Add Descript-Audio-Codec model

docs/source/en/_toctree.yml

@@ -696,6 +696,8 @@
         title: Bark
       - local: model_doc/clap
         title: CLAP
+      - local: model_doc/dac
+        title: dac
       - local: model_doc/encodec
         title: EnCodec
       - local: model_doc/hiera

docs/source/en/index.md

@@ -105,6 +105,7 @@ Flax), PyTorch, and/or TensorFlow.
 |                       [CPM-Ant](model_doc/cpmant)                        |       ✅        |         ❌         |      ❌      |
 |                          [CTRL](model_doc/ctrl)                          |       ✅        |         ✅         |      ❌      |
 |                           [CvT](model_doc/cvt)                           |       ✅        |         ✅         |      ❌      |
+|                           [DAC](model_doc/dac)                           |       ✅        |         ❌         |      ❌      |
 |                   [Data2VecAudio](model_doc/data2vec)                    |       ✅        |         ❌         |      ❌      |
 |                    [Data2VecText](model_doc/data2vec)                    |       ✅        |         ❌         |      ❌      |
 |                   [Data2VecVision](model_doc/data2vec)                   |       ✅        |         ✅         |      ❌      |

docs/source/en/model_doc/dac.md

@@ -0,0 +1,80 @@
+<!--Copyright 2024 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.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# DAC
+
+## Overview
+
+
+The DAC model was proposed in [Descript Audio Codec: High-Fidelity Audio Compression with Improved RVQGAN](https://arxiv.org/abs/2306.06546) by Rithesh Kumar, Prem Seetharaman, Alejandro Luebs, Ishaan Kumar, Kundan Kumar.
+
+The Descript Audio Codec (DAC) model is a powerful tool for compressing audio data, making it highly efficient for storage and transmission. By compressing 44.1 KHz audio into tokens at just 8kbps bandwidth, the DAC model enables high-quality audio processing while significantly reducing the data footprint. This is particularly useful in scenarios where bandwidth is limited or storage space is at a premium, such as in streaming applications, remote conferencing, and archiving large audio datasets.
+
+The abstract from the paper is the following:
+
+*Language models have been successfully used to model natural signals, such as images, speech, and music. A key component of these models is a high quality neural compression model that can compress high-dimensional natural signals into lower dimensional discrete tokens. To that end, we introduce a high-fidelity universal neural audio compression algorithm that achieves ~90x compression of 44.1 KHz audio into tokens at just 8kbps bandwidth. We achieve this by combining advances in high-fidelity audio generation with better vector quantization techniques from the image domain, along with improved adversarial and reconstruction losses. We compress all domains (speech, environment, music, etc.) with a single universal model, making it widely applicable to generative modeling of all audio. We compare with competing audio compression algorithms, and find our method outperforms them significantly. We provide thorough ablations for every design choice, as well as open-source code and trained model weights. We hope our work can lay the foundation for the next generation of high-fidelity audio modeling.*
+
+This model was contributed by [Kamil Akesbi](https://huggingface.co/kamilakesbi).
+The original code can be found [here](https://github.com/descriptinc/descript-audio-codec/tree/main?tab=readme-ov-file).
+
+
+## Model structure
+
+The Descript Audio Codec (DAC) model is structured into three distinct stages:
+
+1. Encoder Model: This stage compresses the input audio, reducing its size while retaining essential information.
+2. Residual Vector Quantizer (RVQ) Model: Working in tandem with the encoder, this model quantizes the latent codes of the audio, refining the compression and ensuring high-quality reconstruction.
+3. Decoder Model: This final stage reconstructs the audio from its compressed form, restoring it to a state that closely resembles the original input.
+
+## Usage example 
+
+Here is a quick example of how to encode and decode an audio using this model: 
+
+```python 
+>>> from datasets import load_dataset, Audio
+>>> from transformers import DacModel, AutoProcessor
+>>> librispeech_dummy = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+
+>>> model = DacModel.from_pretrained("descript/dac_16khz")
+>>> processor = AutoProcessor.from_pretrained("descript/dac_16khz")
+>>> librispeech_dummy = librispeech_dummy.cast_column("audio", Audio(sampling_rate=processor.sampling_rate))
+>>> audio_sample = librispeech_dummy[-1]["audio"]["array"]
+>>> inputs = processor(raw_audio=audio_sample, sampling_rate=processor.sampling_rate, return_tensors="pt")
+
+>>> encoder_outputs = model.encode(inputs["input_values"])
+>>> # Get the intermediate audio codes
+>>> audio_codes = encoder_outputs.audio_codes
+>>> # Reconstruct the audio from its quantized representation
+>>> audio_values = model.decode(encoder_outputs.quantized_representation)
+>>> # or the equivalent with a forward pass
+>>> audio_values = model(inputs["input_values"]).audio_values
+```
+
+## DacConfig
+
+[[autodoc]] DacConfig
+
+## DacFeatureExtractor
+
+[[autodoc]] DacFeatureExtractor
+    - __call__
+
+## DacModel
+
+[[autodoc]] DacModel
+    - decode
+    - encode
+    - forward

