapTextAttention(config)
+        self.is_decoder = config.is_decoder
+        self.add_cross_attention = config.add_cross_attention
+        if self.add_cross_attention:
+            if not self.is_decoder:
+                raise ValueError(f"{self} should be used as a decoder model if cross attention is added")
+            self.crossattention = ClapTextAttention(config, position_embedding_type="absolute")
+        self.intermediate = ClapTextIntermediate(config)
+        self.output = ClapTextOutput(config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
+        )
+        attention_output = self_attention_outputs[0]
+
+        # if decoder, the last output is tuple of self-attn cache
+        if self.is_decoder:
+            outputs = self_attention_outputs[1:-1]
+            present_key_value = self_attention_outputs[-1]
+        else:
+            outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        cross_attn_present_key_value = None
+        if self.is_decoder and encoder_hidden_states is not None:
+            if not hasattr(self, "crossattention"):
+                raise ValueError(
+                    f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers"
+                    " by setting `config.add_cross_attention=True`"
+                )
+
+            # cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
+            cross_attention_outputs = self.crossattention(
+                attention_output,
+                attention_mask,
+                head_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                cross_attn_past_key_value,
+                output_attentions,
+            )
+            attention_output = cross_attention_outputs[0]
+            outputs = outputs + cross_attention_outputs[1:-1]  # add cross attentions if we output attention weights
+
+            # add cross-attn cache to positions 3,4 of present_key_value tuple
+            cross_attn_present_key_value = cross_attention_outputs[-1]
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        layer_output = apply_chunking_to_forward(
+            self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output
+        )
+        outputs = (layer_output,) + outputs
+
+        # if decoder, return the attn key/values as the last output
+        if self.is_decoder:
+            outputs = outputs + (present_key_value,)
+
+        return outputs
+
+    def feed_forward_chunk(self, attention_output):
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        return layer_output
+
+
+# Copied from transformers.models.bert.modeling_bert.BertEncoder with Bert->ClapText
+class ClapTextEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList([ClapTextLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = False,
+        output_hidden_states: Optional[bool] = False,
+        return_dict: Optional[bool] = True,
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPastAndCrossAttentions]:
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
+
+        next_decoder_cache = () if use_cache else None
+        for i, layer_module in enumerate(self.layer):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+            past_key_value = past_key_values[i] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+                if use_cache:
+                    logger.warning(
+                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                    )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, past_key_value, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+                layer_outputs = layer_module(
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    past_key_value,
+                    output_attentions,
+                )
+
+            hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1],)
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+                if self.config.add_cross_attention:
+                    all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+# Copied from transformers.models.bert.modeling_bert.BertPooler
+class ClapTextPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+class ClapPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = ClapTextConfig
+    base_model_prefix = "clap"
+    supports_gradient_checkpointing = False
+    _keys_to_ignore_on_load_missing = [r"position_ids", r"logit_scale_a", r"logit_scale_t"]
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        factor = self.config.initializer_factor
+
+        if isinstance(module, ClapTextEmbeddings):
+            module.position_embeddings.weight.data.normal_(mean=0.0, std=factor * 0.02)
+            module.token_type_embeddings.weight.data.normal_(mean=0.0, std=factor * 0.02)
+        elif isinstance(module, ClapModel):
+            nn.init.normal_(module.logit_scale_a, std=factor * 0.02)
+            nn.init.normal_(module.logit_scale_t, std=factor * 0.02)
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=factor * 0.02)
+
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        elif isinstance(module, (nn.Conv2d, nn.Linear)):
+            in_proj_std = (self.config.hidden_size**-0.5) * ((2 * self.config.num_hidden_layers) ** -0.5) * factor
+            nn.init.normal_(module.weight, std=in_proj_std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, ClapTextEncoder):
+            module.gradient_checkpointing = value
+
+
+class ClapAudioModel(ClapPreTrainedModel):
+    config_class = ClapAudioConfig
+    main_input_name = "input_features"
+
+    def __init__(self, config: ClapAudioConfig):
+        super().__init__(config)
+        self.audio_encoder = ClapAudioEncoder(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.audio_encoder.patch_embed.proj
+
+    @add_start_docstrings_to_model_forward(CLAP_AUDIO_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=ClapAudioConfig)
+    def forward(
+        self,
+        input_features: Optional[torch.FloatTensor] = None,
+        is_longer: Optional[torch.BoolTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from datasets import load_dataset
+        >>> from transformers import AutoProcessor, ClapAudioModel
+
+        >>> dataset = load_dataset("ashraq/esc50")
+        >>> audio_sample = dataset["train"]["audio"][0]["array"]
+
+        >>> model = ClapAudioModel.from_pretrained("laion/clap-htsat-fused")
+        >>> processor = AutoProcessor.from_pretrained("laion/clap-htsat-fused")
+
+        >>> inputs = processor(audios=audio_sample, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        >>> last_hidden_state = outputs.audio_emmbeds
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+
+        return self.audio_encoder(
+            input_features=input_features,
+            is_longer=is_longer,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+
+class ClapTextModel(ClapPreTrainedModel):
+    """
+
+    The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
+    cross-attention is added between the self-attention layers, following the architecture described in *Attention is
+    all you need*_ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz
+    Kaiser and Illia Polosukhin.
+
+    To behave as an decoder the model needs to be initialized with the `is_decoder` argument of the configuration set
+    to `True`. To be used in a Seq2Seq model, the model needs to initialized with both `is_decoder` argument and
+    `add_cross_attention` set to `True`; an `encoder_hidden_states` is then expected as an input to the forward pass.
+
+    .. _*Attention is all you need*: https://arxiv.org/abs/1706.03762
+
+    """
+
+    config_class = ClapTextConfig
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    # Copied from transformers.models.bert.modeling_bert.BertModel.__init__ with Bert->ClapText
+    def __init__(self, config, add_pooling_layer=True):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = ClapTextEmbeddings(config)
+        self.encoder = ClapTextEncoder(config)
+
+        self.pooler = ClapTextPooler(config) if add_pooling_layer else None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    # Copied from transformers.models.bert.modeling_bert.BertModel.forward
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPoolingAndCrossAttentions]:
+        r"""
+        encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if self.config.is_decoder:
+            use_cache = use_cache if use_cache is not None else self.config.use_cache
+        else:
+            use_cache = False
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        batch_size, seq_length = input_shape
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        # past_key_values_length
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
+        if attention_mask is None:
+            attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)), device=device)
+
+        if token_type_ids is None:
+            if hasattr(self.embeddings, "token_type_ids"):
+                buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(batch_size, seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape)
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if self.config.is_decoder and encoder_hidden_states is not None:
+            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+            encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            token_type_ids=token_type_ids,
+            inputs_embeds=inputs_embeds,
+            past_key_values_length=past_key_values_length,
+        )
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            past_key_values=encoder_outputs.past_key_values,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+@add_start_docstrings(CLAP_START_DOCSTRING)
+class ClapModel(ClapPreTrainedModel):
+    config_class = ClapConfig
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def __init__(self, config: ClapConfig):
+        super().__init__(config)
+
+        if not isinstance(config.text_config, ClapTextConfig):
+            raise ValueError(
+                "config.text_config is expected to be of type ClapTextConfig but is of type"
+                f" {type(config.text_config)}."
