feature/issue-256-convnext-detection #262
Conversation
|
Hello @ANAMASGARD, Thanks for this contribution ! |
Implements ConvNeXt-based object detection heads using the existing Faster R-CNN pipeline (RPN + ROI heads) with ConvNeXt+FPN backbones. - Adds tiny/small/base ConvNeXt FPN backbones - Exposes model_convnext_tiny_detection(), _small_detection(), _base_detection() - Keeps output format compatible with existing Faster R-CNN models - Makes all ConvNeXt detection tests pass (TDD Phase 2) Refs mlverse#256
|
@cregouby Can you please approve and review my PR and give your feedback ! |
|
Sure, I will. Just give me some time.... |
cregouby
left a comment
cregouby
left a comment
There was a problem hiding this comment.
praise: Thanks a lot for this really nice contribution!
suggestion : would you add your person() as contributor ctb in the DESCRIPTION file ?
todo add an example and switch to conditions.R functions for messages (see inline)
|
Sir @cregouby can you please re initialize the workflow so that we can see if there errors there, |
|
Sir @cregouby can you please give feedback on how can these failing test cases be passed. |
|
Hello @ANAMASGARD, Have you tried to run the just-added example ? > topk <- pred$scores$topk(k = 5)[[2]]
> boxes <- pred$boxes[topk, ]
> boxes
torch_tensor
-0.1104 2.7817 1.3124 -4.0643
-0.2005 2.7428 1.7754 -4.1288
0.7580 3.2098 -1.7157 0.5556
0.2230 3.1610 1.3431 -3.9176
-0.0881 2.8218 0.9908 -3.9863
[ CPUFloatType{5,4} ][ grad_fn = <SliceBackward0> ]Plus looking at the normalized image, I doubt the |
…etrained detection weights
|
praise Thanks for the clarification. |
upvote the note on random bboxes restore test on error message bypass lfw test on windows runner
2 participants
Add comprehensive test suite for ConvNeXt object detection models before implementation .
Tests cover:
Status: Tests fail as expected (no implementation yet)
In the upcoming commits I will add the actual implementation.
FIXES #256