src/transformers/__init__.py

@@ -312,6 +312,7 @@
         "CTRLTokenizer",
     ],
     "models.cvt": ["CvtConfig"],
+    "models.dac": ["DacConfig", "DacFeatureExtractor"],
     "models.data2vec": [
         "Data2VecAudioConfig",
         "Data2VecTextConfig",
@@ -1757,6 +1758,12 @@
             "CvtPreTrainedModel",
         ]
     )
+    _import_structure["models.dac"].extend(
+        [
+            "DacModel",
+            "DacPreTrainedModel",
+        ]
+    )
     _import_structure["models.data2vec"].extend(
         [
             "Data2VecAudioForAudioFrameClassification",
@@ -5026,6 +5033,10 @@
         CTRLTokenizer,
     )
     from .models.cvt import CvtConfig
+    from .models.dac import (
+        DacConfig,
+        DacFeatureExtractor,
+    )
     from .models.data2vec import (
         Data2VecAudioConfig,
         Data2VecTextConfig,
@@ -6450,6 +6461,10 @@
             CvtModel,
             CvtPreTrainedModel,
         )
+        from .models.dac import (
+            DacModel,
+            DacPreTrainedModel,
+        )
         from .models.data2vec import (
             Data2VecAudioForAudioFrameClassification,
             Data2VecAudioForCTC,

src/transformers/models/__init__.py

@@ -59,6 +59,7 @@
     cpmant,
     ctrl,
     cvt,
+    dac,
     data2vec,
     dbrx,
     deberta,

src/transformers/models/auto/configuration_auto.py

@@ -73,6 +73,7 @@
         ("cpmant", "CpmAntConfig"),
         ("ctrl", "CTRLConfig"),
         ("cvt", "CvtConfig"),
+        ("dac", "DacConfig"),
         ("data2vec-audio", "Data2VecAudioConfig"),
         ("data2vec-text", "Data2VecTextConfig"),
         ("data2vec-vision", "Data2VecVisionConfig"),
@@ -354,6 +355,7 @@
         ("cpmant", "CPM-Ant"),
         ("ctrl", "CTRL"),
         ("cvt", "CvT"),
+        ("dac", "DAC"),
         ("data2vec-audio", "Data2VecAudio"),
         ("data2vec-text", "Data2VecText"),
         ("data2vec-vision", "Data2VecVision"),

src/transformers/models/auto/feature_extraction_auto.py

@@ -49,6 +49,7 @@
         ("conditional_detr", "ConditionalDetrFeatureExtractor"),
         ("convnext", "ConvNextFeatureExtractor"),
         ("cvt", "ConvNextFeatureExtractor"),
+        ("dac", "DacFeatureExtractor"),
         ("data2vec-audio", "Wav2Vec2FeatureExtractor"),
         ("data2vec-vision", "BeitFeatureExtractor"),
         ("deformable_detr", "DeformableDetrFeatureExtractor"),

src/transformers/models/auto/modeling_auto.py

@@ -73,6 +73,7 @@
         ("cpmant", "CpmAntModel"),
         ("ctrl", "CTRLModel"),
         ("cvt", "CvtModel"),
+        ("dac", "DacModel"),
         ("data2vec-audio", "Data2VecAudioModel"),
         ("data2vec-text", "Data2VecTextModel"),
         ("data2vec-vision", "Data2VecVisionModel"),

