Spaces:
Running
Running
Commit History
Today
-
fix(enhance): normalize FFA-Net input like the official test code Running
FFA-Net (official net/test.py) feeds the network RGB [0,1] input normalized with mean (0.64, 0.6, 0.58) and std (0.14, 0.15, 0.152). Without it the outdoor weights washed foggy images out (gray-level std 64.8 -> 14.4 on RTTS BD_Baidu_046) and RTMDet found nothing. With it the image keeps its contrast (std 67.0) and edges sharpen. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>b123fa1 -
fix(docker): upgrade pip and pin +cpu torch wheels so the torch install step succeeds
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>0cb72a1 -
feat(detect): use MMDetection RTMDet-s as the primary detector
Detection now runs through MMDetection 3.3: COCO-pretrained RTMDet-s from the OpenMMLab model zoo, loaded with mmdet.apis.init_detector and run with inference_detector. Ultralytics YOLOv8 stays as an automatic fallback, and DETECTOR=yolo forces it. - /pipeline and /pipeline_video share one _run_detector helper. - /health reports detector_framework ("mmdetection" or "ultralytics"). - Score threshold defaults to 0.3 (DET_SCORE_THR). - torch pinned to 2.1.2 CPU, numpy < 2 and opencv-headless < 4.11, since the prebuilt mmcv 2.1.0 CPU ops are compiled against torch 2.1. - mmcv/mmengine/mmdet installed with --no-deps so they don't pull in GUI opencv-python; their runtime deps are listed in requirements.txt. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>868be2c
Yesterday
-
fix(enhance): load FFA-Net and Zero-DCE++ weights correctly
FFA-Net: the .pk file is a torch.save training checkpoint, not a pickled module. pickle.load returned the legacy magic number (an int), so startup failed with "'int' object has no attribute 'eval'". Build the official FFA-Net (gps=3, blocks=19, two post convs) and load ckpt['model']. Zero-DCE++: the app defined the original Zero-DCE network, whose layer names do not match the Zero-DCE++ checkpoint. strict=False loaded no weights, so low-light enhancement ran with random weights. Use the official depthwise-separable architecture and curve formula. Both loaders now use strict=True and fall back to CLAHE on mismatch. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>80867d6
-
feat(video): preserve audio track and fix video frame assembly
- Pass original input.mp4 as second ffmpeg input to copy audio stream - Use -c:a aac + -shortest for broad browser compatibility - Push PROJECT_SCOPE.md to repo Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>e75f322 -
feat(video): add /pipeline_video endpoint with frame-by-frame annotation and H264 output
3c7c247 -
Add HF Spaces README with Docker sdk configuration
ea7968c
-
fix(detect): raise confidence threshold 0.25→0.45 and iou 0.45 to cut false positives
713c1ee
-
fix: use /tmp/models for writable model cache, hardcode HF repo
871bbe6 -
fix: rename env vars to alphanumeric-only for HF Spaces secret naming rules
bf6ac6d -
fix(startup): always load pretrained yolov8n fallback regardless of file existence
9363b95 -
fix(docker): replace libgl1-mesa-glx with libgl1 for Debian trixie
911230f -
deploy: support all models (YOLO, FFA-Net, Zero-DCE++, Restormer, buffalo_l)
- HF_MODELS manifest covers all Phase 2-5 trained weights - _enhance() routes by condition: low-light→Zero-DCE++, fog→FFA-Net, else CLAHE - Restormer pulled directly from deepinv/Restormer on HF Hub (already public) - buffalo_l auto-downloads from InsightFace CDN as before - torch added to requirements for enhancement models; CLAHE fallback retained - /health now reports which enhancement models are loaded Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>a467a79 -
deploy: add HF Hub model download on startup
Models are now pulled from HF Hub (HF_MODEL_REPO env var) into /app/models on first Space boot instead of relying on Google Drive links. huggingface_hub added to requirements.txt. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>1e13785
-
fix
31931b6 -
add deploy/ folder — HuggingFace Spaces inference API
Standalone files to push to a HF Space repo (Docker SDK): - app.py: FastAPI matching Spring Boot InferenceClient contract - Dockerfile: port 7860, non-root appuser (HF requirements) - requirements.txt: CPU-only deps, no torch (CLAHE fallback for enhancement) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>03c4042