+            )
+
+        if not isinstance(config.audio_config, ClapAudioConfig):
+            raise ValueError(
+                "config.audio_config is expected to be of type ClapAudioConfig but is of type"
+                f" {type(config.audio_config)}."
+            )
+
+        text_config = config.text_config
+        audio_config = config.audio_config
+
+        self.logit_scale_a = nn.Parameter(torch.ones([]) * np.log(config.logit_scale_init_value))
+        self.logit_scale_t = nn.Parameter(torch.ones([]) * np.log(config.logit_scale_init_value))
+
+        self.projection_dim = config.projection_dim
+
+        self.text_model = ClapTextModel(text_config)
+        self.text_projection = ClapProjectionLayer(text_config)
+
+        self.audio_model = ClapAudioModel(audio_config)
+        self.audio_projection = ClapProjectionLayer(audio_config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(CLAP_TEXT_INPUTS_DOCSTRING)
+    def get_text_features(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> torch.FloatTensor:
+        r"""
+        Returns:
+            text_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The text embeddings obtained by
+            applying the projection layer to the pooled output of [`ClapTextModel`].
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoTokenizer, ClapModel
+
+        >>> model = ClapModel.from_pretrained("laion/clap-htsat-unfused")
+        >>> tokenizer = AutoTokenizer.from_pretrained("laion/clap-htsat-unfused")
+
+        >>> inputs = tokenizer(["the sound of a cat", "the sound of a dog"], padding=True, return_tensors="pt")
+        >>> text_features = model.get_text_features(**inputs)
+        ```"""
+        # Use CLAP model's config for some fields (if specified) instead of those of audio & text components.
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        text_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = text_outputs[1] if return_dict is not None else text_outputs.pooler_output
+        text_features = self.text_projection(pooled_output)
+        text_features = F.normalize(text_features, dim=-1)
+
+        return text_features
+
+    @add_start_docstrings_to_model_forward(CLAP_AUDIO_INPUTS_DOCSTRING)
+    def get_audio_features(
+        self,
+        input_features: Optional[torch.Tensor] = None,
+        is_longer: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> torch.FloatTensor:
+        r"""
+        Returns:
+            audio_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The audio embeddings obtained by
+            applying the projection layer to the pooled output of [`ClapAudioModel`].
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoFeatureExtractor, ClapModel
+        >>> import torch
+
+        >>> model = ClapModel.from_pretrained("laion/clap-htsat-unfused")
+        >>> feature_extractor = AutoFeatureExtractor.from_pretrained("laion/clap-htsat-unfused")
+        >>> random_audio = torch.rand((16_000))
+        >>> inputs = feature_extractor(random_audio, return_tensors="pt")
+        >>> audio_features = model.get_audio_features(**inputs)
+        ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        audio_outputs = self.audio_model(
+            input_features=input_features,
+            is_longer=is_longer,
+            return_dict=return_dict,
+        )
+
+        pooled_output = audio_outputs[1] if not return_dict else audio_outputs.pooler_output
+
+        audio_features = self.audio_projection(pooled_output)
+        audio_features = F.normalize(audio_features, dim=-1)
+
+        return audio_features
+
+    @add_start_docstrings_to_model_forward(CLAP_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=ClapOutput, config_class=ClapConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        input_features: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        return_loss: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, ClapOutput]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from datasets import load_dataset
+        >>> from transformers import AutoProcessor, ClapModel
+
+        >>> dataset = load_dataset("ashraq/esc50")
+        >>> audio_sample = dataset["train"]["audio"][0]["array"]
+
+        >>> model = ClapModel.from_pretrained("laion/clap-htsat-unfused")
+        >>> processor = AutoProcessor.from_pretrained("laion/clap-htsat-unfused")
+
+        >>> input_text = ["Sound of a dog", "Sound of vaccum cleaner"]
+
+        >>> inputs = processor(text=input_text, audios=audio_sample, return_tensors="pt", padding=True)
+
+        >>> outputs = model(**inputs)
+        >>> logits_per_audio = outputs.logits_per_audio  # this is the audio-text similarity score
+        >>> probs = logits_per_audio.softmax(dim=-1)  # we can take the softmax to get the label probabilities
+        ```"""
+        # Use CLAP model's config for some fields (if specified) instead of those of audio & text components.
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        audio_outputs = self.audio_model(
+            input_features=input_features,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        text_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        audio_embeds = audio_outputs[1] if not return_dict else audio_outputs.pooler_output
+        audio_embeds = self.audio_projection(audio_embeds)
+
+        text_embeds = text_outputs[1] if not return_dict else text_outputs.pooler_output
+        text_embeds = self.text_projection(text_embeds)
+
+        # normalized features
+        audio_embeds = audio_embeds / audio_embeds.norm(p=2, dim=-1, keepdim=True)
+        text_embeds = text_embeds / text_embeds.norm(p=2, dim=-1, keepdim=True)
+
+        # cosine similarity as logits
+        logit_scale_text = self.logit_scale_t.exp()
+        logit_scale_audio = self.logit_scale_a.exp()
+        logits_per_text = torch.matmul(text_embeds, audio_embeds.t()) * logit_scale_text
+        logits_per_audio = torch.matmul(audio_embeds, text_embeds.t()) * logit_scale_audio
+
+        loss = None
+        if return_loss:
+            caption_loss = contrastive_loss(logits_per_text)
+            audio_loss = contrastive_loss(logits_per_text.t())
+            loss = (caption_loss + audio_loss) / 2.0
+
+        if not return_dict:
+            output = (logits_per_audio, logits_per_text, text_embeds, audio_embeds, text_outputs, audio_outputs)
+            return ((loss,) + output) if loss is not None else output
+
+        return ClapOutput(
+            loss=loss,
+            logits_per_audio=logits_per_audio,
+            logits_per_text=logits_per_text,
+            text_embeds=text_embeds,
+            audio_embeds=audio_embeds,
+            text_model_output=text_outputs,
+            audio_model_output=audio_outputs,
+        )
+
+
+@add_start_docstrings(
+    """
+    CLAP Text Model with a projection layer on top (a linear layer on top of the pooled output).
