Pretrain

This section documents pretraining experiments (Level 0 visual, Level 1 BC, etc.). Metrics and logs differ from RL: pretrain uses epoch-based training, CSV from PyTorch Lightning (and optionally TensorBoard), and metrics such as train/val loss, overall accuracy, and per-action accuracy (e.g. accel, left+accel, right+accel, brake).

Analysis: Use scripts/analyze_pretrain_bc.py for BC pretrain runs. It reads Lightning CSV from run directories and reports loss/accuracy by epoch and per-action accuracy with human-readable names (not just class 0, 1, 2). See the command doc for full options.

Contents

• BC pretrain: training length & early stop (Level 1)
  • Experiment Overview
  • Results
  • Run Analysis
  • Detailed Metrics Analysis (by Epoch)
  • Configuration Changes
  • Hardware
  • Conclusions
  • Recommendations