Vec3D transition shortcomings hidden by mypy failures

[supersergiy]

@nkemnitz discovered several breaking issues introduced by the Vec3D change.
The most notable one is the breakage of the interpolate flow, addressed by #210, #209.

Several questions remain:

1. _VecND is currently not compatible with the Sequence protocol. This requires annotations like Union[Sequence[float], VecND] even in places that rely only on __len__ and __getitem__. Is there a good solution to this? fix: _VecND inherits from abc.Sequence #213
   
2. How about vec.list()/vec.as_list()/vec.to_list()? Same for tuple
3. Can we include to Vec3D in the [Docs]? It's challenging for the new users to know how to operate it.
4. Most importantly, how come the typeguard error where tensor_ops.interpolate expected sequence but got Vec3D was not caught by mypy? The error is fixed by fix: VecND support for tensor_ops.interpolate #210 , but I expect mypy to be able to catch this.

[nkemnitz]

For runtime checks you can already do Sequence.register(_VecND). Doesn't help the static type checkers, though: python/mypy#2922

Nevermind - Just inheriting from abc.Sequence works. Tried the same with typing.Sequence yesterday and got a strange pylance error

[nkemnitz]

@supersergiy: Regarding 2) Is there an advantage of this over list(vec) or tuple(vec)?

[dodamih]

So 1) is (at least checking for typing bit, not the other things that need to be implemented) fixed by the PR linked - are the other mixin functions actually needed?

For 2), I think list(vec) or tuple(vec) is clearer - it works as is.

For 3), I will write docs for VecND.

For 4), I've looked into this a little and it looks like typing.Sequence and collections.abc.Sequence being different might have something to do with it. Still looking into it.

[supersergiy]

Problem with 4 is more serious than just Sequence: https://github.com/ZettaAI/zetta_utils/pull/217/files#diff-9a1f1394cc41ddf4e36389c4dd6a230a4cef5a09496e37eeb13a294e06776068R69

At the same time, reveal_type(flow_schema) gives:

interpolate_flow.py:61: note: Revealed type is 
"zetta_utils.mazepa_layer_processing.common.chunked_apply_flow.ChunkedApplyFlowSchema[
[src: torch._tensor.Tensor, 
scale_factor: Union[builtins.float, typing.Sequence[builtins.float]], 
mode: Union[Literal['nearest'], Literal['nearest-exact'], Literal['linear'], Literal['bilinear'], Literal['bicubic'], Literal['trilinear'], Literal['area'], Literal['img'], Literal['field'], Literal['mask'], Literal['segmentation']], 
mask_value_thr: builtins.float =], zetta_utils.layer.volumetric.index.VolumetricIndex, 
None]"

but reveal_type(flow_schema.__call__): Revealed type is "def (idx: IndexT?, *args: P.args?, **kwargs: P.kwargs?) -> zetta_utils.mazepa.flows.Flow"

[supersergiy]

This used to work. I wonder if something broke our mypy plugin