+    """,
+    CLAP_START_DOCSTRING,
+)
+class ClapTextModelWithProjection(ClapPreTrainedModel):
+    config_class = ClapTextConfig
+
+    def __init__(self, config: ClapTextConfig):
+        super().__init__(config)
+        self.text_model = ClapTextModel(config)
+        self.text_projection = ClapProjectionLayer(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.text_model.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.text_model.embeddings.word_embeddings = value
+
+    @add_start_docstrings_to_model_forward(CLAP_TEXT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=ClapTextModelOutput, config_class=ClapTextConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, ClapTextModelOutput]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoTokenizer, ClapTextModelWithProjection
+
+        >>> model = ClapTextModelWithProjection.from_pretrained("laion/clap-htsat-unfused")
+        >>> tokenizer = AutoTokenizer.from_pretrained("laion/clap-htsat-unfused")
+
+        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        >>> text_embeds = outputs.text_embeds
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        text_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = text_outputs[1] if not return_dict else text_outputs.pooler_output
+
+        text_embeds = self.text_projection(pooled_output)
+
+        if not return_dict:
+            outputs = (text_embeds, text_outputs[0]) + text_outputs[2:]
+            return tuple(output for output in outputs if output is not None)
+
+        return ClapTextModelOutput(
+            text_embeds=text_embeds,
+            last_hidden_state=text_outputs.last_hidden_state,
+            hidden_states=text_outputs.hidden_states,
+            attentions=text_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    CLAP Audio Model with a projection layer on top (a linear layer on top of the pooled output).
+    """,
+    CLAP_START_DOCSTRING,
+)
+class ClapAudioModelWithProjection(ClapPreTrainedModel):
+    config_class = ClapAudioConfig
+    main_input_name = "input_features"
+
+    def __init__(self, config: ClapAudioConfig):
+        super().__init__(config)
+        self.audio_model = ClapAudioModel(config)
+        self.audio_projection = ClapProjectionLayer(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.audio_model.audio_encoder.patch_embed.proj
+
+    @add_start_docstrings_to_model_forward(CLAP_AUDIO_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=ClapAudioModelOutput, config_class=ClapAudioConfig)
+    def forward(
+        self,
+        input_features: Optional[torch.FloatTensor] = None,
+        is_longer: Optional[torch.BoolTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, ClapAudioModelOutput]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from datasets import load_dataset
+        >>> from transformers import ClapAudioModelWithProjection, ClapProcessor
+
+        >>> model = ClapAudioModelWithProjection.from_pretrained("laion/clap-htsat-fused")
+        >>> processor = ClapProcessor.from_pretrained("laion/clap-htsat-fused")
+
+        >>> dataset = load_dataset("ashraq/esc50")
+        >>> audio_sample = dataset["train"]["audio"][0]["array"]
+
+        >>> inputs = processor(audios=audio_sample, return_tensors="pt")
+        >>> outputs = model(**inputs)
+        >>> audio_embeds = outputs.audio_embeds
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+
+        audio_outputs = self.audio_model(
+            input_features=input_features,
+            is_longer=is_longer,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = audio_outputs[1] if not return_dict else audio_outputs.pooler_output
+
+        audio_embeds = self.audio_projection(pooled_output)
+
+        if not return_dict:
+            outputs = (audio_embeds, audio_outputs[0]) + audio_outputs[2:]
+            return tuple(output for output in outputs if output is not None)
+
+        return ClapAudioModelOutput(
+            audio_embeds=audio_embeds,
+            last_hidden_state=audio_outputs.last_hidden_state,
+            attentions=audio_outputs.attentions,
+            hidden_states=audio_outputs.hidden_states,
+        )

src/transformers/models/clap/processing_clap.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Audio/Text processor class for CLAP
+"""
+
+from ...processing_utils import ProcessorMixin
+from ...tokenization_utils_base import BatchEncoding
+
+
+class ClapProcessor(ProcessorMixin):
+    r"""
+    Constructs a CLAP processor which wraps a CLAP feature extractor and a RoBerta tokenizer into a single processor.
+
+    [`ClapProcessor`] offers all the functionalities of [`ClapFeatureExtractor`] and [`RobertaTokenizerFast`]. See the
+    [`~ClapProcessor.__call__`] and [`~ClapProcessor.decode`] for more information.
+
+    Args:
+        feature_extractor ([`ClapFeatureExtractor`]):
+            The audio processor is a required input.
+        tokenizer ([`RobertaTokenizerFast`]):
+            The tokenizer is a required input.
+    """
+    feature_extractor_class = "ClapFeatureExtractor"
+    tokenizer_class = ("RobertaTokenizer", "RobertaTokenizerFast")
+
+    def __init__(self, feature_extractor, tokenizer):
+        super().__init__(feature_extractor, tokenizer)
+
+    def __call__(self, text=None, audios=None, return_tensors=None, **kwargs):
+        """
+        Main method to prepare for the model one or several sequences(s) and audio(s). This method forwards the `text`
+        and `kwargs` arguments to RobertaTokenizerFast's [`~RobertaTokenizerFast.__call__`] if `text` is not `None` to
+        encode the text. To prepare the audio(s), this method forwards the `audios` and `kwrags` arguments to
+        ClapFeatureExtractor's [`~ClapFeatureExtractor.__call__`] if `audios` is not `None`. Please refer to the
+        doctsring of the above two methods for more information.
+
+        Args:
+            text (`str`, `List[str]`, `List[List[str]]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            audios (`np.ndarray`, `torch.Tensor`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The audio or batch of audios to be prepared. Each audio can be NumPy array or PyTorch tensor. In case
+                of a NumPy array/PyTorch tensor, each audio should be of shape (C, T), where C is a number of channels,
+                and T the sample length of the audio.
+
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors of a particular framework. Acceptable values are:
+
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+                - `'jax'`: Return JAX `jnp.ndarray` objects.
+
+        Returns:
+            [`BatchEncoding`]: A [`BatchEncoding`] with the following fields:
+
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+              `None`).
+            - **audio_features** -- Audio features to be fed to a model. Returned when `audios` is not `None`.
+        """
+        sampling_rate = kwargs.pop("sampling_rate", None)
+
+        if text is None and audios is None:
+            raise ValueError("You have to specify either text or audios. Both cannot be none.")
