Properly handle maskrcnn and keypoints w.r.t. V2 in detection references #7742
Conversation
NicolasHug
commented
Jul 13, 2023
- a bunch of cleanups and better error messages
Co-authored-by: Philip Meier <github.pmeier@posteo.de>
|
|
||
| ds = ds_fn(p, image_set=image_set, transforms=get_transform(is_train, args), use_v2=args.use_v2) | ||
| num_classes, mode = {"coco": (91, "instances"), "coco_kp": (2, "person_keypoints")}[args.dataset] | ||
| with_masks = "mask" in args.model |
There was a problem hiding this comment.
Maybe we should remove the get_masks parameter of get_coco() and just make it part of the mode parameter, i.e. allow "instances", "instances_masks", and "keypoints".
ooooor just completely change the parametrization of get_coco() to something more friendly.
In addition to that, the way we magically specify keypoints or masks (through dataset name or through model name) is pretty terrible right now.
I feel like we should address that for good, but I'd prefer doing it in another PR
There was a problem hiding this comment.
Yes, I agree that it would be better that args.dataset could parametrize data type instead of the model name.
args.dataset can be "coco, "coco_kp" and "coco_masks", for example.
|
Hey @NicolasHug! You merged this PR, but no labels were added. The list of valid labels is available at https://github.com/pytorch/vision/blob/main/.github/process_commit.py |
…n references (#7742) Summary: Co-authored-by: Philip Meier <github.pmeier@posteo.de> Reviewed By: matteobettini Differential Revision: D48642313 fbshipit-source-id: 2bf2bdbaef94dc7851646581bb922fc69fc42e77