Working with large Numpy arrays, is there a more efficient way? (Running out of RAM)

Question

I am extracting features using EfficientNetB1 and this results in big numpy arrays which google colab PRO with 27GB of RAM can not handle so i am wondering if there is a bet