src/transformers/models/dac/__init__.py

@@ -0,0 +1,60 @@
+# coding=utf-8
+# Copyright 2024 Descript and 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.
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_dac": ["DacConfig"],
+    "feature_extraction_dac": ["DacFeatureExtractor"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_dac"] = [
+        "DacModel",
+        "DacPreTrainedModel",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_dac import (
+        DacConfig,
+    )
+    from .feature_extraction_dac import DacFeatureExtractor
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_dac import (
+            DacModel,
+            DacPreTrainedModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/dac/configuration_dac.py

@@ -0,0 +1,111 @@
+# coding=utf-8
+# Copyright 2024 Descript and 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.
+"""Dac model configuration"""
+
+import math
+
+import numpy as np
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class DacConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of an [`DacModel`]. It is used to instantiate a
+    Dac 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
+    [descript/dac_16khz](https://huggingface.co/descript/dac_16khz) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        encoder_hidden_size (`int`, *optional*, defaults to 64):
+            Intermediate representation dimension for the encoder.
+        downsampling_ratios (`List[int]`, *optional*, defaults to `[2, 4, 8, 8]`):
+            Ratios for downsampling in the encoder. These are used in reverse order for upsampling in the decoder.
+        decoder_hidden_size (`int`, *optional*, defaults to 1536):
+            Intermediate representation dimension for the decoder.
+        n_codebooks (`int`, *optional*, defaults to 9):
+            Number of codebooks in the VQVAE.
+        codebook_size (`int`, *optional*, defaults to 1024):
+            Number of discrete codes in each codebook.
+        codebook_dim (`int`, *optional*, defaults to 8):
+            Dimension of the codebook vectors. If not defined, uses `encoder_hidden_size`.
+        quantizer_dropout (`bool`, *optional*, defaults to 0):
+            Whether to apply dropout to the quantizer.
+        commitment_loss_weight (float, *optional*, defaults to 0.25):
+            Weight of the commitment loss term in the VQVAE loss function.
+        codebook_loss_weight (float, *optional*, defaults to 1.0):
+            Weight of the codebook loss term in the VQVAE loss function.
+        sampling_rate (`int`, *optional*, defaults to 16000):
+            The sampling rate at which the audio waveform should be digitalized expressed in hertz (Hz).
+    Example:
+
+    ```python
+    >>> from transformers import DacModel, DacConfig
+
+    >>> # Initializing a "descript/dac_16khz" style configuration
+    >>> configuration = DacConfig()
+
+    >>> # Initializing a model (with random weights) from the "descript/dac_16khz" style configuration
+    >>> model = DacModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "dac"
+
+    def __init__(
+        self,
+        encoder_hidden_size=64,
+        downsampling_ratios=[2, 4, 8, 8],
+        decoder_hidden_size=1536,
+        n_codebooks=9,
+        codebook_size=1024,
+        codebook_dim=8,
+        quantizer_dropout=0,
+        commitment_loss_weight=0.25,
+        codebook_loss_weight=1.0,
+        sampling_rate=16000,
+        **kwargs,
+    ):
+        self.encoder_hidden_size = encoder_hidden_size
+        self.downsampling_ratios = downsampling_ratios
+        self.decoder_hidden_size = decoder_hidden_size
+        self.upsampling_ratios = downsampling_ratios[::-1]
+        self.n_codebooks = n_codebooks
+        self.codebook_size = codebook_size
+        self.codebook_dim = codebook_dim
+        self.quantizer_dropout = quantizer_dropout
+        self.sampling_rate = sampling_rate
+
+        self.hidden_size = encoder_hidden_size * (2 ** len(downsampling_ratios))
+
+        self.hop_length = int(np.prod(downsampling_ratios))
+        self.commitment_loss_weight = commitment_loss_weight
+        self.codebook_loss_weight = codebook_loss_weight
+
+        super().__init__(**kwargs)
+
+    @property
+    def frame_rate(self) -> int:
+        hop_length = np.prod(self.upsampling_ratios)
+        return math.ceil(self.sampling_rate / hop_length)