Video Swin Transformers

This summer I participated in Google Summer of Code 2022 as a contributor, and my mentors were Luiz GUStavo Martins and Sayak Paul.

This repository presents my works on TensorFlow 2 implementations of the Video Swin Transformer models, convertion of PyTorch weights to TensorFlow 2 models and notebook for fine-tuning the models on UCF101 dataset.

You can find the final report of the project here.

Outcomes

• A collection of feature-extractor from Video Swin Transformer models on TensorFlow Hub
• Fine-tuning Video Swin Transformer with Dummy Inputs notebook for fine-tuning the Video Swin Transformer backbones after adding I3D video classification model as head
• A notebook for converting PyTorch weights to TensorFlow 2 model for your desired input shape
• Custom LearningRateSchedule for Cosine Decay with WarmUp