+
+        if text is not None:
+            encoding = self.tokenizer(text, return_tensors=return_tensors, **kwargs)
+
+        if audios is not None:
+            audio_features = self.feature_extractor(
+                audios, sampling_rate=sampling_rate, return_tensors=return_tensors, **kwargs
+            )
+
+        if text is not None and audios is not None:
+            encoding["input_features"] = audio_features.input_features
+            return encoding
+        elif text is not None:
+            return encoding
+        else:
+            return BatchEncoding(data=dict(**audio_features), tensor_type=return_tensors)
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to RobertaTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to RobertaTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer
+        to the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    @property
+    def model_input_names(self):
+        tokenizer_input_names = self.tokenizer.model_input_names
+        feature_extractor_input_names = self.feature_extractor.model_input_names
+        return list(dict.fromkeys(tokenizer_input_names + feature_extractor_input_names))

src/transformers/utils/dummy_pt_objects.py

@@ -1464,6 +1464,58 @@ class ChineseCLIPVisionModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+CLAP_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class ClapAudioModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class ClapAudioModelWithProjection(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class ClapFeatureExtractor(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class ClapModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class ClapPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class ClapTextModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class ClapTextModelWithProjection(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 CLIP_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

tests/models/clap/test_feature_extraction_clap.py

+# coding=utf-8
+# Copyright 2023 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import itertools
+import random
+import unittest
+
+import numpy as np
+
+from transformers import ClapFeatureExtractor
+from transformers.testing_utils import require_torch, require_torchaudio
+from transformers.utils.import_utils import is_torch_available
+
+from ...test_sequence_feature_extraction_common import SequenceFeatureExtractionTestMixin
+
+
+if is_torch_available():
+    import torch
+
+global_rng = random.Random()
+
+
+# Copied from tests.models.whisper.test_feature_extraction_whisper.floats_list
+def floats_list(shape, scale=1.0, rng=None, name=None):
+    """Creates a random float32 tensor"""
+    if rng is None:
+        rng = global_rng
+
+    values = []
+    for batch_idx in range(shape[0]):
+        values.append([])
+        for _ in range(shape[1]):
+            values[-1].append(rng.random() * scale)
+
+    return values
+
+
+@require_torch
+@require_torchaudio
+# Copied from tests.models.whisper.test_feature_extraction_whisper.WhisperFeatureExtractionTester with Whisper->Clap
+class ClapFeatureExtractionTester(unittest.TestCase):
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        min_seq_length=400,
+        max_seq_length=2000,
+        feature_size=10,
+        hop_length=160,
+        chunk_length=8,
+        padding_value=0.0,
+        sampling_rate=4_000,
+        return_attention_mask=False,
+        do_normalize=True,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.min_seq_length = min_seq_length
+        self.max_seq_length = max_seq_length
+        self.seq_length_diff = (self.max_seq_length - self.min_seq_length) // (self.batch_size - 1)
+        self.padding_value = padding_value
+        self.sampling_rate = sampling_rate
+        self.return_attention_mask = return_attention_mask
+        self.do_normalize = do_normalize
+        self.feature_size = feature_size
+        self.chunk_length = chunk_length
+        self.hop_length = hop_length
+
+    def prepare_feat_extract_dict(self):
+        return {
+            "feature_size": self.feature_size,
+            "hop_length": self.hop_length,
+            "chunk_length": self.chunk_length,
+            "padding_value": self.padding_value,
+            "sampling_rate": self.sampling_rate,
+            "return_attention_mask": self.return_attention_mask,
+            "do_normalize": self.do_normalize,
+        }
+
+    def prepare_inputs_for_common(self, equal_length=False, numpify=False):
+        def _flatten(list_of_lists):
+            return list(itertools.chain(*list_of_lists))
+
+        if equal_length:
+            speech_inputs = [floats_list((self.max_seq_length, self.feature_size)) for _ in range(self.batch_size)]
+        else:
+            # make sure that inputs increase in size
+            speech_inputs = [
+                floats_list((x, self.feature_size))
+                for x in range(self.min_seq_length, self.max_seq_length, self.seq_length_diff)
+            ]
+        if numpify:
+            speech_inputs = [np.asarray(x) for x in speech_inputs]
+        return speech_inputs
+
+
+@require_torch
+@require_torchaudio
+# Copied from tests.models.whisper.test_feature_extraction_whisper.WhisperFeatureExtractionTest with Whisper->Clap
+class ClapFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest.TestCase):
+    feature_extraction_class = ClapFeatureExtractor
+
+    def setUp(self):
+        self.feat_extract_tester = ClapFeatureExtractionTester(self)
+
+    def test_call(self):
+        # Tests that all call wrap to encode_plus and batch_encode_plus
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
+        # create three inputs of length 800, 1000, and 1200
+        speech_inputs = [floats_list((1, x))[0] for x in range(800, 1400, 200)]
+        np_speech_inputs = [np.asarray(speech_input) for speech_input in speech_inputs]
+
+        # Test feature size
+        input_features = feature_extractor(np_speech_inputs, padding="max_length", return_tensors="np").input_features
+        self.assertTrue(input_features.ndim == 4)
+
+        # Test not batched input
+        encoded_sequences_1 = feature_extractor(speech_inputs[0], return_tensors="np").input_features
+        encoded_sequences_2 = feature_extractor(np_speech_inputs[0], return_tensors="np").input_features
+        self.assertTrue(np.allclose(encoded_sequences_1, encoded_sequences_2, atol=1e-3))
+
+        # Test batched
+        encoded_sequences_1 = feature_extractor(speech_inputs, return_tensors="np").input_features
+        encoded_sequences_2 = feature_extractor(np_speech_inputs, return_tensors="np").input_features
+        for enc_seq_1, enc_seq_2 in zip(encoded_sequences_1, encoded_sequences_2):
+            self.assertTrue(np.allclose(enc_seq_1, enc_seq_2, atol=1e-3))
+
+    def test_double_precision_pad(self):
+        import torch
+
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
+        np_speech_inputs = np.random.rand(100, 32).astype(np.float64)
+        py_speech_inputs = np_speech_inputs.tolist()
+
+        for inputs in [py_speech_inputs, np_speech_inputs]:
+            np_processed = feature_extractor.pad([{"input_features": inputs}], return_tensors="np")
+            self.assertTrue(np_processed.input_features.dtype == np.float32)
+            pt_processed = feature_extractor.pad([{"input_features": inputs}], return_tensors="pt")
+            self.assertTrue(pt_processed.input_features.dtype == torch.float32)
+
+    def _load_datasamples(self, num_samples):
+        from datasets import load_dataset
+
+        ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+        # automatic decoding with librispeech
+        speech_samples = ds.sort("id").select(range(num_samples))[:num_samples]["audio"]
+
+        return [x["array"] for x in speech_samples]
+
+    def integration_test_fusion(self):
+        # fmt: off
+        EXPECTED_INPUT_FEATURES = torch.tensor(
+            [
+                [
+                    -30.2194, -22.4424, -18.6442, -17.2452, -22.7392, -32.2576, -36.1404,
