[BUG] Remove res["pi"].detach().cpu() about fugw

[pbarbarant]

Fixed by #43 and #44 .

[alexisthual]

Sorry, I'm a bit late in the game here, but I think it's important to store the final object using cpu, otherwise you might not be able to load it on machines which do not have a gpu (I might be wrong but I recall something of that kind happening to me).

[pbarbarant]

Sorry, I'm a bit late in the game here, but I think it's important to store the final object using cpu, otherwise you might not be able to load it on machines which do not have a gpu (I might be wrong but I recall something of that kind happening to me).

Wouldn't it be better to upgrade save_mapping / load_mapping utilities to by send the mapping object to the cpu by default before dumping it ?

[pbarbarant]

I'm afraid that by detaching to the cpu right after fitting, we might introduce a large memory overhead when computing barycenters.

[alexisthual]

Sorry, I'm a bit late in the game here, but I think it's important to store the final object using cpu, otherwise you might not be able to load it on machines which do not have a gpu (I might be wrong but I recall something of that kind happening to me).

Wouldn't it be better to upgrade save_mapping / load_mapping utilities to by send the mapping object to the cpu by default before dumping it ?

That sounds like a very good idea :)

[pbarbarant]

Sorry, I'm a bit late in the game here, but I think it's important to store the final object using cpu, otherwise you might not be able to load it on machines which do not have a gpu (I might be wrong but I recall something of that kind happening to me).

Wouldn't it be better to upgrade save_mapping / load_mapping utilities to by send the mapping object to the cpu by default before dumping it ?

That sounds like a very good idea :)

Reopening this issue