Skip to content

DEPLOY.md

Latest commit

 

History

History
91 lines (64 loc) · 3.15 KB

DEPLOY.md

File metadata and controls

91 lines (64 loc) · 3.15 KB
Raw

Deploying the demo

Hosting reality check (August 2026)

Hugging Face Spaces no longer runs Docker or Gradio Spaces on the free tier: both now return 402 Payment Required and require a PRO subscription. Only static Spaces are free, and a static Space cannot run PyTorch server-side.

So the recommended host is Streamlit Community Cloud: free, permanent, and it runs this app from the GitHub repo with no code changes.

1. Streamlit Community Cloud (recommended, free)

Everything in the repo is already prepared. You only need to do the OAuth step, which requires your GitHub login.

  1. Go to https://share.streamlit.io and sign in with GitHub.
  2. Click Create app → Deploy a public app from GitHub.
  3. Fill in:
    • Repository: aghasalim/explainable-defect-detector
    • Branch: main
    • Main file path: app.py
  4. Open Advanced settings and set Python version: 3.12.
  5. Click Deploy.

First build takes ~5 minutes (it installs CPU-only PyTorch and downloads the WideResNet50-2 weights once). You get a permanent URL like https://explainable-defect-detector.streamlit.app.

Then tell me the URL and I will add the live badge to both READMEs.

Why the requirements pin +cpu

On linux-amd64 the default PyPI torch is the CUDA build, roughly 800 MB of nvidia-* wheels that would exhaust the free tier for a demo that never touches a GPU. requirements.txt pins torch==2.13.0+cpu against the PyTorch CPU index for that reason.

2. Docker (verified working)

The image is built and tested, it serves on port 7860 and scores every bundled sample correctly on CPU.

docker build -t defect-detector .
docker run --rm -p 7860:7860 defect-detector

Image is ~2.5 GB, mostly PyTorch plus the backbone weights baked in at build time so the first request does not stall on a 100 MB download. This runs unchanged on Render, Fly.io, Railway, or any container host, and on HF Spaces if you ever take a PRO subscription.

3. Hugging Face Spaces (requires PRO, ~$9/month)

Kept ready in case you subscribe. spaces/README.md already carries the correct YAML frontmatter (Docker SDK, port 7860).

hf auth login                                   # write token
./scripts/deploy_space.sh aghasalim

The script stages only what the Space needs and uploads via hf upload, so no git-lfs setup is required. Without PRO, hf repos create --type space --sdk docker fails with 402 Payment Required.

Updating the model artefacts

models/*.pt hold the memory bank and calibrated threshold, they are what make the image self-contained. Regenerate and refresh the demo samples with:

uv run python src/edd/export.py bottle          # -> models/bottle.pt
uv run python src/edd/verify_threshold.py       # audit the shipped threshold
uv run python src/edd/samples.py                # re-pick demo samples from real scores

samples.py chooses each sample by scoring the real test split with the exported artefact, so a _MISSED sample is a genuine false negative rather than a broken demo. If you export all 15 categories, use Git LFS:

git lfs install && git lfs track "models/*.pt"