pk1308/23f2000254-t12026-genre-classifier

Music Genre Classifier - 23f2000254-t12026

Hierarchical Vision Transformer for music genre classification using Mel spectrogram and CQT features.

Model Architecture

• Type: Hierarchical Vision Transformer (ViT)
• Input: 7-second audio segments with 2-second overlap (6 segments)
• Features:
  • Mel spectrogram (128 bins)
  • Constant-Q Transform (84 bins)
• Encoder: Factorized attention with 256 embed dim, 8 heads, 4 layers
• Temporal: Transformer encoder with CLS token aggregation

Training Details

• Dataset: 10,000 synthetic mashups (GTZAN-based)
• Classes: 10 genres (blues, classical, country, disco, hiphop, jazz, metal, pop, reggae, rock)
• Best Val F1: 0.763
• Best Val Accuracy: 76.6%

Files

• best_model.pt: Model checkpoint with weights, config, and genre mapping
• history.json: Training history