Allow register_kernel() to take dispatcher name as input #7796
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/7796
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New Failure, 2 Unrelated FailuresAs of commit 12fb8e1: NEW FAILURE - The following job has failed:
BROKEN TRUNK - The following jobs failed but were present on the merge base f3c89cc:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
| import torchvision.transforms.v2.functional # noqa | ||
|
|
||
| try: | ||
| return next( |
There was a problem hiding this comment.
Not the most efficient, we could store the mapping somewhere. I don't think we care anyway since this is just executed during registration, not when calling the dispatcher.
There was a problem hiding this comment.
Not sure I understand the iteration here. Can't we just do
try:
return getattr(torchvision.transforms.v2.functional, name)
except AttributeError:
raise ValueError(...) from None
There was a problem hiding this comment.
I'm ok with this.
Side note: I thought about allowing the transform name as well e.g.
@register_kernel("Resize", MyDatapoint), but I don't think we should do that because that won't work in general anyway, as some transforms rely on more than one dispatcher.
👍 for the reason you mentioned.
| import torchvision.transforms.v2.functional # noqa | ||
|
|
||
| try: | ||
| return next( |
There was a problem hiding this comment.
Not sure I understand the iteration here. Can't we just do
try:
return getattr(torchvision.transforms.v2.functional, name)
except AttributeError:
raise ValueError(...) from None
plus some minor tests for
register_kernel.Writing the tutorials in #7795 I thought that we should allow to register kernels by name, not just by callables. Typically if I'm a dev and I just care about
Resize()(the transform class), I should just be able to register my kernel with@register_kernel("resize", MyDatapoint)instead of having to do@register_kernel(resize, MyDatapoint)which forces me to import the functional dispatcher for no obvious reason.Side note: I thought about allowing the transform name as well e.g.
@register_kernel("Resize", MyDatapoint), but I don't think we should do that because that won't work in general anyway, as some transforms rely on more than one dispatcher.cc @vfdev-5