+                    -35.6120, -29.6229, -29.0454, -32.2157, -36.7664, -29.4436, -26.7825,
+                    -31.1811, -38.3918, -38.8749, -43.4485, -47.6236, -38.7528, -31.8574,
+                    -39.0591, -41.3190, -32.3319, -31.4699, -33.4502, -36.7412, -34.5265,
+                    -35.1091, -40.4518, -42.7346, -44.5909, -44.9747, -45.8328, -47.0772,
+                    -46.2723, -44.3613, -48.6253, -44.9551, -43.8700, -44.6104, -48.0146,
+                    -42.7614, -47.3587, -47.4369, -45.5018, -47.0198, -42.8759, -47.5056,
+                    -47.1567, -49.2621, -49.5643, -48.4330, -48.8495, -47.2512, -40.8439,
+                    -48.1234, -49.1218, -48.7222, -50.2399, -46.8487, -41.9921, -50.4015,
+                    -50.7827
+                ],
+                [
+                    -89.0141, -89.1411, -88.8096, -88.5480, -88.3481, -88.2038,
+                    -88.1105, -88.0647, -88.0636, -88.1051, -88.1877, -88.1110,
+                    -87.8613, -88.6679, -88.2685, -88.9684, -88.7977, -89.6264,
+                    -89.9299, -90.3184, -91.1446, -91.9265, -92.7267, -93.6099,
+                    -94.6395, -95.3243, -95.5923, -95.5773, -95.0889, -94.3354,
+                    -93.5746, -92.9287, -92.4525, -91.9798, -91.8852, -91.7500,
+                    -91.7259, -91.7561, -91.7959, -91.7070, -91.6914, -91.5019,
+                    -91.0640, -90.0807, -88.7102, -87.0826, -85.5956, -84.4441,
+                    -83.8461, -83.8605, -84.6702, -86.3900, -89.3073, -93.2926,
+                    -96.3813, -97.3529, -100.0000, -99.6942, -92.2851, -87.9588,
+                    -85.7214, -84.6807, -84.1940, -84.2021
+                ],
+                [
+                    -51.6882, -50.6852, -50.8198, -51.7428, -53.0325, -54.1619, -56.4903,
+                    -59.0314, -60.7996, -60.5164, -59.9680, -60.5393, -62.5796, -65.4166,
+                    -65.6149, -65.1409, -65.7226, -67.9057, -72.5089, -82.3530, -86.3189,
+                    -83.4241, -79.1279, -79.3384, -82.7335, -79.8316, -80.2167, -74.3638,
+                    -71.3930, -75.3849, -74.5381, -71.4504, -70.3791, -71.4547, -71.8820,
+                    -67.3885, -69.5686, -71.9852, -71.0307, -73.0053, -80.8802, -72.9227,
+                    -63.8526, -60.3260, -59.6012, -57.8316, -61.0603, -67.3403, -67.1709,
+                    -60.4967, -60.5079, -68.3345, -67.5213, -70.6416, -79.6219, -78.2198,
+                    -74.6851, -69.5718, -69.4968, -70.6882, -66.8175, -73.8558, -74.3855,
+                    -72.9405
+                ]
+            ]
+        )
+        # fmt: on
+        MEL_BIN = [963, 963, 161]
+        input_speech = self._load_datasamples(1)
+        feaure_extractor = ClapFeatureExtractor()
+        for padding, EXPECTED_VALUES, idx_in_mel in zip(
+            ["repeat", "repeatpad", None], EXPECTED_INPUT_FEATURES, MEL_BIN
+        ):
+            input_features = feaure_extractor(input_speech, return_tensors="pt", padding=padding).input_features
+            self.assertTrue(torch.allclose(input_features[0, idx_in_mel], EXPECTED_VALUES, atol=1e-4))
+
+    def integration_test_rand_trunc(self):
+        # TODO in this case we should set the seed and use a longer audio to properly see the random truncation
+        # fmt: off
+        EXPECTED_INPUT_FEATURES = torch.tensor(
+            [
+                [
+                    -42.3330, -36.2735, -35.9231, -43.5947, -48.4525, -46.5227, -42.6477,
+                    -47.2740, -51.4336, -50.0846, -51.8711, -50.4232, -47.4736, -54.2275,
+                    -53.3947, -55.4904, -54.8750, -54.5510, -55.4156, -57.4395, -51.7385,
+                    -55.9118, -57.7800, -63.2064, -67.0651, -61.4379, -56.4268, -54.8667,
+                    -52.3487, -56.4418, -57.1842, -55.1005, -55.6366, -59.4395, -56.8604,
+                    -56.4949, -61.6573, -61.0826, -60.3250, -63.7876, -67.4882, -60.2323,
+                    -54.6886, -50.5369, -47.7656, -45.8909, -49.1273, -57.4141, -58.3201,
+                    -51.9862, -51.4897, -59.2561, -60.4730, -61.2203, -69.3174, -69.7464,
+                    -65.5861, -58.9921, -59.5610, -61.0584, -58.1149, -64.4045, -66.2622,
+                    -64.4610
+                ],
+                [
+                    -41.2298, -38.4211, -39.8834, -45.9950, -47.3839, -43.9849, -46.0371,
+                    -52.5490, -56.6912, -51.8794, -50.1284, -49.7506, -53.9422, -63.2854,
+                    -56.5754, -55.0469, -55.3181, -55.8115, -56.0058, -57.9215, -58.7597,
+                    -59.1994, -59.2141, -64.4198, -73.5138, -64.4647, -59.3351, -54.5626,
+                    -54.7508, -65.0230, -60.0270, -54.7644, -56.0108, -60.1531, -57.6879,
+                    -56.3766, -63.3395, -65.3032, -61.5202, -63.0677, -68.4217, -60.6868,
+                    -54.4619, -50.8533, -47.7200, -45.9197, -49.0961, -57.7621, -59.0750,
+                    -51.9122, -51.4332, -59.4132, -60.3415, -61.6558, -70.7049, -69.7905,
+                    -66.9104, -59.0324, -59.6138, -61.2023, -58.2169, -65.3837, -66.4425,
+                    -64.4142
+                ],
+                [
+                    -51.6882, -50.6852, -50.8198, -51.7428, -53.0325, -54.1619, -56.4903,
+                    -59.0314, -60.7996, -60.5164, -59.9680, -60.5393, -62.5796, -65.4166,
+                    -65.6149, -65.1409, -65.7226, -67.9057, -72.5089, -82.3530, -86.3189,
+                    -83.4241, -79.1279, -79.3384, -82.7335, -79.8316, -80.2167, -74.3638,
+                    -71.3930, -75.3849, -74.5381, -71.4504, -70.3791, -71.4547, -71.8820,
+                    -67.3885, -69.5686, -71.9852, -71.0307, -73.0053, -80.8802, -72.9227,
+                    -63.8526, -60.3260, -59.6012, -57.8316, -61.0603, -67.3403, -67.1709,
+                    -60.4967, -60.5079, -68.3345, -67.5213, -70.6416, -79.6219, -78.2198,
+                    -74.6851, -69.5718, -69.4968, -70.6882, -66.8175, -73.8558, -74.3855,
+                    -72.9405
+                ]
+            ]
+        )
+        # fmt: on
+
+        input_speech = self._load_datasamples(1)
+        feaure_extractor = ClapFeatureExtractor()
+        for padding, EXPECTED_VALUES in zip(["repeat", "repeatpad", None], EXPECTED_INPUT_FEATURES):
+            input_features = feaure_extractor(
+                input_speech, return_tensors="pt", truncation="rand_trunc", padding=padding
+            ).input_features
+            self.assertTrue(torch.allclose(input_features[0, 0, :30], EXPECTED_VALUES, atol=1e-4))

tests/models/clap/test_modeling_clap.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Testing suite for the PyTorch CLAP model. """
+
+
+import inspect
+import os
+import tempfile
+import unittest
+
+import numpy as np
+from datasets import load_dataset
+
+from transformers import ClapAudioConfig, ClapConfig, ClapProcessor, ClapTextConfig
+from transformers.testing_utils import require_torch, slow, torch_device
+from transformers.utils import is_torch_available
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import (
+    ModelTesterMixin,
+    _config_zero_init,
+    floats_tensor,
+    ids_tensor,
+    random_attention_mask,
+)
+
+
+if is_torch_available():
+    import torch
+    from torch import nn
+
+    from transformers import (
+        ClapAudioModel,
+        ClapAudioModelWithProjection,
+        ClapModel,
+        ClapTextModel,
+        ClapTextModelWithProjection,
+    )
+    from transformers.models.clap.modeling_clap import CLAP_PRETRAINED_MODEL_ARCHIVE_LIST
+
+
+class ClapAudioModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=12,
+        image_size=60,
+        num_mel_bins=16,
+        window_size=4,
+        spec_size=64,
+        patch_size=2,
+        patch_stride=2,
+        seq_length=16,
+        freq_ratio=2,
+        num_channels=3,
+        is_training=True,
+        hidden_size=256,
+        patch_embeds_hidden_size=32,
+        projection_dim=32,
+        num_hidden_layers=4,
+        num_heads=[2, 2, 2, 2],
+        intermediate_size=37,
+        dropout=0.1,
