File size: 7,035 Bytes
db48d32 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 | ---
license: other
license_name: non-commercial-inherited
language:
- ko
library_name: pytorch
pipeline_tag: audio-classification
tags:
- audio
- audio-classification
- sound-event-detection
- efficientat
- mobilenetv3
- edge
datasets:
- retrina0678/miimo-audio-dataset
metrics:
- accuracy
- f1
- recall
model-index:
- name: miimo-efficientat-target4
results:
- task:
type: audio-classification
name: Sound Event Classification
dataset:
name: miimo-audio-dataset (4-class, held-out test)
type: retrina0678/miimo-audio-dataset
metrics:
- type: accuracy
value: 0.9774
name: Accuracy (5-fold mean)
- type: recall
value: 0.9763
name: Macro Recall (5-fold mean)
- type: f1
value: 0.9756
name: Macro F1 (5-fold mean)
---
# miimo-efficientat-target4
[EfficientAT](https://github.com/fschmid56/EfficientAT) `mn10_as` ๋ฅผ 4๊ฐ ์ํฅ ์ด๋ฒคํธ๋ก
ํ์ธํ๋ํ ๋ชจ๋ธ. 2์ด ํด๋ฆฝ ๋จ์ ๋ถ๋ฅ์ด๊ณ ์ฃ์ง ๋๋ฐ์ด์ค(๋ผ์ฆ๋ฒ ๋ฆฌํ์ด) ๋ฐฐํฌ๋ฅผ ์ผ๋์ ๋๋ค.
- **ํด๋์ค**: `baby_cry`, `bicycle`, `glass_break`, `gunshot`
- **๋ฐฑ๋ณธ**: `mn10_as` (MobileNetV3, AudioSet ์ฌ์ ํ์ต) + MLP head
- **์
๋ ฅ**: 32,000 Hz mono, 2.0์ด (64,000 ์ํ) โ AugmentMelSTFT 128 mel
- **ํ์ต ๋ฐ์ดํฐ**: [retrina0678/miimo-audio-dataset](https://huggingface.co/datasets/retrina0678/miimo-audio-dataset)
- **์ฒดํฌํฌ์ธํธ ํฌ๊ธฐ**: 17MB
## ์ฑ๋ฅ (held-out test, 5-fold)
| metric | mean | std | min | max |
|---|---|---|---|---|
| accuracy | **0.9774** | 0.0113 | 0.9617 | 0.9909 |
| balanced accuracy | 0.9763 | 0.0113 | 0.9618 | 0.9903 |
| macro precision | 0.9758 | 0.0123 | 0.9594 | 0.9906 |
| macro recall | 0.9763 | 0.0113 | 0.9618 | 0.9903 |
| macro F1 | 0.9756 | 0.0122 | 0.9592 | 0.9904 |
### ํด๋์ค๋ณ recall (5-fold)
| ํด๋์ค | mean | std | ๋น๊ณ |
|---|---|---|---|
| `bicycle` | 0.9980 | 0.0046 | |
| `baby_cry` | 0.9861 | 0.0113 | |
| `gunshot` | 0.9849 | 0.0070 | |
| `glass_break` | **0.9362** | 0.0314 | โ ๏ธ ๊ฐ์ฅ ์ฝํจ โ ์ฃผ๋ก `gunshot` ์ผ๋ก ์ค๋ถ๋ฅ |
์ ์ฒด 5-fold pooled confusion matrix์์ `glass_break` โ `gunshot` ์ค๋ถ๋ฅ๊ฐ 36๊ฑด์ผ๋ก
๊ฐ์ฅ ํฐ ์ค์ฐจ ์์ธ์ด๋ค. ํ์ด์ ๊ณ์ด์ด๋ผ 2์ด ์ฐฝ์์ ํผ๋๋๋ ๊ฒ์ผ๋ก ๋ณด์ธ๋ค.
