https://paperswithcode.com/dataset/hockey-fight-detection-dataset Papers with Code - Hockey Fight Detection Dataset Dataset Whereas the action recognition community has focused mostly on detecting simple actions like clapping, walking or jogging, the detection of fights or in general aggressive behaviors has been comparatively less studied. Such capability may be extremely usef paperswithcode.co..
https://arxiv.org/abs/2203.01305 DN-DETR: Accelerate DETR Training by Introducing Query DeNoising We present in this paper a novel denoising training method to speedup DETR (DEtection TRansformer) training and offer a deepened understanding of the slow convergence issue of DETR-like methods. We show that the slow convergence results from the instabilit arxiv.org
제목에서 언급한 전처리 과정이란 학습과정에서 이미지를 배치데이터로 만들어주는, PyTorch 식 표현으로 collate 하는 과정을 의미한다. 학습 과정시에 transforms를 확인해보면, RandomResize Augmentation이 포함됨을 알 수 있다. 입력 이미지들을 하나의 배치 데이터로 만들어주기 위해서는 사이즈(c, h, w)가 같아야한다. 그런데 RandomResize가 들어가게되면 이 h, w 가 달라질 수 있다. 그래서, DETR에서는 배치 데이터로 만들어줄때, 하나의 배치내에서 가장 큰 h, w를 갖는 이미지를 찾고 그 이미지에 맞게 zero filled image 를 생성하고 거기에 이미지를 top-left 부터 복사한다. 그럼 max size보다 작은 이미지들은 right 과 b..
실험에 사용된 프로젝트 commit id: 243fc4b1fe214ff6c27759dad51c37809db8f7f8 https://github.com/ultralytics/ultralytics GitHub - ultralytics/ultralytics: NEW - YOLOv8 🚀 in PyTorch > ONNX > CoreML > TFLite NEW - YOLOv8 🚀 in PyTorch > ONNX > CoreML > TFLite - GitHub - ultralytics/ultralytics: NEW - YOLOv8 🚀 in PyTorch > ONNX > CoreML > TFLite github.com COCO dataset에 대해 validation 을 진행하여 f1 score가 best score를..
https://github.com/IDEA-Research/DINO =1.5.0 torchvision>=0.6.0 git+https://github.com/cocodataset/panopticapi.git#egg=panopticapi scipy termcolor addict yapf timm 4. Compiling CUDA operators 단계(gpu driver version: 531.61, cuda version: 11.4, pytorch cuda version: 11.7)에서 딱히 빌드 오류는 없었고 python test.py 실행시 메모리가 부족해서 메모리 사이즈 에러는 발생했다. 어쨌든 작은 인풋 사이즈(D= 30, 32, 64, 71) 에서 돌아갔으니 빌드는 잘 된것이니 무시해줬다. 그런 다..
Crowd Counting 목적으로 구축된 ShanghaiTech 데이터셋을 보면 A part/B part 로 데이터셋이 나누어져있다. 이는 목적은 같지만 데이터의 특징이 달라서 그런건데, A, B를 합쳐서 학습을 해보고 싶다는 생각이 들었다. 기존에 데이터셋 클래스를, A part 따로 B part 따로 읽을 수 있게 구현해놨는데 이걸 합쳐서 읽을 수 있게 내가 직접 구현할까 하다가 한 번 찾아봤다. 역시나 있더라... https://stackoverflow.com/a/67163131 Pytorch - Concatenating Datasets before using Dataloader I am trying to load two datasets and use them both for training. P..
음... onnxruntime 1.13 부터인가 정확히는 모르겠는데 onnxruntime을 1.13으로 올리니 GetInputName 과 GetOutputName 이 없다고 오류난다. api가 변경됐다. Name뒤에 Allocated가 붙었다. 찾아보니 기존의 GetInputName 과 GetOutputName 이 메모리 leak 나기 쉬워서 변경됐다는 거 같다. 그래서 기존의 코드에서 GetInputName 과 GetOutputName 을 GetInputNameAllocated 과 GetOutputNameAllocated 로 바꿨는데 빌드 안된다. 반환형도 살짝 바뀌었고 또 이에 맞춰 수정은 해서 빌드는 됐는데 실행해보니 input name이랑 output name을 출력해보니 엉뚱한 값들이 출력됐다...
YOLOv5 모델은 torch.hub.load를 이용하여 Load 가능하다. https://colab.research.google.com/github/ultralytics/yolov5/blob/master/tutorial.ipynb#scrollTo=GMusP4OAxFu6 YOLOv5 Tutorial Run, share, and edit Python notebooks colab.research.google.com # YOLOv5 PyTorch HUB Inference (DetectionModels only) import torch model = torch.hub.load('ultralytics/yolov5', 'yolov5s', force_reload=True) # yolov5n - yolov5x6 or ..
https://github.com/openai/CLIP GitHub - openai/CLIP: Contrastive Language-Image Pretraining Contrastive Language-Image Pretraining. Contribute to openai/CLIP development by creating an account on GitHub. github.com CLIP 모델 파인튜닝 해보려니 1 iteration 학습하니까 바로 Loss가 nan이 뜬다. 당황해서 이슈 찾아보니 이슈가 있었다. https://github.com/openai/CLIP/issues/40 NaN values after a single gradient step · Issue #40 · openai/CLIP ..
요새 NAS(Neural Architecture Search) 마냥 real world dataset 에 대한 BS(Backbone Search)를 진행하고 있다. 실험 해보니, 백본마다 transfer learning시 성능이 크게 다르다. 대체로 모델들 끼리 성능을 봐볼때 이미지넷상에서의 성능을 척도로 이 모델이 좋다 저 모델이 좋다고 주장한다. 나아가서 detection, segmentation, transfer learning등 광범위한 실험을 근거자료로 제시한다. 근데, 실제로 real world dataset으로 이 백본, 저 백본 다 돌려보면 막상 다양한 태스크, 데이터셋에서 성능이 잘나왔다는 모델도 그렇게 성능이 안나오는 경우가 많았다. 또 어떤 백본은 freeze 시켜야 성능이 잘나오고,..
- Total
- Today
- Yesterday
- 가장 긴 증가하는 부분 수열
- 백준 11437
- 조합
- Lowest Common Ancestor
- MOT
- PyCharm
- LCA
- C++ Deploy
- ㅂ
- 단축키
- 문제집
- cosine
- 백준 11053
- 백트래킹
- 인공지능을 위한 선형대수
- 파이참
- 자료구조
- 위상 정렬 알고리즘
- 순열
- 백준 1766
- 백준
- 이분탐색
- FairMOT
일 | 월 | 화 | 수 | 목 | 금 | 토 |
---|---|---|---|---|---|---|
1 | 2 | 3 | 4 | |||
5 | 6 | 7 | 8 | 9 | 10 | 11 |
12 | 13 | 14 | 15 | 16 | 17 | 18 |
19 | 20 | 21 | 22 | 23 | 24 | 25 |
26 | 27 | 28 | 29 | 30 | 31 |