Feat/add instance segmentation #67
Conversation
☂️ Python Coverage
Overall Coverage
New Files
Modified Files
|
Test Results 6 files 6 suites 1h 39m 49s ⏱️ Results for commit d47253a. ♻️ This comment has been updated with latest results. |
| n = len(images) | ||
|
|
||
| for i in range(n): | ||
| boxes = boxes_batch[i].tolist() | ||
| if len(boxes) == 0: | ||
| final_segments.append([]) | ||
| continue | ||
|
|
||
| inputs = self.processor( | ||
| images[i], input_boxes=[boxes], return_tensors="pt" | ||
| ).to(self.device) | ||
|
|
||
| with torch.no_grad(): | ||
| outputs = self.model(**inputs, return_dict=True) |
There was a problem hiding this comment.
SlimSAM doesn't support batched inference?
There was a problem hiding this comment.
The thing is that each image can have different number of detected objects, and in that case the batched inference isn't possible straight away , so that's why I implemented it per image. But now that you've mentioned it, I thought about it again and realized that we could "padd" the bboxes with dummy bboxes, so that we can have batch inference, I'm currrently testing it. Let me know @sokovninn, if you'd find this small hack better.
There was a problem hiding this comment.
Oh, I see. Dummy bboxes is a good solution. However, I am not sure if it will bring any boost in inference speed, but it is worth a try I think.
There was a problem hiding this comment.
Exactly, I'll test it and let you know.
There was a problem hiding this comment.
It turned out not to be faster, so not gonna use it.
examples/generate_instance_segmentation_dataset_and_train_yolo.ipynb
Outdated
Show resolved
Hide resolved
This PR includes adding the option to generate an instance-segmentation dataset.
It includes: