MAINT: work around arg{min,max} not implemented for booleans #63
Conversation
torch_np/_detail/_reductions.py
Outdated
|
|
||
| if tensor.dtype == torch.bool: | ||
| # RuntimeError: "argmax_cpu" not implemented for 'Bool' | ||
| tensor = tensor.view(torch.int8) |
There was a problem hiding this comment.
Is the equivalence guaranteed in pytorch?
There was a problem hiding this comment.
I would cast it to uint8 to be 100% sure, but yeah, I believe it is guaranteed by some work @peterbell10 did some time ago.
There was a problem hiding this comment.
A cast is guaranteed to be 0 or 1, but a view could contain any byte value.
There was a problem hiding this comment.
Oh and of course the work I did doesn't apply to the torch.compile stack.
There was a problem hiding this comment.
In this case, if viewing with uint8 we would get something that's non zero, so argmax would then work, but sure, let's do to, as the compiler should be able to optimise out the copy really
There was a problem hiding this comment.
IIRC argmax always returns the lowest index that's equal to the maximum, so the distinction does matter.
|
By cast you mean tensor.to or tensor.view? |
|
view is fine. |
ev-br
force-pushed
the
argminmax_bool
branch
from
711d3cc to
1ed3d24
Compare
|
Based on #63 (comment) and #63 (comment) let's copy to uint8 to be on the safe side. |
No description provided.