Low accuracy on validation and training

I am training a deep learning video classifier using movinet a0 but I get 40% training accuracy and 6% val accuracy !
I am trying to classify sign language using my own dataset. I converted it to tfrecords. I have used 15 epochs and 1 epoch but the accuracy is nearly the same

This is my code

Any advice would be appreciated

I am using data from the same source for training and testing
but accuracy is still low