How ducbao210/video_deepfake_detection works
This is a video deepfake detection system spanning dataset preparation, model training and evaluation, and interactive inference. The main workflow prepares videos as sampled frames and actor-disjoint splits, trains frame-sequence classifiers, evaluates checkpoints, and serves predictions through a FastAPI backend with a Gradio interface. The graph preserves the documented model families, knowledge distillation, EDA, and Hugging Face checkpoint support. Some pipeline links are documented in the README rather than shown in sampled source; those relationships cite the README.
Main entry points: dataset.py Frame dataset · trainer.py Standard training · dependencies.py Model loading and cache · kd_trainer.py Distillation training
Components 22
Data preparation
- Dataset acquisition
scripts/download_dataset.py - Dataset analysis
notebooks/01_eda.ipynb - Frame extraction
scripts/extract_frames.py - Actor-disjoint splits
scripts/split_dataset.py - Frame dataset
src/data/dataset.py - Image transforms
src/data/transforms.py
Model development
- Model architectures
src/models - Standard training
src/training/trainer.py - Distillation training
src/training/kd_trainer.py - Training and evaluation engine
src/training/engine.py - Optimizer and scheduler
src/training/optim.py - Model evaluation
scripts/evaluate.py - Model checkpoints
Inference services
- FastAPI service
backend/main.py - Model loading and cache
backend/dependencies.py - Video inference
scripts/inference.py - Gradio interface
frontend/app.py
Shared capabilities
- Pipeline configuration
configs/config.yaml - Metrics and plots
src/utils/metrics.py
Other components
- Video dataset
- Hugging Face Hub
- User
Connections 29
- Video dataset → Dataset acquisition: provides videos
README.md - Dataset acquisition → Frame extraction: supplies raw videos
README.md - Dataset acquisition → Dataset analysis: feeds analysis
README.md - Frame extraction → Actor-disjoint splits: precedes splitting
README.md - Actor-disjoint splits → Frame dataset: defines data splits
README.md - Frame dataset → Image transforms: applies transforms
src/data/dataset.py - Pipeline configuration → Frame dataset: configures sampling
- Pipeline configuration → Image transforms: configures preprocessing
src/data/transforms.py - Frame dataset → Standard training: supplies training data
- Frame dataset → Model evaluation: supplies evaluation data
scripts/evaluate.py - Standard training → Training and evaluation engine: runs epochs
src/training/trainer.py - Standard training → Optimizer and scheduler: uses optimization
- Distillation training → Training and evaluation engine: runs distillation steps
src/training/kd_trainer.py - Distillation training → Optimizer and scheduler: uses optimization
scripts/train_kd.py - Model architectures → Standard training: provides classifiers
scripts/train.py - Standard training → Model checkpoints: saves weights
README.md - Distillation training → Model checkpoints: saves weights
README.md - Model checkpoints → Hugging Face Hub: uploads or downloads
README.md - Hugging Face Hub → Model evaluation: provides checkpoints
README.md - Model evaluation → Metrics and plots: computes metrics
scripts/evaluate.py - User → Gradio interface: submits video
README.md - Gradio interface → FastAPI service: requests prediction
README.md - FastAPI service → Model loading and cache: loads model bundle
backend/main.py - Model loading and cache → Model architectures: builds model
backend/dependencies.py - Model loading and cache → Pipeline configuration: loads settings
backend/dependencies.py - FastAPI service → Video inference: extracts video frames
backend/main.py - Model loading and cache → Hugging Face Hub: fetches missing weights
- Video inference → Model architectures: runs prediction
scripts/inference.py - Video inference → Image transforms: preprocesses frames
scripts/inference.py