---
library_name: transformers
tags: []
---

# ConvNext (trained on XCL from BirdSet)

ConvNext trained on the XCL dataset from BirdSet, covering 9736 bird species from Xeno-Canto. Please refer to the [BirdSet Paper](https://arxiv.org/pdf/2403.10380) and the 
[BirdSet Repository](https://github.com/DBD-research-group/BirdSet/tree/main) for further information. 

### Model Details
ConvNeXT is a pure convolutional model (ConvNet), inspired by the design of Vision Transformers, that claims to outperform them.

## How to use
The BirdSet data needs a custom processor that is available in the BirdSet repository. The model does not have a processor available. 
The model accepts a mono image (spectrogram) as input (e.g., `torch.Size([16, 1, 128, 343])`)

- The model is trained on 5-second clips of bird vocalizations.
- num_channels: 1
- pretrained checkpoint: facebook/convnext-base-224-22k
- sampling_rate: 32_000
- normalize spectrogram: mean: -4.268, std: 4.569 (from esc-50)
- spectrogram: n_fft: 1024, hop_length: 320, power: 2.0
- melscale: n_mels: 128, n_stft: 513
- dbscale: top_db: 80

```python
import torch
from transformers import AutoModelForImageClassification
from datasets import load_dataset

dataset = load_dataset("DBD-research-group/BirdSet", "HSN")
```

## Model Source
- **Repository:** [BirdSet Repository](https://github.com/DBD-research-group/BirdSet/tree/main)
- **Paper [optional]:** [BirdSet Paper](https://arxiv.org/pdf/2403.10380)

## Citation