Rudraneel93/bird-sound-identifier

Bird Sound Identifier (EfficientNet-B0, Fold 3)

This repository contains a PyTorch model for bird sound classification, trained on the BirdCLEF dataset using an EfficientNet-B0 backbone. The model was trained for 30 epochs on fold 3 with strong data augmentation and class balancing.

Model Architecture

• Backbone: EfficientNet-B0 (from timm library)
• Input: 1-channel Mel spectrograms (shape: [1, 128, T])
• First Conv Layer: Modified to accept 1 channel instead of 3
• Classifier: Dropout(0.5) + Linear layer (output = number of classes)
• Loss: CrossEntropyLoss with class weights
• Optimizer: Adam (lr=1e-3, weight_decay=1e-3)
• Scheduler: ReduceLROnPlateau (patience=3, factor=0.5)
• Augmentation: SpecAugment, Gaussian noise, mixup
• Early Stopping: patience=5 epochs

Training Results (Fold 3, 30 Epochs)

Epoch	Train Loss	Train Acc	Val Loss	Val Acc
1	5.51	0.014	32.70	0.033
2	5.26	0.077	19.76	0.133
3	4.51	0.202	4.20	0.211
4	3.70	0.281	3.72	0.294
5	3.25	0.332	3.53	0.299
...	...	...	...	...
30	0.643	0.764	2.20	0.629

Final Metrics:

• Final Train Loss: 0.643
• Final Train Accuracy: 0.764
• Final Validation Loss: 2.20
• Final Validation Accuracy: 0.629

Analysis

• The model shows rapid improvement in both training and validation accuracy in the first 10 epochs, with validation accuracy rising from 3% to over 40%.
• After epoch 10, both losses decrease steadily and accuracies improve, with validation accuracy plateauing around 62-63%.
• The gap between train and validation accuracy remains reasonable, indicating no severe overfitting.
• The lowest validation loss and highest validation accuracy are achieved in the final epochs, suggesting stable convergence.

Training & Validation Curves

Usage

To load the model weights in PyTorch:

import torch
import timm
from pathlib import Path

# Load model architecture
def get_efficientnet_b0(num_classes):
    model = timm.create_model('efficientnet_b0', pretrained=False)
    in_ch = model.conv_stem.in_channels
    if in_ch != 1:
        model.conv_stem = torch.nn.Conv2d(1, model.conv_stem.out_channels, kernel_size=model.conv_stem.kernel_size,
                                          stride=model.conv_stem.stride, padding=model.conv_stem.padding, bias=False)
    in_features = model.classifier.in_features
    model.classifier = torch.nn.Sequential(
        torch.nn.Dropout(0.5),
        torch.nn.Linear(in_features, num_classes)
    )
    return model

model = get_efficientnet_b0(num_classes=...)
model.load_state_dict(torch.load('best_efficientnetb0_fold3.pth', map_location='cpu'))
model.eval()

Citation

If you use this model, please cite both the BirdCLEF dataset and this repository:

BirdCLEF Dataset:

Kahl, S., Stöter, F.-R., Thakur, A., Klinck, H., Goëau, H., Glotin, H., Vellinga, W.-P., Planqué, R., Joly, A., & Lostanlen, V. (2023). BirdCLEF 2023: Bird sound recognition in complex acoustic environments. Proceedings of the Working Notes of CLEF 2023.

This Model:

Das, R. (2026). Bird Sound Identifier (EfficientNet-B0, Fold 3) [Model]. Hugging Face. https://huggingface.co/Rudraneel93/bird-sound-identifier

Author

• Rudraneel Das

---

For questions or issues, please open an issue on the Hugging Face Hub.