Skip to content

Latest commit

 

History

History
26 lines (22 loc) · 2.21 KB

dev_scratchpad.md

File metadata and controls

26 lines (22 loc) · 2.21 KB

Dev Scratchpad

Runtime overview

  • Entry point main.py spins up FastAPI with lifespan that loads three fastai learners. Run by python main.py or uvicorn main:app --host 0.0.0.0 --port 9020.
  • Async endpoints call FastAI Learner.predict via run_in_threadpool; uploads arrive as UploadFile.
  • Upload guardrails: MIME whitelist (jpeg/png/tiff), 10 MB cap enforced via seek (no double read), centralized logging + exception hooks.
  • Preprocessing: imageio.v3.imread -> drop alpha -> exposure.rescale_intensity(..., out_range=np.float64) -> img_as_ubyte -> PILImage.create.
  • /healthz reports readiness (ok vs degraded) based on which learners are loaded.

Models & data

  • Models expected at models/xtype-simple_resnet18_fp16_01, models/lateral_fliprot_resnet18_fp16_07, models/frontal_fliprot_resnet18_fp16_03 (fastai exported learners). They are treated as files (not dirs); ensure git-lfs or manual drop.
  • label_func stub exists solely to satisfy learner export requirements (fastai dataloaders reference it).

API surface

  • / welcome ping.
  • /xray-info chained classification: global type + optional flip/rotation for lateral/frontal (uses map_fliprot_prediction).
  • /xray-class type-only; /lateral-fliprot + /frontal-fliprot call classify_specific_model directly.
  • All endpoints reuse validate_file_size + validate_image; responses include raw prediction, probability, full vocab w/ probabilities.

Operational notes

  • Logging now stdout-only (single StreamHandler), container-friendly.
  • CORS wide open.
  • Dockerfile: multi-stage (python:3.13-slim base) installs torch CPU wheels, fastai, scikit-image, imageio, fastapi. Models copied via scratch stage; BFD9020.html copied for /test route.
  • Compose file (docker-compose.yml) builds from repo root, publishes 9020:9020, sets LOG_LEVEL=INFO.
  • Tester available at /test (serves BFD9020.html); standalone file also works locally. TIFF thumbnails are rendered via <canvas> using vendored /static/pako.min.js + /static/UTIF.js (FastAPI mounts StaticFiles(directory="static")).
  • No requirements.txt/pyproject—deps inferred from image setup (fastapi, fastai, torch/vision/audio, scikit-image, imageio, numpy, uvicorn).