vit-aws

Outline -

• Loading ViT.
• Loading training data and after Preprocessing pushed to AWS S# bucket.
• Training ViT on AWS SageMaker.
• Torch JIT (Just In Time) Scripting, Quantizing and converting the model to Pytorch Lite Format which helps in lesser memory requirements hence efficient memory usage,fast inference time hence lower latency.
  • These optimizations are done so that the model can perform well if decided to deployed to production environment.
• Deploying the model with AWS SageMaker endpoint for public use.

Further Improvements -

• Loading the PyTorch model using ONNX format ,i.e the C++ format for more faster inference.
• Extend the deployed SageMaker endpoint with Flask APIs.
• More rigorous evaluation metrics.