+        attention_dropout=0.1,
+        initializer_range=0.02,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.image_size = image_size
+        self.num_mel_bins = num_mel_bins
+        self.window_size = window_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.is_training = is_training
+        self.hidden_size = hidden_size
+        self.projection_dim = projection_dim
+        self.num_hidden_layers = num_hidden_layers
+        self.num_heads = num_heads
+        self.num_attention_heads = num_heads[0]
+        self.seq_length = seq_length
+        self.spec_size = spec_size
+        self.freq_ratio = freq_ratio
+        self.patch_stride = patch_stride
+        self.patch_embeds_hidden_size = patch_embeds_hidden_size
+        self.intermediate_size = intermediate_size
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.initializer_range = initializer_range
+        self.scope = scope
+
+    def prepare_config_and_inputs(self):
+        input_features = floats_tensor([self.batch_size, 1, self.hidden_size, self.num_mel_bins])
+        config = self.get_config()
+
+        return config, input_features
+
+    def get_config(self):
+        return ClapAudioConfig(
+            image_size=self.image_size,
+            patch_size=self.patch_size,
+            num_mel_bins=self.num_mel_bins,
+            window_size=self.window_size,
+            num_channels=self.num_channels,
+            hidden_size=self.hidden_size,
+            patch_stride=self.patch_stride,
+            projection_dim=self.projection_dim,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_heads,
+            intermediate_size=self.intermediate_size,
+            dropout=self.dropout,
+            attention_dropout=self.attention_dropout,
+            initializer_range=self.initializer_range,
+            spec_size=self.spec_size,
+            freq_ratio=self.freq_ratio,
+            patch_embeds_hidden_size=self.patch_embeds_hidden_size,
+        )
+
+    def create_and_check_model(self, config, input_features):
+        model = ClapAudioModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        with torch.no_grad():
+            result = model(input_features)
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    def create_and_check_model_with_projection(self, config, input_features):
+        model = ClapAudioModelWithProjection(config=config)
+        model.to(torch_device)
+        model.eval()
+        with torch.no_grad():
+            result = model(input_features)
+        self.parent.assertEqual(result.audio_embeds.shape, (self.batch_size, self.projection_dim))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_features = config_and_inputs
+        inputs_dict = {"input_features": input_features}
+        return config, inputs_dict
+
+
+@require_torch
+class ClapAudioModelTest(ModelTesterMixin, unittest.TestCase):
+    """
+    Here we also overwrite some of the tests of test_modeling_common.py, as CLAP does not use input_ids, inputs_embeds,
+    attention_mask and seq_length.
+    """
+
+    all_model_classes = (ClapAudioModel, ClapAudioModelWithProjection) if is_torch_available() else ()
+    fx_compatible = False
+    test_pruning = False
+    test_resize_embeddings = False
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = ClapAudioModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=ClapAudioConfig, has_text_modality=False, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    @unittest.skip(reason="ClapAudioModel does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    def test_model_common_attributes(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            self.assertIsInstance(model.get_input_embeddings(), (nn.Module))
+            x = model.get_output_embeddings()
+            self.assertTrue(x is None or isinstance(x, nn.Linear))
+
+    def test_hidden_states_output(self):
+        def check_hidden_states_output(inputs_dict, config, model_class):
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+
+            hidden_states = outputs.hidden_states
+
+            expected_num_layers = getattr(
+                self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1
+            )
+            self.assertEqual(len(hidden_states), expected_num_layers)
+
+            self.assertListEqual(
+                list(hidden_states[0].shape[-2:]),
+                [self.model_tester.patch_embeds_hidden_size, self.model_tester.patch_embeds_hidden_size],
+            )
+
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            inputs_dict["output_hidden_states"] = True
+            check_hidden_states_output(inputs_dict, config, model_class)
+
+            # check that output_hidden_states also work using config
+            del inputs_dict["output_hidden_states"]
+            config.output_hidden_states = True
+
+            check_hidden_states_output(inputs_dict, config, model_class)
+
+    @unittest.skip(reason="ClapAudioModel does not output any loss term in the forward pass")
+    def test_retain_grad_hidden_states_attentions(self):
+        pass
+
+    def test_forward_signature(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            signature = inspect.signature(model.forward)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = [*signature.parameters.keys()]
+
+            expected_arg_names = ["input_features"]
+            self.assertListEqual(arg_names[:1], expected_arg_names)
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_model_with_projection(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model_with_projection(*config_and_inputs)
+
+    @unittest.skip(reason="ClapAudioModel does not output any loss term in the forward pass")
+    def test_training(self):
+        pass
+
+    @unittest.skip(reason="ClapAudioModel does not output any loss term in the forward pass")
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    @unittest.skip(reason="ClapAudioModel has no base class and is not available in MODEL_MAPPING")
+    def test_save_load_fast_init_from_base(self):
+        pass
+
+    @unittest.skip(reason="ClapAudioModel has no base class and is not available in MODEL_MAPPING")
+    def test_save_load_fast_init_to_base(self):
+        pass
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in CLAP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = ClapAudioModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+    @slow
+    def test_model_with_projection_from_pretrained(self):
+        for model_name in CLAP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = ClapAudioModelWithProjection.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+            self.assertTrue(hasattr(model, "visual_projection"))
+
+
+class ClapTextModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=12,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=32,
+        projection_dim=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        dropout=0.1,
+        attention_dropout=0.1,
+        max_position_embeddings=512,
+        initializer_range=0.02,
+        scope=None,
+        projection_hidden_act="relu",
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_input_mask = use_input_mask
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.projection_dim = projection_dim
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.scope = scope
+        self.projection_hidden_act = projection_hidden_act
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+
+        if input_mask is not None:
+            batch_size, seq_length = input_mask.shape
+            rnd_start_indices = np.random.randint(1, seq_length - 1, size=(batch_size,))
+            for batch_idx, start_index in enumerate(rnd_start_indices):
+                input_mask[batch_idx, :start_index] = 1
+                input_mask[batch_idx, start_index:] = 0
+
+        config = self.get_config()
+
+        return config, input_ids, input_mask
+
+    def get_config(self):
+        return ClapTextConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            projection_dim=self.projection_dim,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            dropout=self.dropout,