### ์ถ์ฒ๋ณ recall
| ํด๋์ค | ์ถ์ฒ | recall | n |
|---|---|---|---|
| baby_cry | AI Hub (์ฆ๊ฐ) | 0.9744 | 585 |
| baby_cry | ESC-50 (์ฆ๊ฐ) | 1.0000 | 305 |
| baby_cry | donateacry | 1.0000 | 190 |
| bicycle | AI Hub (์ฆ๊ฐ) | 0.9980 | 490 |
| glass_break | AI Hub (์ฆ๊ฐ) | 0.9362 | 580 |
| gunshot | AI Hub (์ฆ๊ฐ) | 0.9849 | 595 |
## ํ์ผ ๊ตฌ์ฑ
```
best_model.pt ์ต์ข
์ฒดํฌํฌ์ธํธ (fold 2, best by macro_recall)
config.json ํ์ต ํ์ดํผํ๋ผ๋ฏธํฐ
labels.json ํด๋์ค ์์ / label2id
all_metrics.json fold๋ณ ์ ์ฒด ์งํ
kfold_summary.csv fold๋ณ ์์ฝ
recall_by_source.csv ์ถ์ฒ๋ณ recall
report.md ์์ธ ๋ฆฌํฌํธ (ํผ๋ํ๋ ฌ ํฌํจ)
fold_01..05/ fold๋ณ best.pt + ์์ธกยทํผ๋ํ๋ ฌ
```
`best_model.pt` ๋ dict์ด๊ณ ํค๋ ๋ค์๊ณผ ๊ฐ๋ค:
`model_state_dict`, `config`, `labels`, `label2id`, `id2label`, `fold`, `epoch`, `val_metrics`
## ์ฌ์ฉ๋ฒ
EfficientAT ์ ์ฅ์๊ฐ ํ์ํ๋ค.
```bash
git clone https://github.com/fschmid56/EfficientAT
pip install torch torchaudio huggingface_hub
```
```python
import sys, torch, torchaudio
from huggingface_hub import hf_hub_download
sys.path.insert(0, "EfficientAT")
from models.mn.model import get_model as get_mn
from models.preprocess import AugmentMelSTFT
from helpers.utils import NAME_TO_WIDTH
ckpt_path = hf_hub_download("retrina0678/miimo-efficientat-target4", "best_model.pt")
ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=False)
labels = ckpt["labels"] # ['baby_cry','bicycle','glass_break','gunshot']
model = get_mn(num_classes=len(labels), pretrained_name="mn10_as",
width_mult=NAME_TO_WIDTH("mn10_as"), head_type="mlp")
model.load_state_dict(ckpt["model_state_dict"])
model.eval()
# ํ์ต๊ณผ ๋์ผํ ์ ์ฒ๋ฆฌ โ freqm/timem ์ ์ถ๋ก ์ 0 ์ผ๋ก ๋๋ค
mel = AugmentMelSTFT(n_mels=128, sr=32000, win_length=800, hopsize=320, freqm=0, timem=0)
mel.eval()
wav, sr = torchaudio.load("clip.wav")
if sr != 32000:
wav = torchaudio.functional.resample(wav, sr, 32000)
wav = wav.mean(0, keepdim=True)[:, :64000] # mono, 2์ด
with torch.no_grad():
logits = model(mel(wav).unsqueeze(1))
if isinstance(logits, (tuple, list)):
logits = logits[0]
probs = logits.flatten(1).softmax(-1)[0]
for lbl, p in sorted(zip(labels, probs.tolist()), key=lambda x: -x[1]):
print(f"{lbl:<12} {p:.4f}")
```
## ํ์ต ์ค์
| ํญ๋ชฉ | ๊ฐ |
|---|---|
| ๋ฐฑ๋ณธ | `mn10_as` (AudioSet ์ฌ์ ํ์ต) |
| head | `mlp` |
| epochs / batch | 20 / 16 |
| optimizer | lr 1e-4, weight decay 1e-4 |
| freeze backbone | ์ 2 epoch |
| CV | StratifiedGroupKFold 5-fold, test_size 0.2 |
| group column | `group_id` (= `aug_<source_file>`) |
| selection metric | `macro_recall` |
| class weight | balanced + weighted sampler |
| seed | 42 |
| AMP | on |
**๋ฐ์ดํฐ ๋์ ๋ฐฉ์ง**: ๊ฐ์ ์๋ณธ ์์์์ ๋์จ ํด๋ฆฝ์ 50% ์ค๋ฒ๋ฉ + ์ฆ๊ฐ์ผ๋ก ์๋ก ๊ฒน์น๋ฏ๋ก,
ํด๋ฆฝ์ด ์๋๋ผ `source_file` ๋จ์(`group_id`)๋ก ๋ถํ ํ๋ค.
