IsSynth-Cars / src /README.md
DanielNguyen121's picture
Publish idempotent Cars data recovery scripts and dashboard
24b1962 verified
|
Raw History Blame Contribute Delete
1.3 kB
# IsSynth-Cars source
1. In a Molab runtime with an authorized non-commercial research purpose and `HF_TOKEN` configured, install the versions recorded in `configs/environment.json`.
2. Clone SPC commit `96834b22d5e59c852e53d713d434cbabb2f0d165` to `/marimo/SPC`. Clone Qwen3-VL-Embedding commit `393e2978d27852b0d0230d6994f37f9c15bed73c` under that source tree and remove the no-grad decorator from the embedder `forward` method as the Caltech runner does.
3. Run `python src/bootstrap_cars_data.py` to fetch authentic original Stanford train/test archives from the documented academic mirror, stage official annotations/splits, validate per-image hashes and few-shot support, and validate the existing synthetic ZIP. This keeps original images local.
4. Run `python src/download_model.py` to stage the Qwen base model, or let the runner download it automatically.
5. Run `python src/run_cars_issynth_supervisor.py`. It detects its own active process, resumes from the latest compatible checkpoint in the public model repository, and evaluates the best validation checkpoint after epoch 40.
Original images are never uploaded to the model repository. The source archive and per-image hashes, class map, support selection, and corrected validation manifest are public reproducibility artifacts.