How setsuna25th3/graduation-thesis works
This is an image-based deepfake detection application. Its primary workflow scans an image folder, loads the selected GenConViT inference mode and checkpoint, classifies images, computes metrics when folder names reveal ground truth, and exports JSON results. The repository also supports training and testing separate ED and VAE variants. Source excerpts show the model and training calls; some prediction-script sections are omitted, so connections beyond those visible in the excerpts or documented in the README are not asserted.
Main entry points: train.py Training orchestration · prediction.py Prediction CLI · genconvit.py Mode dispatcher · pred_func.py Inference utilities
Components 15
Image inference
- Prediction CLI
prediction.py - Inference utilities
model/pred_func.py - Metrics and export
prediction.py
Detection models
- Model configuration
model/config.py - Mode dispatcher
model/genconvit.py - ED detector
model/genconvit_ed.py - VAE detector
model/genconvit_vae.py - Hybrid embedding
model/model_embedder.py
Model training
- Training orchestration
train.py - ED train loop
train/train_ed.py - VAE train loop
train/train_vae.py
Data and artifacts
- Image datasets
dataset/loader.py - Model checkpoints
weight - Prediction results
result
Other components
- User / image folder
Connections 19
- User / image folder → Prediction CLI: selects folder
README.md - Prediction CLI → Inference utilities: loads model
prediction.py - Prediction CLI → Model configuration: loads configuration
prediction.py - Inference utilities → Image datasets: normalizes frames
model/pred_func.py - Inference utilities → Mode dispatcher: constructs model
model/pred_func.py - Mode dispatcher → ED detector: dispatches ED
model/genconvit.py - Mode dispatcher → VAE detector: dispatches VAE
model/genconvit.py - ED detector → Hybrid embedding: uses embedding
model/genconvit_ed.py - VAE detector → Hybrid embedding: uses embedding
model/genconvit_vae.py - Prediction CLI → Metrics and export: evaluates results
prediction.py - Metrics and export → Prediction results: writes JSON
prediction.py - Training orchestration → Image datasets: loads datasets
train.py - Training orchestration → ED detector: builds ED model
train.py - Training orchestration → VAE detector: builds VAE model
train.py - Training orchestration → ED train loop: selects ED loop
train.py - Training orchestration → VAE train loop: selects VAE loop
train.py - Training orchestration → Model checkpoints: saves weights
train.py - Training orchestration → Model configuration: loads configuration
train.py - Mode dispatcher → Model checkpoints: loads weights
model/genconvit.py