Video vision transformer overfitting

[Badal-Shrestha]

I am training hybrid vision transformer model with Dense121 model in custom dataset which contains 25 classes. The model get overfit from the start, the validation loss remains constant doesn't changes.

Each class has 60 video of length 3-5 second. I am following this article for training https://keras.io/examples/vision/video_transformers/

Is there any way to solve this issues ??