AudioForge — Parameter-Efficient Audio Classification

Built and evaluated an audio-classification pipeline on FSD50K, covering 36K+ training clips across 200 multi-label sound categories.

I compared a CNN trained from scratch against a pretrained Audio Spectrogram Transformer adapted using LoRA.

The LoRA setup trained only 0.52% of the AST’s 86.8M parameters while improving mean Average Precision from 0.302 to 0.557 — an 84% improvement.

Beyond model training, I handled the GPU infrastructure end-to-end: VRAM sizing, AWS EC2 provisioning, quota management, spot-instance fallback, cost-controlled start/stop workflows, and debugging failed training runs with smoke tests before full experiments.

What I worked on

Audio Spectrogram Transformer fine-tuning

LoRA / parameter-efficient training

Multi-label audio classification

PyTorch training and evaluation pipelines

GPU memory and infrastructure optimization

Experiment debugging and validation

Stack: Python · PyTorch · Hugging Face · AWS EC2