How ouc-yer/sepl-deepfakedetection works
DeepfakeBench’s SePL implementation trains and evaluates a face-forgery detector. The principal workflow is dataset sampling and image preparation, SePL prediction and objective computation, then optimization and metric-based evaluation; the included t-SNE script provides a separate feature-analysis workflow. README documents cross-dataset testing and pretrained-weight evaluation, but the supplied tree omits the referenced test entry point and configuration files. Several internal calls are therefore shown only where supported by the excerpts.
Main entry points: train.py Training entry · sepl_detector.py SePL detector · trainer.py Trainer · abstract_dataset.py Dataset loader
Components 14
Data preparation
- Dataset loader
DeepfakeBench/training/dataset/abstract_dataset.py - Forgery augmentation
DeepfakeBench/training/dataset
SePL detection
- SePL detector
DeepfakeBench/training/detectors/sepl_detector.py - Prompt modules
DeepfakeBench/training/detectors/sepl_modules.py - Alignment objectives
DeepfakeBench/training/loss/sepl_loss.py
Training and evaluation
- Training entry
DeepfakeBench/training/train.py - Trainer
DeepfakeBench/training/trainer/trainer.py - Optimizers
DeepfakeBench/training/optimizor - Detection metrics
DeepfakeBench/training/metrics/base_metrics_class.py - Training logger
DeepfakeBench/training/logger.py
Feature analysis
- Feature analysis
DeepfakeBench/analysis/tsne_sepl.py
Other components
- Researcher
- Face datasets
- Pretrained CLIP
Connections 18
- Researcher → Training entry: starts training
- Face datasets → Dataset loader: provides samples
- Training entry → Dataset loader: prepares loader
- Dataset loader → Forgery augmentation: uses transforms
- Dataset loader → Training entry: yields batches
- Training entry → SePL detector: constructs model
- SePL detector → Prompt modules: uses prompts
- SePL detector → Alignment objectives: computes objectives
- SePL detector → Pretrained CLIP: loads pretrained model
- Training entry → Trainer: starts training
- Training entry → Optimizers: selects optimizer
- Trainer → SePL detector: runs batches
- Trainer → Optimizers: updates parameters
- Trainer → Detection metrics: records scores
- Trainer → Training logger: logs progress
- Researcher → Feature analysis: requests analysis
- Feature analysis → Dataset loader: prepares test loaders
- Feature analysis → SePL detector: extracts features