What is the most efficient way to load a large numpy `.npy` file as PyTorch dataset?

Question

I have a large numpy .npy file, with shape like (100000, 200, 200) for ConvNet. The first dimension is sample size and the last two dimensions can