
Update: 02/18/2024 https://huggingface.co/docs/transformers/v4.20.1/en/perf_train_gpu_many#:~:text=DP%20copies%20data%20within%20the,not%20the%20case%20with%20DP. ————/ DDP를 써라! 핵심은 DP는 single-process multi-thread이고, DDP는 multi-process라는 것. python GIL overhead를 피하려면 multi-thread가 아닌 multi-process를 쓰는게 효율적임.https://pytorch.org/docs/master/notes/cuda.html#use-nn-parallel-distributeddataparallel-in..

Update 2021.12.27 torch로 train할 때 segmentation fault (core dump) 버그가 뜨는데, osmesa import할 때만 떴음. 원인은 아직 불명. Update 2021.08.16 the problem was osmesa. egl has no problem os.environ['PYOPENGL_PLATFORM'] = 'egl' but normally osmesa has no bugs... what should I do python debugger sometimes gives segmentation fault. Most of cases seem to be related to dependencies between python libraries. In my case,..

전에 코드는 잘이해안갔는데, 이건 바로 이해감. https://github.com/zju3dv/neuralbody/blob/master/lib/utils/if_nerf/if_nerf_data_utils.py def get_near_far(bounds, ray_o, ray_d): """calculate intersections with 3d bounding box""" norm_d = np.linalg.norm(ray_d, axis=-1, keepdims=True) viewdir = ray_d / norm_d viewdir[(viewdir -1e-10)] = 1e-5 viewdir[(viewdir > -1e-5) & (viewdir < 1e-10)] = -1e-5 tm..

사람을 중심으로 camera 위치를 변형해 novel view synthesis를 하는 일을 하고 있다. 이 때 camera 위치 변화에 따라 camera parameter를 어떻게 바꿀지가 중요한데, 이 포스트는 코드와 함께 개념을 되짚어본다. ㄹ그림에서 보듯이 world coordinate 상에 사람이 저렇게 있을 때 camera의 위치를 높이는 안바꾸고 z축 기준으로만 회전시켜 novel view를 얻는 게 내 목표다. (왼쪽에 있는 빨간 점이 원래 카메라 위치, 오른쪽에 있는 빨간 점이 바뀐 카메라 위치) 카메라의 orientation은 당연히 z축이 사람을 향해 있고 (이건 가정이 아니고 사실), y축은 아래로, x 축은 오른쪽을 향해 있다고 가정하자. (내 가정이 본인 데이터셋과 맞는지는 본인..

초기실험들. Neuralbody command python train_net.py --cfg_file configs/zju_mocap_exp/latent_xyzc_multi.yaml exp_name xyzc_multi_debug resume False python run.py --type evaluate --cfg_file configs/zju_mocap_exp/latent_xyzc_multi.yaml exp_name xyzc_multi_debug eval True IBRNet Command python -m torch.distributed.launch --nproc_per_node=1 train.py --config configs/h36m.txt --num_source_views 3 --expname ..

Human3.6M에 part segmentation annotation이 있었다. 없는 줄 알고 예전에 densepose돌려서 얻었었는데... 퀄리티가 더 좋은지는 모르겠으나 괜찮아 보인다. 다만 README(로그인 안되어있으면 안 열릴수도)에도 part label에 대한 설명이 없어서 직접 분석함. 하는 김에 원본 데이터와 경식이형 parsing 데이터 관계도 정리함. 원본 데이터와 경식이형 parsing 데이터 명칭 관계 cam_name_mapper = {'01': '54138969', '02': '55011271', '03': '58860488', '04': '60457274'} action_name_mapper = { '01': {'act_02_subact_01': 'Directions 1', '..

https://andaeng.tistory.com/40 [Ubuntu 18.04] 우분투에서 무선랜카드 사용하기 (feat : iptime A3000UA) 바로 리눅스 환경에 iptimeA3000UA를 설치해보도록 하자 1. USB 연결 제품의 USB단자를 PC에 연결한다. 2. 드라이버 설치 Windows환경에서는 iptime 공식홈페이지에 이 제품에 제공되는 드라이버가 있어서 andaeng.tistory.com iptimeA3000UA-2를 사서 이분대로 하니까 바로 됐다. 2가 붙어서 다른 건 아닌가, 내가 잘못 산건 아닌가 걱정했는데 괜한 걱정. 이렇게 생겼다.

https://itzone.tistory.com/694 암호 없이 ssh-keygen & ssh-copy-id 사용하여 SSH 로그인 수행 3단계 원격 리눅스 서버에 암호 입력 없이 로그인 할 수 있습니다. 간단한 3단계 절차를 진행하면 되는데 ssky-keygen 사용과 이 글에서 설명하고 있는 ssh-copy-id 입니다. ssh-keygen 은 공용과 개인 키들을 생 itzone.tistory.com https://askubuntu.com/questions/311447/how-do-i-ssh-to-machine-a-via-b-in-one-command How do I SSH to machine A via B in one command? I want to access a computer, say ma..
- Total
- Today
- Yesterday
- 비전
- camera coordinate
- deep learning
- 컴퓨터비젼
- 인터뷰
- Interview
- Docker
- spin
- densepose
- Pose2Mesh
- 피트니스
- 에디톨로지
- nohup
- 컴퓨터비전
- VAE
- pyrender
- 머신러닝
- part segmentation
- Machine Learning
- 2d pose
- nerf
- Transformation
- demo
- 헬스
- world coordinate
- Virtual Camera
- pytorch
- Generative model
- focal length
- 문경식
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 |