Download server/README.md from Manojd99/LumenX: direct link, hf CLI and curl.
- Browser
- Download file 4.74 kB
-
https://huggingface.co/spaces/Manojd99/LumenX/resolve/main/server/README.md
- Command line
-
hf download hf://spaces/Manojd99/LumenX/server/README.md
-
curl -L -o README.md https://huggingface.co/spaces/Manojd99/LumenX/resolve/main/server/README.md
LumenX Server
This folder contains the local FastAPI application, browser interface, demo gallery, and dataset evaluation tools. The default configuration uses FakeEngine for a working UI preview. The real CPU/GPU pipeline is enabled with LUMENX_ENGINE=real.
Install and launch
From the repository root in PowerShell:
core\.venv\Scripts\python.exe -m pip install -e "core[vision]" -e server
core\.venv\Scripts\python.exe server\scripts\create_demo_gallery.py
$env:LUMENX_MOCK = "1"
core\.venv\Scripts\python.exe -m uvicorn lumenx_server.main:app --app-dir server --host 127.0.0.1 --port 8000
For a real engine, install the optional model runtime, download weights, clear mock mode, and set LUMENX_ENGINE=real as described in the root README. By default the app runs locally at port 8000. Set LUMENX_SERVER_CONFIG to a YAML file to override upload limits, paths, queue depth, and retention. The engine's vision and model thresholds are configured separately in core/lumenx_core/config/default.yaml.
What the app includes
- Drag/drop upload with orientation correction, alpha preservation, 16-bit grayscale conversion, upload-size checks, and disclosed downscaling.
- Progressive SSE events with maps, candidate thumbnails, diagnostics, final result, and run retrieval.
- Before/after comparison, map overlays, a WebGL relight preview, candidate cards, and quality/timing readouts.
- Classical/model/hybrid comparison, one bounded live-camera session, synthetic local gallery, and benchmark summary view.
- Bounded single-worker processing lane, one active model instance, atomic run artifacts, and JSON error responses.
Synthetic gallery scenes are illustrative UI assets only. They are not benchmark or training data.
Evaluation workflow
Download the datasets from the source links in datasets/README.md. Do not place third-party data into version control. The evaluator discovers common Low/High and Low/Normal layouts or accepts explicit folders.
Paired LOL example:
core\.venv\Scripts\python.exe server\eval\evaluate.py --dataset LOL-v1 --root D:\datasets\LOL --paired --modes classical,dl,hybrid --ablations no_tuner,no_fusion,no_classical,no_dl,no_gain_denoise,no_recoverability,no_guardrails,no_color_correction
core\.venv\Scripts\python.exe server\eval\aggregate.py
core\.venv\Scripts\python.exe server\eval\mine_failures.py
For a dataset with no references, omit --paired. Use --input-dir and --reference-dir if automatic discovery does not match its layout. The evaluator writes resumable per-image rows, progress, paired PSNR/SSIM/LPIPS where available, no-reference IQA where available, clipping/colour/edge/noise metrics, confidence-class errors, guardrail failures, and timings under server/results/. It validates each core run before recording it. Re-run the same command to resume; existing keys are skipped. Add --save-images to retain final outputs under server/results/previews/ for manual review. Run the aggregate command again after each evaluation batch.
Measure actual API latency separately with a running server:
core\.venv\Scripts\python.exe server\eval\benchmark_server.py --images D:\datasets\LOL\eval\Low --mode hybrid --warmup 1
The first request is excluded from timing by default. Measurements include image upload, server processing, artifact publication, and response serialization. Record device, Python version, image dimensions, warm-up policy, and checkpoint hash with any reported latency.
Dataset path and split discipline
Keep training, validation, and test paths separate. Use the published test split unchanged for final reporting; do not fine-tune on benchmark references. The training script expects unpaired low-light training images and is not invoked by the evaluator. For previously unseen sets such as DICM, LIME, MEF, NPE, and VV, report no-reference and custom measurements only. Record input filenames and any resize policy.
Operational notes
GET /healthreports service readiness, engine readiness, loaded model, and lane state.GET /configreturns the public-safe configuration; absolute paths and secrets are omitted.GET /runs/{id}returns the complete result or current run status.server/runs/andserver/results/are local, ignored runtime data. Use a local disk with enough space for uploaded originals and intermediate images.- Browser camera access requires localhost or HTTPS and explicit browser permission.
See docs/API.md, docs/ARCHITECTURE.md, and docs/FAILURE_ANALYSIS.md.