## ํด๋์ค ๊ท ํ
์ต์ ํด๋์ค(`bicycle` 3,747)์ ๋ง์ถฐ ํด๋์ค๋น 3,747๊ฐ๋ก downsample โ ์ด 14,988 ํด๋ฆฝ.
## ํ๊ณ
- **2์ด ๊ณ ์ ์ฐฝ**์ด๋ผ ๊ทธ๋ณด๋ค ๊ธด ์ด๋ฒคํธ์ ๋ฌธ๋งฅ์ ๋ชป ๋ณธ๋ค.
- **`glass_break` recall์ด 0.936**์ผ๋ก ๊ฐ์ฅ ๋ฎ๊ณ fold ๊ฐ ํธ์ฐจ(std 0.031)๋ ํฌ๋ค. `gunshot` ๊ณผ์ ํผ๋์ด ์ฃผ ์์ธ.
- ํ์ต ๋ฐ์ดํฐ๊ฐ **AI Hub ์ค์ฌ**์ด๋ผ ๋
น์ ํ๊ฒฝ์ด ๋ค๋ฅธ ์ค์ฌ์ฉ ํ๊ฒฝ์์๋ ์ฑ๋ฅ์ด ๋จ์ด์ง ์ ์๋ค. `recall_by_source.csv` ์ฐธ๊ณ .
- 4๊ฐ ํด๋์ค๋ง ๋ค๋ฃจ๊ณ , ๊ทธ ์ธ ์๋ฆฌ๋ **๊ฐ์ฅ ๊ฐ๊น์ด ํด๋์ค๋ก ๊ฐ์ ๋ถ๋ฅ**๋๋ค. ์ค์ฌ์ฉ ์ ํ๋ฅ ์๊ณ๊ฐ(threshold)์ ๋๊ณ reject ์ฒ๋ฆฌํ๋ ๊ฒ์ ๊ถํ๋ค.
## ๋ผ์ด์ ์ค
โ ๏ธ **์์
์ ์ด์ฉ ๋ถ๊ฐ.**
ํ์ต ๋ฐ์ดํฐ์ [ESC-50](https://github.com/karolpiczak/ESC-50)(**CC BY-NC 3.0**)๊ณผ
AI Hub ์ ๊ณต ๋ฐ์ดํฐ(**์ฌ๋ฐฐํฌ ์ ํ**)๊ฐ ํฌํจ๋์ด ์๋ค. ์ด ๊ฐ์ค์น๋ ํด๋น ๋ฐ์ดํฐ์์
ํ์๋์์ผ๋ฏ๋ก ์ ๋ฐ์ดํฐ์ ์ ์ฝ์ ๊ทธ๋๋ก ์น๊ณํ๋ค.
- ์์
์ ์ฉ๋๋ก๋ ์ฌ์ฉํ ์ ์๋ค.
- ๋ฐฑ๋ณธ `mn10_as` ์์ฒด์ ๋ผ์ด์ ์ค๋ [EfficientAT ์ ์ฅ์](https://github.com/fschmid56/EfficientAT)๋ฅผ ๋ฐ๋ฅธ๋ค.
- ์์ธํ ์ถ์ฒ๋ [๋ฐ์ดํฐ์
์นด๋](https://huggingface.co/datasets/retrina0678/miimo-audio-dataset)๋ฅผ ์ฐธ๊ณ .
## ์ธ์ฉ
```
EfficientAT: F. Schmid, K. Koutini, G. Widmer.
"Efficient Large-Scale Audio Tagging via Transformer-to-CNN Knowledge Distillation." ICASSP 2023.
ESC-50: K. J. Piczak. "ESC: Dataset for Environmental Sound Classification." ACM MM 2015.
```
|