WindOneSound 300M: FSD50K Part 5 Training (Grand Finale)
Overview
This script fine-tunes the existing Part 4 Primitive Extractor checkpoint on the final FSD50K Part 5 dataset, completing the full 5-part training pipeline. The goal is to maximize the model's acoustic knowledge base and achieve peak generalization without catastrophic forgetting.
Methodology & Configuration
- Base Checkpoint:
best_fsd50k_primitive_part4.pt - Dataset: FSD50K Part 5 (7,370 training samples, 819 validation samples)
- Dual-Loss Objective:
- Weighted BCE Loss: Handles multi-label classification with dynamically calculated class weights.
- MSE Reconstruction Loss: Forces the model to rebuild the input BEATs features from the primitives, preserving the acoustic feature space.
- Total Loss:
BCE_Loss + (0.1 * MSE_Loss)
- Optimizer & Scheduler: AdamW (lr=2e-4, weight_decay=1e-4) with a 10% Linear Warmup followed by Cosine Annealing decay.
- Epochs: 10
- Mobile Optimization: Harmless PyTorch warnings are suppressed. The training loop outputs a clean, structured table after each epoch.
Training Script
(See the Python code block executed in Colab for the exact implementation)
Date: Wednesday, 8th of October