lab backup (tier-1)

Archive of the lab workspace: curated cystoscopy datasets, run code (src_snapshot), final checkpoints (results / best_weights / models / weights), docs, KB, and the cascade-w34 inference server.

Layout

  • data_cystoscopy-core_<X>.tar โ€” dataset subdirs (yolo_roi, B, C, D, H, I, seg_v1, v1/v2/v3_build, usable_v1). Extract under data/cystoscopy-core/<X>/.
  • cystoscopy-core_small/ โ€” remaining dataset files (parquet manifests, checksums, README).
  • runs/ โ€” per-run src_snapshot/ (code), results/, best_weights/, models/, weights/ (final checkpoints). YOLO keeps best.pt + last.pt only.
  • inference/cascade_w34_server/ โ€” FastAPI server. Set LAB to the extract root before running.
  • writing/, kb/, lib/, prepare/, reports/ โ€” docs, notes, helpers.

Restore (any Linux box)

  1. Download this repo (huggingface-cli download milkyroad/file2 --local-dir lab_backup).

  2. mkdir -p lab/data/cystoscopy-core and extract each data_cystoscopy-core_<X>.tar into lab/data/cystoscopy-core/<X>/. Copy cystoscopy-core_small/ files into lab/data/cystoscopy-core/.

  3. Copy runs/, inference/, lib/, kb/, writing/, prepare/ to lab/.

  4. Point the server at that root:

    export LAB=/path/to/lab
    # or edit inference/cascade_w34_server/cascade_w34_server.py
    # LAB = pathlib.Path("<LAB_ROOT>")  โ†’  pathlib.Path("/path/to/lab")
    
  5. Python deps (CUDA recommended): torch, opencv-python, numpy, fastapi, uvicorn, pillow, pydantic, transformers, timm, ultralytics. See inference/cascade_w34_server/requirements.txt.

  6. Accept the gated encoder google/medsiglip-448 (Health AI Developer Foundations terms) and huggingface-cli login. Depth also pulls depth-anything/Depth-Anything-V2-Small-hf on first use.

  7. python inference/cascade_w34_server/cascade_w34_server.py (port 8001).

Excluded (on purpose)

  • Hugging Face model caches (google/medsiglip-448 and other hub snapshots).
  • cache/, __pycache__/, per-run venv/.venv.
  • data/hf/ (re-downloadable Hub datasets).
  • Upstream Marigold-v2 repo clone (re-clone).
  • Intermediate YOLO epoch*.pt; 3D reconstruction frame/video dumps (LAN only).

LAN sibling copy (full progress minus the excludes above) lives on the other Spark and is not part of this repo.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support