Code implementation for Efficient Quantized Vision Transformers for Deepfake Detection

Paper submission currently under review

• models/bitlinear.py: code implementations of the BitLinear layer, and RMSNorm

• models/dataloader.py: custom dataloader class for faster training

• models/deepfake.py: custom deepfake classifier, ViT wrapper, and EfficientNet baseline detector

• models/distill.py: code implementation for knowledge distillation with hard-label loss

• models/utils.py: utility functions for parameter stats and model size estimation

• models/vit.py: code implementation of LinViT (ViT) and BitViT models using SPT, sincos2d, and GAP

• train.py: example training/fine-tuning code

• pretrain.ipynb: example pre-training config

• finetune.ipynb: example fine-tuning config

Models:

• torch >= v2.1.0
• einops >= 0.8.0

See environment.yml for full environment used for the thesis research.

MIT