Pavit

🚀 Pathway Vision Transformer (PaViT)

PaViT is an image recognition model based on the Pathway Vision Transformer architecture. Inspired by Google's PaLM, it showcases the potential of few - shot learning in image recognition.

🚀 Quick Start

The model can be used for image recognition tasks by leveraging the trained weights available in the repository. You can modify the code to utilize custom datasets. Additionally, enhancing the model's performance is possible by increasing the number of self - attention heads and linear layers.

Basic Usage

#import Libraries
!pip install huggingface_hub["tensorflow"]
import matplotlib.pyplot as plt
import cv2
from huggingface_hub import from_pretrained_keras

Advanced Usage

#load model
model=from_pretrained_keras('Ajibola/PaViT')
#load image
image=cv2.imread('image_path')
image=cv2.resize(image, (224, 224)) #224 is the default image size
image=image/image.max() #Normalize the image to [0-1]
prediction=model.predict(image)
prediction=np.argmax(prediction, axis=-1) #Get Highest probability class

✨ Features

• Few - Shot Learning Potential: Inspired by Google's PaLM, PaViT demonstrates the potential of few - shot learning techniques in image recognition tasks.
• High Performance on CPU: Trained on a 4GB RAM CPU, it achieved 88% accuracy with 4 self - attention heads and 96% accuracy with 12 self - attention heads and 12 linearly stacked linear layers, even on a relatively small dataset.

📦 Installation

To use the model, you need to install the necessary libraries:

!pip install huggingface_hub["tensorflow"]

🔧 Technical Details

PaViT was trained on a 4GB RAM CPU using a dataset of 15000 Kaggle images of 15 classes. The uploaded weight was trained on an image dataset of 3 classes (Cat, Dog, and Wild animal). The model shows impressive performance and fast training speed on a CPU, despite the relatively small dataset.

🤝 Contribution

The author believes that PaViT has the potential to outperform existing Vision Transformer models and welcomes contributions from developers and other contributors. You can contribute to the project through pull requests. Developers can also report issues or suggest new features for the project.

📄 License

This project is licensed under the MIT License.