From https://discuss.pytorch.org/t/data-type-when-change-tensor-array-to-numpy-array/9786

getsizeof internally calls the __sizeof__ method on the object, then it adds the garbage collection overhead to it and returns. Apparently __sizeof__ on PyTorch tensors doesn’t take the size of the underlying buffer into account:

torch.ones(10,10).__sizeof__()  # 40
torch.ones(100,100).__sizeof__()  # 40
np.ones((10,10)).__sizeof__() # 912
np.ones((100,100)).__sizeof__() # 80112

I can take care of this a few days from now.