Metadata-Version: 2.1
Name: dasheng
Version: 0.0.1
Author-email: Heinrich Dinkel <dinkelheinrich@xiaomi.com>, Junbo Zhang <zhangjunbo1@xiaomi.com>
Maintainer-email: Heinrich Dinkel <dinkelheinrich@xiaomi.com>, Junbo Zhang <zhangjunbo1@xiaomi.com>
License: MIT License
        
        Copyright (c) 2024 Xiaomi
        
        Permission is hereby granted, free of charge, to any person obtaining a copy
        of this software and associated documentation files (the "Software"), to deal
        in the Software without restriction, including without limitation the rights
        to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
        copies of the Software, and to permit persons to whom the Software is
        furnished to do so, subject to the following conditions:
        
        The above copyright notice and this permission notice shall be included in all
        copies or substantial portions of the Software.
        
        THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
        IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
        FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
        AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
        LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
        OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
        SOFTWARE.
        
Project-URL: Homepage, https://github.com/Richermans/dasheng
Project-URL: Documentation, https://github.com/Richermans/dasheng
Project-URL: Repository, https://github.com/Richermans/dasheng
Project-URL: Issues, https://github.com/Richermans/dasheng/issues
Classifier: Development Status :: 3 - Alpha
Classifier: Intended Audience :: Developers
Classifier: Topic :: Software Development :: Build Tools
Classifier: License :: OSI Approved :: MIT License
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.8
Classifier: Programming Language :: Python :: 3.9
Classifier: Programming Language :: Python :: 3.10
Classifier: Programming Language :: Python :: 3.11
Classifier: Environment :: GPU :: NVIDIA CUDA :: 11.4
Classifier: Environment :: GPU :: NVIDIA CUDA :: 12
Classifier: Topic :: Multimedia :: Sound/Audio :: Speech
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: einops
Requires-Dist: numpy
Requires-Dist: pytorch_ignite
Requires-Dist: torch
Requires-Dist: torchaudio
Provides-Extra: train
Requires-Dist: accelerate>=0.28.0; extra == "train"
Requires-Dist: bitsandbytes>=0.35.4; extra == "train"
Requires-Dist: webdataset>=0.2.86; extra == "train"
Requires-Dist: braceexpand>=0.1.7; extra == "train"
Requires-Dist: fire>=0.5.0; extra == "train"
Requires-Dist: loguru>=0.7.2; extra == "train"
Requires-Dist: numpy>=1.24.1; extra == "train"
Requires-Dist: pytorch_ignite>=0.4.13; extra == "train"
Requires-Dist: PyYAML>=6.0.1; extra == "train"
Requires-Dist: torch>=2.1.1; extra == "train"
Requires-Dist: torchaudio>=2.1.1; extra == "train"
Requires-Dist: tqdm>=4.66.1; extra == "train"
Requires-Dist: pandas>=2.0; extra == "train"
Provides-Extra: all
Requires-Dist: dasheng[train]; extra == "all"

<div align="center">
    <h1>
    Dasheng (大声)
    </h1>
    <p>
    Official PyTorch code for <b>D</b>eep <b>A</b>udio-<b>S</b>ignal <b>H</b>olistic <b>E<\b>mbeddi<b>ng</b>s <br>
    <b><em>Scaling up masked audio encoder learning for general audio classification</em></b>
    </p>
    <p>
    <img src="src/logo.png" alt="Dasheng Logo" style="width: 240px; height: 240px;">
    </p>
    <p>
    (Logo generated by DALL·E 3)
    </p>
    <a href="https://github.com/richermans/dasheng"><img src="https://img.shields.io/badge/Platform-linux-lightgrey" alt="version"></a>
    <a href="https://github.com/richermans/dasheng"><img src="https://img.shields.io/badge/Python-3.8+-orange" alt="version"></a>
    <a href="https://github.com/richermans/dasheng"><img src="https://img.shields.io/badge/PyTorch-1.13+-brightgreen" alt="python"></a>
    <a href="https://github.com/richermans/dasheng"><img src="https://img.shields.io/badge/License-MIT-red.svg" alt="mit"></a>
</div>

# TL;DR


```
pyton3 -m pip install dasheng
python3 -c "from dasheng import dasheng_base; import torch; model = dasheng_base().eval(); features=model(torch.randn(1, 16000))"
```


This repo provides checkpoints for the Interspeech 2024 paper `Scaling up masked audio encoder learning for general audio classification`.
The goal of this work is to investigate the scalability of masked autoencoders for audio.
Prior work did not scale beyond 10,000 hours of audio, while Dasheng used 272,000 hours of training data.



# Models

Dasheng models have been trained on 272k hours of general audio, mainly [VGGSound](https://www.robots.ox.ac.uk/~vgg/data/vggsound/), [Audioset](https://research.google.com/audioset/), [MTG-Jamendo](https://mtg.github.io/mtg-jamendo-dataset/) and [ACAV100M](https://acav100m.github.io/).

