MasalaDosai/resnet_genre

🎵 ResNet34 Music Genre Classifier

📌 Model Description

This model uses a ResNet34 architecture trained on mel spectrograms generated from audio files to classify music genres.

Audio input is first converted into a spectrogram using librosa, then resized and passed into the CNN.

---

⚙️ Model Details

• Architecture: ResNet34
• Framework: PyTorch
• Input: Mel Spectrogram (224 × 224)
• Output: Genre class index
• File: best_resnet34_librosa.pth

---

🎧 Input Pipeline (IMPORTANT)

The model does NOT take raw audio directly.

Steps:

1. Load audio using librosa (sr=22050)
2. Convert to mel spectrogram
3. Convert to decibel scale
4. Resize to 224×224
5. Normalize
6. Feed into model

---

🧠 Classes

Replace this with your actual labels:

0 → Genre_1  
1 → Genre_2  
2 → Genre_3  
3 → Genre_4  
4 → Genre_5  
5 → Genre_6  
6 → Genre_7  
7 → Genre_8  
8 → Genre_9  
9 → Genre_10  

---

🚀 Usage

import torch
import torchvision.models as models

NUM_CLASSES = 10

model = models.resnet34()
model.fc = torch.nn.Linear(model.fc.in_features, NUM_CLASSES)

checkpoint = torch.load("best_resnet34_librosa.pth", map_location="cpu")
model.load_state_dict(checkpoint)

model.eval()

---

📊 Performance

Fill this properly:

• Train Accuracy: XX%
• Validation Accuracy: XX%

---

⚠️ Limitations

• Performance depends heavily on preprocessing consistency
• Sensitive to noise and audio quality
• Not robust to different sampling rates

---

📦 Requirements

torch
torchvision
librosa
numpy

---

💡 Notes

• This is not a Hugging Face Transformers model
• It requires custom preprocessing before inference
• For best results, match training pipeline exactly

---

🔗 Future Improvements

• Add real-time audio classification UI (Gradio)
• Improve robustness with augmentation
• Export to TorchScript / ONNX for deployment