Deployment & configuration

Environment variables

Web app and proxy

Variable	Required	Where	Purpose
HF_TOKEN	✓	server only	Hugging Face token with read access to the private Space. Never prefix it with VITE_, or Vite embeds it in the browser bundle
VITE_SPACE_ID	—	build	Space id (owner/name); default akashch1512/SingleViewHeigthEstimation
HF_SPACE_URL	—	server	Override the Space URL derived from VITE_SPACE_ID
PORT	—	server	Port for npm run serve (default 8080)
OVERPASS_USER_AGENT	—	server	User-Agent for the Overpass relay

FastAPI service

Variable	Purpose
DW_CKPT	PyTorch checkpoint to load at import
DW_ONNX	ONNX graph to load at import (takes precedence over the checkpoint)
DW_JOBS_DIR	Job output directory (default outputs/jobs)

Deploy the web app

Vercel

Frontend/vercel.json maps the proxy routes onto serverless functions:

Frontend/vercel.json

{
  "rewrites": [
    { "source": "/hf-space/:path*", "destination": "/api/hf-space?__path=:path*" },
    { "source": "/overpass/:mirror", "destination": "/api/overpass?__mirror=:mirror" }
  ],
  "functions": {
    "api/hf-space.js": { "maxDuration": 300 },
    "api/overpass.js": { "maxDuration": 60 }
  }
}

1. Import the repository in Vercel and set Root Directory to Frontend.
2. Add HF_TOKEN (and optionally VITE_SPACE_ID) as environment variables.
3. Deploy. The build runs npm run build and serves dist/.

Docker

Frontend/Dockerfile (summary)

FROM node:22-alpine AS build     # npm ci && vite build
FROM node:22-alpine              # copies dist/ + server/, runs as user "node"
ENV NODE_ENV=production PORT=8080
CMD ["node", "server/serve.mjs"]

docker build -t depthwizard-frontend Frontend/
docker run -e HF_TOKEN=hf_xxx -p 8080:8080 depthwizard-frontend

Node

cd Frontend
npm ci && npm run build
HF_TOKEN=hf_xxx PORT=8080 npm run serve

The model Space

The hosted model is a private Gradio Space on ZeroGPU. Each call reserves a GPU slot (@spaces.GPU) for 45 s, or 120 s with TTA. GPU time is billed against the daily ZeroGPU quota of the account that owns HF_TOKEN. When it runs out, the app shows GPU quota used up.

Deploy the FastAPI service

cd Model_Traning/v5
pip install -r requirements.txt
python -m geo.dem --prime scene.tif                      # optional: cache DEMs for offline use
DW_ONNX=/models/depthwizard.onnx uvicorn serve.app:app --host 0.0.0.0 --port 8000

Caution

The service has no authentication. Run it behind a reverse proxy that handles TLS and access control, and give DW_JOBS_DIR enough disk: a georeferenced 1024² job with absolute products and meshes takes about 80 MB.

Deploy these docs

The docs site is a static Astro Starlight build in Docs-Site/:

cd Docs-Site
npm install
npm run dev       # http://localhost:4321
npm run build     # → dist/  (fails on broken internal links)

On Vercel, set Root Directory to Docs-Site. The included vercel.json selects the Astro framework preset. Node 22.12 or later is required.

CI

The repository has one GitHub Actions workflow, pr-agent.yml, an automated PR reviewer. It has no build, test or deploy workflows yet.