Sync from GitHub via hub-sync

README.md

@@ -1,74 +1,42 @@
-# Enhanced Deepfake Detector
+---
+title: Deepfake Detector API
+emoji: 🔍
+colorFrom: blue
+colorTo: green
+sdk: docker
+app_port: 7860
+pinned: false
+license: mit
+---
 
-Research-oriented deepfake detection pipeline based on an LRCN + ViT hybrid with blink-aware temporal modeling and attentive adversarial training.
+# Deepfake Detector API
 
-## Features
-- Eye-region sequence extraction from videos (MediaPipe + OpenCV)
-- LRCN-style temporal model with ViT/CNN backbone
-- Adversarial training hooks (FGSM/PGD)
-- Evaluation suite (Accuracy, Precision, Recall, F1, AUC)
-- Ablation runner and explainability utilities
+FastAPI inference service for the [Enhanced Deepfake Detector](https://github.com/oyingrace/deepfake-detection) (LRCN + ViT, blink-aware temporal modeling).
 
-## Project Structure
-- `configs/` experiment and model configs
-- `src/data/` dataset preparation and preprocessing
-- `src/models/` backbones and model architecture
-- `src/train/` training and adversarial routines
-- `src/eval/` evaluation and ablation utilities
-- `src/viz/` attention map visualization helpers
-- `scripts/` local/cloud convenience runners
+> **GitHub project docs:** see [.github/README.md](.github/README.md)
 
-## Quick Start
+## Endpoints
 
-### 1. Environment (Python 3.11 only)
+| Method | Path | Description |
+|--------|------|-------------|
+| `GET` | `/health` | Service status and model load state |
+| `POST` | `/predict` | Upload a video (multipart form field `file`) |
 
-```bash
-PYENV_VERSION=3.11.11 python -m venv .venv311
-source .venv311/bin/activate
-pip install -r requirements-phase1.txt
-pip install mediapipe==0.10.21 --no-deps
-pip install absl-py attrs flatbuffers "protobuf>=4.25,<5" jax jaxlib matplotlib sounddevice
-```
-
-### 2. HuggingFace login (one-time)
-
-```bash
-hf auth login
-```
-
-### 3. Phase 1 — Stream dataset and build `.npz` files
-
-Uses HuggingFace **streaming** (`load_dataset(..., streaming=True)` + `decode(False)`): one video at a time in RAM, no 17GB zip copied into `data/processed/`. Only `.npz` files are saved locally.
+## Example
 
 ```bash
-source .venv311/bin/activate
-python -m src.data.stream_ff_dataset \
-  --out-root data/processed \
-  --metadata-csv data/metadata.csv \
-  --num-real 200 \
-  --num-fake 200
+curl https://devqueen-deepfake-server.hf.space/health
+curl -X POST https://devqueen-deepfake-server.hf.space/predict -F "file=@video.mp4"
 ```
 
-Smoke test (2+2 videos):
-
-```bash
-NUM_REAL=2 NUM_FAKE=2 bash scripts/run_local.sh
-```
-
-### 4. Phase 2 — Train (after Phase 1 completes)
-
-```bash
-pip install torch torchvision timm scikit-learn matplotlib seaborn
-python -m src.train.train --config configs/train/aat_pgd.yaml
-python -m src.eval.evaluate --checkpoint outputs/best.pt --config configs/train/aat_pgd.yaml
-```
+## Runtime variables
 
-## Datasets
-- FaceForensics++
-- Celeb-DF
+| Variable | Description |
+|----------|-------------|
+| `ALLOWED_ORIGINS` | Comma-separated frontend URLs for CORS (e.g. your Vercel app) |
+| `DEVICE` | `cpu` (default) |
 
-Use identity-disjoint splits to avoid leakage.
+## Deploy notes
 
-## Notes
-- Local machine: preprocessing/debug
-- Cloud GPU: full training and ablation sweeps
+- Synced from GitHub via `.github/workflows/sync-to-hf-space.yml`
+- Full setup guide: [huggingface/README.md](huggingface/README.md)