+            attention_dropout=self.attention_dropout,
+            max_position_embeddings=self.max_position_embeddings,
+            initializer_range=self.initializer_range,
+            projection_hidden_act=self.projection_hidden_act,
+        )
+
+    def create_and_check_model(self, config, input_ids, input_mask):
+        model = ClapTextModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        with torch.no_grad():
+            result = model(input_ids, attention_mask=input_mask)
+            result = model(input_ids)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    def create_and_check_model_with_projection(self, config, input_ids, input_mask):
+        model = ClapTextModelWithProjection(config=config)
+        model.to(torch_device)
+        model.eval()
+        with torch.no_grad():
+            result = model(input_ids, attention_mask=input_mask)
+            result = model(input_ids)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+        self.parent.assertEqual(result.text_embeds.shape, (self.batch_size, self.projection_dim))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, input_mask = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
+        return config, inputs_dict
+
+
+@require_torch
+class ClapTextModelTest(ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (ClapTextModel, ClapTextModelWithProjection) if is_torch_available() else ()
+    fx_compatible = False
+    test_pruning = False
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = ClapTextModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=ClapTextConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_model_with_projection(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model_with_projection(*config_and_inputs)
+
+    @unittest.skip(reason="ClapTextModel does not output any loss term in the forward pass")
+    def test_training(self):
+        pass
+
+    @unittest.skip(reason="ClapTextModel does not output any loss term in the forward pass")
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    @unittest.skip(reason="ClapTextModel does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="ClapTextModel has no base class and is not available in MODEL_MAPPING")
+    def test_save_load_fast_init_from_base(self):
+        pass
+
+    @unittest.skip(reason="ClapTextModel has no base class and is not available in MODEL_MAPPING")
+    def test_save_load_fast_init_to_base(self):
+        pass
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in CLAP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = ClapTextModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+    @slow
+    def test_model_with_projection_from_pretrained(self):
+        for model_name in CLAP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = ClapTextModelWithProjection.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+            self.assertTrue(hasattr(model, "text_projection"))
+
+
+class ClapModelTester:
+    def __init__(self, parent, text_kwargs=None, audio_kwargs=None, is_training=True):
+        if text_kwargs is None:
+            text_kwargs = {}
+        if audio_kwargs is None:
+            audio_kwargs = {}
+
+        self.parent = parent
+        self.text_model_tester = ClapTextModelTester(parent, **text_kwargs)
+        self.audio_model_tester = ClapAudioModelTester(parent, **audio_kwargs)
+        self.is_training = is_training
+
+    def prepare_config_and_inputs(self):
+        _, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
+        _, input_features = self.audio_model_tester.prepare_config_and_inputs()
+
+        config = self.get_config()
+
+        return config, input_ids, attention_mask, input_features
+
+    def get_config(self):
+        return ClapConfig.from_text_audio_configs(
+            self.text_model_tester.get_config(), self.audio_model_tester.get_config(), projection_dim=64
+        )
+
+    def create_and_check_model(self, config, input_ids, attention_mask, input_features):
+        model = ClapModel(config).to(torch_device).eval()
+        with torch.no_grad():
+            result = model(input_ids, input_features, attention_mask)
+        self.parent.assertEqual(
+            result.logits_per_audio.shape, (self.audio_model_tester.batch_size, self.text_model_tester.batch_size)
+        )
+        self.parent.assertEqual(
+            result.logits_per_text.shape, (self.text_model_tester.batch_size, self.audio_model_tester.batch_size)
+        )
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, attention_mask, input_features = config_and_inputs
+        inputs_dict = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "input_features": input_features,
+            "return_loss": True,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class ClapModelTest(ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (ClapModel,) if is_torch_available() else ()
+    fx_compatible = False
+    test_head_masking = False
+    test_pruning = False
+    test_resize_embeddings = False
+    test_attention_outputs = False
+
+    def setUp(self):
+        self.model_tester = ClapModelTester(self)
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    @unittest.skip(reason="Hidden_states is tested in individual model tests")
+    def test_hidden_states_output(self):
+        pass
+
+    @unittest.skip(reason="Inputs_embeds is tested in individual model tests")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="Retain_grad is tested in individual model tests")
+    def test_retain_grad_hidden_states_attentions(self):
+        pass
+
+    @unittest.skip(reason="ClapModel does not have input/output embeddings")
+    def test_model_common_attributes(self):
+        pass
+
+    # override as the `logit_scale` parameter initilization is different for CLAP
+    def test_initialization(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        configs_no_init = _config_zero_init(config)
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            for name, param in model.named_parameters():
+                if param.requires_grad:
+                    # check if `logit_scale` is initilized as per the original implementation
+                    if name == "logit_scale":
+                        self.assertAlmostEqual(
+                            param.data.item(),
+                            np.log(1 / 0.07),
+                            delta=1e-3,
+                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                        )
+                    else:
+                        self.assertIn(
+                            ((param.data.mean() * 1e9).round() / 1e9).item(),
+                            [0.0, 1.0],
+                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                        )
+
+    def _create_and_check_torchscript(self, config, inputs_dict):
+        if not self.test_torchscript:
+            return
+
+        configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
+        configs_no_init.torchscript = True
+        configs_no_init.return_dict = False
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            model.to(torch_device)
+            model.eval()
+
+            try:
+                input_ids = inputs_dict["input_ids"]
+                input_features = inputs_dict["input_features"]  # CLAP needs input_features
+                traced_model = torch.jit.trace(model, (input_ids, input_features))
+            except RuntimeError:
+                self.fail("Couldn't trace module.")
+
+            with tempfile.TemporaryDirectory() as tmp_dir_name:
+                pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
+
+                try:
+                    torch.jit.save(traced_model, pt_file_name)
+                except Exception:
+                    self.fail("Couldn't save module.")
+
+                try:
+                    loaded_model = torch.jit.load(pt_file_name)
+                except Exception:
+                    self.fail("Couldn't load module.")
