Vision Transformer (ViT) – PyTorch

Scratch Implementation of Vision Transformer and training setup in PyTorch.

Setup

pip install -r requirements.txt

Train

./train.sh

or directly:

python train.py

Monitoring

TensorBoard:

tensorboard --logdir runs

Logs are saved in logs/ Checkpoints are saved in checkpoints/

Structure

• train.py – training entry point
• models/ – ViT implementation
• data/ – dataloaders & transforms
• utils/ – logging, checkpoints, TensorBoard
• train.sh – bash launcher