It means any dimensions preceeding the ones listed. In this context it means that
Grayscale() can process tensors such as:
A single image with shape [3, H, W] (no leading dimensions)
A batch of images of shape [B, 3, H, W] (here B is a leading dimension)
A number of batches of shape [N, B, 3, H, W] (here N and B are leading dimensions)
And the list goes on.