+
+            model.to(torch_device)
+            model.eval()
+
+            loaded_model.to(torch_device)
+            loaded_model.eval()
+
+            model_state_dict = model.state_dict()
+            loaded_model_state_dict = loaded_model.state_dict()
+
+            self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))
+
+            models_equal = True
+            for layer_name, p1 in model_state_dict.items():
+                p2 = loaded_model_state_dict[layer_name]
+                if p1.data.ne(p2.data).sum() > 0:
+                    models_equal = False
+
+            self.assertTrue(models_equal)
+
+    def test_load_audio_text_config(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        # Save ClapConfig and check if we can load ClapAudioConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            audio_config = ClapAudioConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.audio_config.to_dict(), audio_config.to_dict())
+
+        # Save ClapConfig and check if we can load ClapTextConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            text_config = ClapTextConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.text_config.to_dict(), text_config.to_dict())
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in CLAP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = ClapModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+@slow
+@require_torch
+class ClapModelIntegrationTest(unittest.TestCase):
+    paddings = ["repeatpad", "repeat", "pad"]
+
+    def test_integration_unfused(self):
+        EXPECTED_MEANS_UNFUSED = {
+            "repeatpad": 0.0024,
+            "pad": 0.0020,
+            "repeat": 0.0023,
+        }
+
+        librispeech_dummy = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+        audio_sample = librispeech_dummy[-1]
+
+        model_id = "laion/clap-htsat-unfused"
+
+        model = ClapModel.from_pretrained(model_id).to(torch_device)
+        processor = ClapProcessor.from_pretrained(model_id)
+
+        for padding in self.paddings:
+            inputs = processor(audios=audio_sample["audio"]["array"], return_tensors="pt", padding=padding).to(
+                torch_device
+            )
+
+            audio_embed = model.get_audio_features(**inputs)
+            expected_mean = EXPECTED_MEANS_UNFUSED[padding]
+
+            self.assertTrue(
+                torch.allclose(audio_embed.cpu().mean(), torch.tensor([expected_mean]), atol=1e-3, rtol=1e-3)
+            )
+
+    def test_integration_fused(self):
+        EXPECTED_MEANS_FUSED = {
+            "repeatpad": 0.00069,
+            "repeat": 0.00196,
+            "pad": -0.000379,
+        }
+
+        librispeech_dummy = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+        audio_sample = librispeech_dummy[-1]
+
+        model_id = "laion/clap-htsat-fused"
+
+        model = ClapModel.from_pretrained(model_id).to(torch_device)
+        processor = ClapProcessor.from_pretrained(model_id)
+
+        for padding in self.paddings:
+            inputs = processor(
+                audios=audio_sample["audio"]["array"], return_tensors="pt", padding=padding, truncation="fusion"
+            ).to(torch_device)
+
+            audio_embed = model.get_audio_features(**inputs)
+            expected_mean = EXPECTED_MEANS_FUSED[padding]
+
+            self.assertTrue(
+                torch.allclose(audio_embed.cpu().mean(), torch.tensor([expected_mean]), atol=1e-3, rtol=1e-3)
+            )

tests/models/clap/test_processor_clap.py

+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import shutil
+import tempfile
+import unittest
+
+from transformers import ClapFeatureExtractor, ClapProcessor, RobertaTokenizer, RobertaTokenizerFast
+from transformers.testing_utils import require_sentencepiece, require_torchaudio
+
+from .test_feature_extraction_clap import floats_list
+
+
+@require_torchaudio
+@require_sentencepiece
+class ClapProcessorTest(unittest.TestCase):
+    def setUp(self):
+        self.checkpoint = "laion/clap-htsat-unfused"
+        self.tmpdirname = tempfile.mkdtemp()
+
+    def get_tokenizer(self, **kwargs):
+        return RobertaTokenizer.from_pretrained(self.checkpoint, **kwargs)
+
+    def get_feature_extractor(self, **kwargs):
+        return ClapFeatureExtractor.from_pretrained(self.checkpoint, **kwargs)
+
+    def tearDown(self):
+        shutil.rmtree(self.tmpdirname)
+
+    def test_save_load_pretrained_default(self):
+        tokenizer = self.get_tokenizer()
+        feature_extractor = self.get_feature_extractor()
+
+        processor = ClapProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        processor.save_pretrained(self.tmpdirname)
+        processor = ClapProcessor.from_pretrained(self.tmpdirname)
+
+        self.assertEqual(processor.tokenizer.get_vocab(), tokenizer.get_vocab())
+        self.assertIsInstance(processor.tokenizer, RobertaTokenizerFast)
+
+        self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor.to_json_string())
+        self.assertIsInstance(processor.feature_extractor, ClapFeatureExtractor)
+
+    def test_save_load_pretrained_additional_features(self):
+        processor = ClapProcessor(tokenizer=self.get_tokenizer(), feature_extractor=self.get_feature_extractor())
+        processor.save_pretrained(self.tmpdirname)
+
+        tokenizer_add_kwargs = self.get_tokenizer(bos_token="(BOS)", eos_token="(EOS)")
+        feature_extractor_add_kwargs = self.get_feature_extractor(do_normalize=False, padding_value=1.0)
+
+        processor = ClapProcessor.from_pretrained(
+            self.tmpdirname, bos_token="(BOS)", eos_token="(EOS)", do_normalize=False, padding_value=1.0
+        )
+
+        self.assertEqual(processor.tokenizer.get_vocab(), tokenizer_add_kwargs.get_vocab())
+        self.assertIsInstance(processor.tokenizer, RobertaTokenizerFast)
+
+        self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor_add_kwargs.to_json_string())
+        self.assertIsInstance(processor.feature_extractor, ClapFeatureExtractor)
+
+    def test_feature_extractor(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = ClapProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        raw_speech = floats_list((3, 1000))
+
+        input_feat_extract = feature_extractor(raw_speech, return_tensors="np")
+        input_processor = processor(audios=raw_speech, return_tensors="np")
+
+        for key in input_feat_extract.keys():
+            self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2)
+
+    def test_tokenizer(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = ClapProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        input_str = "This is a test string"
+
+        encoded_processor = processor(text=input_str)
+
+        encoded_tok = tokenizer(input_str)
+
+        for key in encoded_tok.keys():
+            self.assertListEqual(encoded_tok[key], encoded_processor[key])
+
+    def test_tokenizer_decode(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = ClapProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        predicted_ids = [[1, 4, 5, 8, 1, 0, 8], [3, 4, 3, 1, 1, 8, 9]]
+
+        decoded_processor = processor.batch_decode(predicted_ids)
+        decoded_tok = tokenizer.batch_decode(predicted_ids)
+
+        self.assertListEqual(decoded_tok, decoded_processor)
+
+    def test_model_input_names(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = ClapProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        self.assertListEqual(
+            processor.model_input_names[2:],
+            feature_extractor.model_input_names,
+            msg="`processor` and `feature_extractor` model input names do not match",
+        )

utils/check_repo.py

@@ -172,6 +172,10 @@ TEST_FILES_WITH_NO_COMMON_TESTS = [
 # should **not** be the rule.
 IGNORE_NON_AUTO_CONFIGURED = PRIVATE_MODELS.copy() + [
     # models to ignore for model xxx mapping
+    "ClapTextModel",
+    "ClapTextModelWithProjection",
+    "ClapAudioModel",
+    "ClapAudioModelWithProjection",
     "Blip2ForConditionalGeneration",
     "Blip2QFormerModel",
     "Blip2VisionModel",

utils/documentation_tests.txt

@@ -40,6 +40,8 @@ src/transformers/models/bloom/configuration_bloom.py
 src/transformers/models/camembert/configuration_camembert.py
 src/transformers/models/canine/configuration_canine.py
 src/transformers/models/canine/modeling_canine.py
+src/transformers/models/clap/configuration_clap.py
+src/transformers/models/clap/modeling_clap.py
 src/transformers/models/clip/configuration_clip.py
 src/transformers/models/clipseg/modeling_clipseg.py
 src/transformers/models/codegen/configuration_codegen.py