What about having torch.IntTensor(10).random_(4) returning integers between 0 and 3 instead of 1 and 4, to make it consistent with numpy.random.randint?