[Feature Request] Add min / max reduction

[fzimmermann89]

Motivation

For supporting einops, tensordict would need a min/max reduction.

Solution

Similiar to mean, etc. the min and max should be implemented in https://github.com/pytorch/tensordict/blob/main/tensordict/base.py by calling _cast_reduction

I did not find any issue or explaination why this is missing..

[vmoens]

on it!

[vmoens]

I'm working on this and I have a question for you:
what would be the most intuitive API:

1. min(dim=dim) returns only the values. This breaks with torch.min(tensor, dim) which returns a namedtuple but feels "natural" otherwise
2. min(dim=dim) returns a tensordict where the leaves are ("key0", "key1", "values") for the values and ("key0", "key1", "indices") for the indices
3. min(dim=dim) returns a tensordict where the leaves are ("values", "key0", "key1") for the values and ("indices", "key0", "key1") for the indices
4. min(dim=dim) returns a tensorclass with fields indices and values, where the leaves are ("key0", "key1") for each
5. A namedtuple, as it is the case with pytorch

[fzimmermann89]

If it were up to me, I would prefer .amin and .amax, both supporting multiple dimensions and only returning the value.

[vmoens]

Cool I tried to do that. LMK if these changes make sense

[fzimmermann89]

Is amin() here min().values? In torch, those two behave differently in the backward pass afaik

[vmoens]

Oh good point, I should use amin directly then