https://arxiv.org/pdf/2204.07370.pdf LSP, FLIC, MPII, COCO, AIC-HKD, CrowdedPose, PennAction, J-HMDB, HiEve, and PoseTrack mmpose 데이터셋 관련 config 파일 모음 폴더 https://github.com/open-mmlab/mmpose/tree/main/configs/_base_/datasets
https://onnxruntime.ai/docs/execution-providers/TensorRT-ExecutionProvider.html#c-api-example NVIDIA - TensorRT Instructions to execute ONNX Runtime on NVIDIA GPUs with the TensorRT execution provider onnxruntime.ai 20230725 기준 ONNXRuntime Official Docs 에 안내된 TensorRT 사용법으로는 TensorRT를 완전히 활용할 수 없는 걸로 파악된다. 여기서 완전히 활용 못한다는 의미는 Dynamic Batch 사이즈를 갖는 데이터에 대한 Inference 를 의미한다. 왜냐! 처음에 모델 Build할때 min..
https://github.com/developer0hye/onnxruntime-cuda-cpp-example/blob/main/examples/dynamic_batch_yolov8.cpp 테스트 모델: YOLOv8n Class Score 가 0.25 이상인 바운딩 박스만 출력 CPU Output size: 6 x1: 0.0581711 y1: 254.459 x2: 32.5574 y2: 324.874 score: 0.255069 class_id: 11 class_name: stopsign x1: 19.2957 y1: 230.079 x2: 793.328 y2: 756.118 score: 0.870351 class_id: 5 class_name: bus x1: 669.096 y1: 393.621 x2: 809..
onnx 모델에는 operations 뿐만아니라 metadata 도 저장할 수 있다. 이를 잘 활용하면 하나의 onnx 파일로 별도의 config 파일 없이도 손쉽게 모델을 배포할 수 있을 것이다. 이를 C++ 에서 읽어올 수 있는 방법은 아래와 같다. *onnxruntime 1.15.1 기준이다. #include #include #include #include #include #include #include #include #include #include #include #include int main(int argc, char *argv[]) { std::string model_file = "/app/models/yolov8n.onnx"; Ort::Env env(ORT_LOGGING_LEVEL_WAR..
https://developer0hye.tistory.com/593 [ONNXRuntime] Ort::detail::SessionImpl::Run 에 대하여 https://onnxruntime.ai/docs/api/c/struct_ort_1_1detail_1_1_session_impl.html#af3f7d273f5e468ef21ef0d56d48c32fd ONNX Runtime: Ort::detail::SessionImpl Struct Template Reference #include std::vector Run (const RunOptions &run_options, const cha developer0hye.tistory.com 위 글에서 ONNXRuntime Outp..
비디오 입력 모델은 대개 디코딩된 비디오로부터 일정한 간격으로 N 프레임을 샘플링하고 이 샘플링된 프레임이 입력으로 들어간다. 그런데, 비디오 디코딩은 연산량이 많이 필요한 작업중에 하나이다. 그러다보니, 학습 과정에서 매 번 비디오를 Full 디코딩하고 거기서 또 프레임을 샘플링하는 것은 데이터 전처리 과정으로써 병목을 발생시키기 쉬운 작업이다. 그래서, 이러한 문제를 완화하고자 나는 `cache`라는 폴더를 생성하고, 학습 과정에서 비디오를 Full 디코딩하면서 학습에 사용되는 프레임들을 `cache/비디오 파일 이름/%05d.jpg` 이러한 형식으로 저장하게 했다. 다시 또 디코딩이 필요한 jpg로 저장한 이유는 numpy array로 바로 저장하기란 내 PC의 용량이 충분하지 않았다. 하지만, 여..
https://github.com/pytorch/ignite/pull/1027 Issue 856 : Enable multi params group for LRScheduler by sdesrozis · Pull Request #1027 · pytorch/ignite Fixes #856 Description: LRScheduler can handle scheduler based on optimizer with one param_groups. This PR enables multi param_groups of optimizer. The main modification is the return of the method... github.com warmup_end_value 를 None로 해두면 optimize..
https://github.com/ultralytics/ultralytics/commit/8940a27bdb26895f09a1554514a9a46312aa89c3 `ultralytics 8.0.114` automatic optimizer selection (#3037) · ultralytics/ultralytics@8940a27 Co-authored-by: pre-commit-ci[bot] Co-authored-by: Burhan github.com 모델이 업데이트 되는 이터레이션수에 따라서 그 수가 큰 경우 SGD를 사용하도록, 아닌 경우 AdamW로 optimizer 가 세팅되도록 바뀌었다. 꽤나 큰 변화 같아서 기록해둔다. 실험적 결과가 있는지 코멘트를 달기는 했다.
https://github.com/LAION-AI/CLIP_benchmark/blob/main/benchmark/results.ipynb GitHub - LAION-AI/CLIP_benchmark: CLIP-like model evaluation CLIP-like model evaluation. Contribute to LAION-AI/CLIP_benchmark development by creating an account on GitHub. github.com CLIP 모델(모델이라고 하는게 맞을까...? 아무튼)은 크게 OpenAI 에서 학습된 모델과 LAION 에서 학습된 모델이 존재한다. 각 모델별로 LAION 에서 벤치마크 결과를 정리해놓았다. 위 ipynb파일을 보면된다.
import timm model = timm.create_model('vit_small_patch32_384', pretrained=pretrained, num_classes=0, img_size=[640, 640]) img_size 를 정의해주면 된다. https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/vision_transformer.py#L389 GitHub - huggingface/pytorch-image-models: PyTorch image models, scripts, pretrained weights -- ResNet, ResNeXT, EfficientNet, E PyTorch image models, scr..
- Total
- Today
- Yesterday
- 백준 11437
- FairMOT
- 순열
- 백트래킹
- 인공지능을 위한 선형대수
- 백준
- 백준 11053
- 위상 정렬 알고리즘
- C++ Deploy
- 파이참
- cosine
- PyCharm
- Lowest Common Ancestor
- 백준 1766
- 이분탐색
- 조합
- 단축키
- 자료구조
- 문제집
- LCA
- ㅂ
- MOT
- 가장 긴 증가하는 부분 수열
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 |