Models with their linear evaluation results on the [HEAR benchmark](https://hearbenchmark.com/), averaged across different domains.

| Model | Parameters (M) | Environment Sounds | Speech  | Music |
|------|-------|-------|-------| ------ |
| Dasheng-Base| 86   | 80.2 | 72.5 | 84.0 |
|Dasheng-0.6B | 600    | 82.4 | 74.9 | 84.0 |
| Dasheng-1.2B | 1200    | **83.2** | **75.7** | **84.9** | 
| [AudioMAE](https://github.com/facebookresearch/AudioMAE) | 86 | 61.7 | 38.7 | 72.7 |
| [Whisper-Base-V1](https://github.com/openai/whisper) | 74 | 52.5 | 73.1 | 69.1 |
| [WavLM-Large](https://github.com/microsoft/unilm/tree/master/wavlm) |  330 | 71.4 |  72.2 | 65.0 |
| [Wav2vec-large-100k-voxpopuli](https://huggingface.co/facebook/wav2vec2-large-100k-voxpopuli) | 300 | 62.5	| 63.6 | 69.5 | 
| [Data2Vec-Audio-Large](https://huggingface.co/facebook/data2vec-audio-large) | 300 |41.1 |  60.5 | 55.0 | 

<img src="metadata/hear_capabilities.png" alt="Hear capabiltiies" style="width: 200px; height: 200px;">

## K-Nearest Neighbor results

Performance of features without parameterized training.

|                          | ESC50 | FSDKaggle18 | NSynth Instrument | Speech Commands 1  | Speech Commands 2   | US8k  | VoxCeleb1 | RAVDESS-Speech | FluentSpeechCommands   |
|--------------------------|-------|--------|-------------|-------|-------|-------|-----------|---------|-------|
| [MSM-MAE](https://github.com/nttcslab/msm-mae) | 2     | 2.18   | 20.58       | 3.7   | 1.5   | 11.5  | 0.12      | 6.77    | 1.85  |
| MelSpec                  | 18.4  | 38.5   | 35.5        | 3.7   | 1.5   | 40.39 | 5.26      | 29.65   | 9.97  |
| [CED-Base](https://github.com/RicherMans/CED)                      | 95.35 | 85.06  | 74.41       | 79.78 | 62.66 | 87.06 | 7.02      | 52.78   | 16.61 |
| [AudioMAE](https://github.com/facebookresearch/AudioMAE) | 53.05 | 43.38  | 67.21       | 56.87 | 5.9   | 58.18 | 2.9       | 28.68   | 7.59  |
| [WavLM-Large](https://github.com/microsoft/unilm/tree/master/wavlm) | 51.3  | 60.87  |             | 96.97 | 92.69 | 58.67 | 28.54     | 51.39   | 83.28 |
| [Wav2vec-large-100k-voxpopuli](https://huggingface.co/facebook/wav2vec2-large-100k-voxpopuli) | 44    | 59.5   | 60.42       | 80.86 | 66.61 | 59.84 | 18.22     | 45.76   | 30.48 |
| Dasheng-Base             | 61.9  | 70.31  | 70.02       | 93.55 | 86    | 73.87 | 34.21     | 58.12   | 52.33 |
| Dasheng-0.6B             | 66.55 | 72.06  | 70.87       | 93.36 | 87.27 | 75.92 | 37.78     | 61.81   | 57.63 |
| Dasheng-1.2B             | 68.55 | 72.06  | 71.19       | 95.9  | 90.9  | 77.71 | 39.39     | 61.94   | 62.38 |


## 1. Installation (Recommended for inference)

Install the package.

```bash
python3 -m pip install dasheng
```


### 1.2 Installation for Training

```bash
python3 -m pip install dasheng[train]
```

## 2. Usage


```python
# The three models of the paper
from dasheng import dasheng_base, dasheng_06B, dasheng_12B

model = dasheng_base()
```

Forward some audio data (note should be 16khz)

```python
import torch
model = model.eval()
features = model(torch.randn(1, 16000))
print(features.shape)
```


## 3. Training

Install dependencies:

```bash
python3 -m pip install dasheng[train]
```

### 3.1 Prepare data



We rely on the excellent (webdataset)[https://github.com/webdataset] library for I/O.
Thus one simply needs to pack their data into a bunch of `.tar` files.

A simple example of such a file would be:
```bash
find DIR -type f -name '*flac' |  tar -rvf data.tgz -T -
```

We also provide a simple script [wavlist_to_tar] that automates this process, which is installed with the package.

```bash
wavlist_to_tar your_data.tsv shards/
```

Creating `your_data.tsv` is simple:

```
find data -type f  | awk 'BEGIN{"filename"}{print}' > your_data.tsv
```


###  3.2 Training from source

To train one should first adjust the config in `dasheng/train/config/*yaml` accordingly, by adding their training data.

```bash
python3 dasheng/train/train.py dasheng/train/config/dasheng_base.yaml
```


MultiGPU support is realized using [Accelerate](https://huggingface.co/docs/accelerate/index)

```bash
accelerate launch --mixed_precision='bf16' dasheng/train/train.py dasheng/train/config/dasheng_base.yaml
```



## Citation

```
@inproceedings{dinkel2024dasheng,
  title={Scaling up masked audio encoder learning for general audio classification},
  author={Dinkel, Heinrich and Yan, Zhiyong and Wang, Yongqing and Zhang, Junbo and Wang, Yujun and Wang, Bin},
  booktitle={Interspeech 2024},
  year={2024}
}
```
