Handling a large dataset consisting of npy files

I have a high number of npy files (448 files) each consisting of around 12k frames (150×150 RGB images) which together make the input to my neural network (X). However, since it is impossible to load all of the files into a single array, and the fact that it is necessary to shuffle all of the samples to avoid bias, how do I create the input and feed it to the network? Someone suggested creating a dummy array to represent indices, shuffle that, create chunks based on the array size and the indices and then feed the chunks to the neural network. However, I was wondering if there is another simpler method.
So in a word, I would like to do this step but with a high number of large npy files:

X_train_filenames, X_val_filenames, y_train, y_val = train_test_split(...)

Note1: Some suggested using TFRecords but I could not find out how to convert and use them.