Drop barrel, add held-out stop-sign distance, add the GPU runbook

Quantise the ONNX export, and record how the model performs

Score checkpoints end to end and keep the results

Phase 3: per-format benchmark with stage timing

Document src/depth for transcription

Phase 2: distance from focal length and pixel geometry

Mirror checkpoints to Drive, and move the logic out of the notebook

Make the gradient-norm probe safe to leave on

Log train loss, val loss and gradient norm to Weights & Biases

Keep the training recipe identical to the reference notebook

Publish the dataset to the Hub and verify the round trip

Fix two cross-dataset convention bugs found by scoring against ground truth

Verify the pipeline on nuScenes v1.0-mini

Phase 1: unified data pipeline, QA and training