Efficiently storing 2D heat maps for loading at runtime in Keras

Question

I would like to train an autoencoder (with convolutional blocks) on a massive amount of heat maps using Keras. The heat maps are 2D and of size 1000 x 1000 x 1 (one color channe