README.md

---
license: mit
library_name: mlx
pipeline_tag: audio-classification
tags:
  - mlx
  - apple-silicon
  - efficientat
  - mobilenetv3
  - audioset
  - audio-tagging
  - audio-classification
  - schism-mlx
---

# EfficientAT mn40_as_ext (AudioSet) — MLX

EfficientAT `mn40_as_ext` (MobileNetV3, width 4.0, 68.43M
params, AudioSet mAP 0.487,
[Schmid et al. 2022](https://arxiv.org/abs/2211.04772)) converted to
[MLX](https://github.com/ml-explore/mlx) for Apple Silicon. Source
checkpoint: `mn40_as_ext_mAP_487.pt` from the
[fschmid56/EfficientAT](https://github.com/fschmid56/EfficientAT) GitHub
Releases (v0.0.1). The models were distilled from PaSST transformers on
AudioSet (527 classes).

This build uses the `mn40_as_ext` weights (extended training, 300
epochs) — the best single static MobileNet in the release table
(mAP 0.487 vs 0.484 for the standard `mn40_as` run); same
architecture, so it loads into the same config-driven model class.

Runs with the open [`schism-mlx`](https://github.com/schism-audio/schism-mlx)
package:

```python
from schism_mlx.classify.efficient_at import EfficientAT

model = EfficientAT.from_pretrained("schism-audio/efficient-at-mn40-mlx")
for label, score in model.classify("song.wav", top_k=5):
    print(f"{score:.3f}  {label}")
```

Handles **arbitrary-length audio** in a single pass — the network is fully
convolutional up to the global average pool, so no windowing or truncation
is needed.

## Conversion & verification

- Weights are the original float32 checkpoint, tensor-renamed and with convs
  permuted to MLX's channels-last layout. No quantization, no finetuning.
- Verified end to end against the genuine reference implementation
  (fschmid56/EfficientAT, `models/mn` + `AugmentMelSTFT`) on synthetic
  signals and real drum recordings from
  [schism-audio/e-gmd](https://huggingface.co/datasets/schism-audio/e-gmd):
  **max logit difference 4.8e-6** with identical features (2.5e-5 end to end through the numpy frontend), identical top-5 everywhere. ~5.4 ms per 10-second clip on an M5 Max (RTF 0.0005).
- The mel frontend (pre-emphasis, 32 kHz, n_fft 1024 / win 800 / hop 320,
  128 Kaldi-style mel bins, 0–15000 Hz, `log(mel + 1e-5)`, `(x + 4.5) / 5`
  normalization) is reimplemented in pure numpy from the reference's own
  `AugmentMelSTFT` code. No torch/torchaudio needed at inference time.

**Precision note:** on M5-class GPUs, MLX silently runs float32 matmuls at
TF32-class precision ([ml-explore/mlx#3860](https://github.com/ml-explore/mlx/issues/3860)).
`schism_mlx` disables this on import; if you load these weights with your own
MLX code, set `MLX_ENABLE_TF32=0` to match the reference implementation.

## License

MIT, inherited from the source repository (code and weights). Model by
Florian Schmid, Khaled Koutini, and Gerhard Widmer (CP JKU Linz); MLX
conversion by [schism-audio](https://huggingface.co/schism-audio).

## Citation

```bibtex
@inproceedings{schmid2023efficient,
  title={Efficient Large-Scale Audio Tagging Via Transformer-to-CNN Knowledge Distillation},
  author={Schmid, Florian and Koutini, Khaled and Widmer, Gerhard},
  booktitle={ICASSP 2023},
  year={2023}
}
```