README.md

---
license: cc-by-4.0
library_name: timm
---

# Model card for vit_base_patch16_1024_128.audiomae_as2m

A Vision Transformer (ViT) for audio. Pretrained on AudioSet-2M with Self-Supervised Masked Autoencoder (MAE) method.

- This is a port of AudioMAE ViT-B/16 weights for usage with `timm`. The naming convention is adopted from other `timm`'s ViT models.
- See the original repo here: https://github.com/facebookresearch/AudioMAE
- For the AudioSet-20k fine-tuned checkpoint, see https://huggingface.co/gaunernst/vit_base_patch16_1024_128.audiomae_as2m_ft_as20k

NOTE: this model does not have a classification head.

## Model Details
- **Model Type:** Audio feature backbone
- **Papers:**
  - Masked Autoencoders that Listen: https://arxiv.org/abs/2207.06405
- **Pretrain Dataset:** AudioSet-2M
- **Original:** https://github.com/facebookresearch/AudioMAE

## Model Usage
### Audio Embeddings

```python
import timm
import torch
import torch.nn.functional as F
from torchaudio.compliance import kaldi

# for fine-tuning, you can pass `num_classes={your number of classes}`
model = timm.create_model("hf_hub:gaunernst/vit_base_patch16_1024_128.audiomae_as2m", pretrained=True)
model = model.eval()

MEAN = -4.2677393
STD = 4.5689974

audio = torch.randn(1, 10 * 16_000)  # make sure input is 16kHz
melspec = kaldi.fbank(audio, htk_compat=True, window_type="hanning", num_mel_bins=128)  # shape (n_frames, 128)

# AudioMAE only accepts 1024-frame input
if melspec.shape[0] < 1024:
    melspec = F.pad(melspec, (0, 0, 0, 1024 - melspec.shape[0]))
else:
    melspec = melspec[:1024]
melspec = (melspec - MEAN) / (STD * 2)

melspec = melspec.view(1, 1, 1024, 128)  # add batch dim and channel dim
output = model(melspec)  # embeddings with shape (1, 768)

# to get frame level embeddings
output = model.forward_features(melspec)  # shape (1, 513, 768)
output = output[:, 1:]  # remove [CLS] token
output = output.unflatten(1, (1024 // 16, 128 // 16))  # (1, 64, 8, 768) -> 2D patches
output = output.mean(2)  # (1, 64, 768) -> mean pooling across mel dimension
```

## Citation
```bibtex
@inproceedings{huang2022amae,
  title = {Masked Autoencoders that Listen},
  author = {Huang, Po-Yao and Xu, Hu and Li, Juncheng and Baevski, Alexei and Auli, Michael and Galuba, Wojciech and Metze, Florian and Feichtenhofer, Christoph}
  booktitle = {NeurIPS},
  year = {2022}
}
```
```bibtex
@misc{rw2019timm,
  author = {Ross Wightman},
  title = {PyTorch Image Models},
  year = {2019},
  publisher = {GitHub},
  journal = {GitHub repository},
  doi = {10.5281/zenodo.4414861},
  howpublished = {\url{https://github.com/huggingface/pytorch-image-models}}
}
```