Welcome to the Vision Transformers (ViT) repository, dedicated to the exploration and implementation of Vision Transformer models for both Image Classification and Time Series Classification tasks.
This repository focuses on leveraging Vision Transformer architectures, originally introduced for image classification, to achieve impressive results on diverse datasets. In particular, we have fine-tuned pre-trained ViT models using ImageNet weights on the Caltech101 dataset, achieving a remarkable 94% test accuracy across 102 classes.
To get started with this repository, ensure you have the following dependencies installed:
- Python == 3.11
- PyTorch: torch == "2.1.0" & torchvision == "0.16.0"
-
Clone the Repository:
git clone https://github.com/Shyamal-Dharia/vision-transformers cd vision-transformers -
Train the ViT model for image classification
python train.py --epochs 5 --batch_size 32 --learning_rate 0.001 --data_folder "/path to your data folder/" -
Visualize the attention map
python visualize_attention_map.py --image_path "/path to your image/" --img_size 224 -
Explore Models:
- Navigate to the
modelsdirectory to access different Vision Transformer architectures (without pretrained weight). - Check out the architecture specifications and choose the one that best suits your task.
- Navigate to the
-
Train model without pretrained weights:
- Check out
examples.py, it has some examples on how to use ViT for images as well as time-series classification.
- Check out
After fine-tuning ViT models on the Caltech101 dataset, we have achieved an impressive 94% test accuracy, showcasing the effectiveness of Vision Transformers in diverse classification tasks.
Training the model for 5 epochs.
Prediction from the model: cup
Contributions to this repository are welcome! Whether you want to add new ViT architectures, share your fine-tuning results, or improve the documentation.
This project is licensed under the MIT License - see the LICENSE file for details.