Description
If I understand correctly sum(tensor_list) will allocate and keep O(N) intermediate tensors (same with a for loop) where N is number of tensors, which can be quite large in the case of big DenseNet. I propose to maybe generalize torch.add to support more than two tensors as input.
Currently one can do: functools.reduce(lambda acc, x: acc.add_(x), tensor_list, torch.zeros_like(tensor_list[0])), so it's not super-urging, but a more idiomatic, TorchScript-able way may be nice