Chen Yulin's Blog

Posted 2025-04-22Updated 2025-05-23Note4 minutes read (About 570 words)

Official repo:
https://github.com/liuhengyue/fcsgg
Our repo:
https://github.com/PSGBOT/KAF-Generation

My venv: fcsgg

Installation

Environment Preparation

git clone git@github.com:liuhengyue/fcsgg.git
cd fcsgg
git submodule init
git submodule update
conda create --name fcsgg
conda create -n fcsgg python=3.10
conda install nvidia/label/cuda-11.8.0::cuda-toolkit -c nvidia/label/cuda-11.8.0
conda install cudatoolkit
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

# for building detectron
conda install -c conda-forge gcc=11.2.0
conda install -c conda-forge gxx=11.2.0

conda env config vars set LD_LIBRARY_PATH="/home/cyl/miniconda3/envs/fcsgg/lib/"
conda env config vars set CPATH="/home/cyl/miniconda3/envs/fcsgg/include/"
conda env config vars set CUDA_HOME="/home/cyl/miniconda3/envs/fcsgg/"

conda deactivate
conda activate fcsgg

export CC=$CONDA_PREFIX/bin/gcc
export CXX=$CONDA_PREFIX/bin/g++

pip install -r requirements.txt
python -m pip install -e detectron2

Downloads

Datasets:

cd ~/Reconst
wget https://cs.stanford.edu/people/rak248/VG_100K_2/images.zip -P ./Data/vg/
wget https://cs.stanford.edu/people/rak248/VG_100K_2/images2.zip -P ./Data/vg/
unzip -j ./Data/vg/images.zip -d ./Data/vg/VG_100K
unzip -j ./Data/vg/images2.zip -d ./Data/vg/VG_100K

Download the scene graphs and extract them to datasets/vg/VG-SGG-with-attri.h5.

Issues

1

1	AttributeError: module 'PIL.Image' has no attribute 'LINEAR'. Did you mean: 'BILINEAR'?

LINEAR-> BILINEAR: commit

2

在尝试训练的过程中报错：

1
2
3

  File "/home/cyl/Reconst/fcsgg/fcsgg/data/detection_utils.py", line 432, in generate_score_map
    masked_fmap = torch.max(masked_fmap, gaussian_mask * k)
RuntimeError: The size of tensor a (55) must match the size of tensor b (56) at non-singleton dimension 1

modify detection_utils.py: commit

Training

首先更改训练的配置文件./config/quick_schedules/Quick-FCSGG-HRNet-W32.yaml, (原文件使用预训练的参数)

MODEL:
  META_ARCHITECTURE: "CenterNet"
  HRNET:
    WEIGHTS: "output/FasterR-CNN-HR32-3x.pth"

更改为train from scratch

MODEL:
  META_ARCHITECTURE: "CenterNet"
  HRNET:
    WEIGHTS: ""  # Empty string to train from scratch

再运行：

1	python tools/train_net.py --num-gpus 1 --config-file configs/quick_schedules/Quick-FCSGG-HRNet-W32.yaml

成功训练✌

...
[04/23 10:21:01] d2.utils.events INFO:  eta: 0:05:37  iter: 1159  total_loss: 1.042  loss_cls: 0.7593  loss_box_wh: 0.08827  loss_center_reg: 0.02485  loss_raf: 0.1754    time: 0.4042  last_time: 0.4519  data_time: 0.0043  last_data_time: 0.0044   lr: 0.001  max_mem: 4141M
[04/23 10:21:09] d2.utils.events INFO:  eta: 0:05:29  iter: 1179  total_loss: 1.028  loss_cls: 0.7208  loss_box_wh: 0.09246  loss_center_reg: 0.02669  loss_raf: 0.1625    time: 0.4042  last_time: 0.4035  data_time: 0.0041  last_data_time: 0.0044   lr: 0.001  max_mem: 4141M
[04/23 10:21:17] d2.utils.events INFO:  eta: 0:05:21  iter: 1199  total_loss: 1.01  loss_cls: 0.671  loss_box_wh: 0.1038  loss_center_reg: 0.02432  loss_raf: 0.1635    time: 0.4042  last_time: 0.3943  data_time: 0.0042  last_data_time: 0.0043   lr: 0.001  max_mem: 4141M
[04/23 10:21:25] d2.utils.events INFO:  eta: 0:05:13  iter: 1219  total_loss: 0.9737  loss_cls: 0.6887  loss_box_wh: 0.0929  loss_center_reg: 0.02574  loss_raf: 0.1749    time: 0.4041  last_time: 0.4101  data_time: 0.0041  last_data_time: 0.0042   lr: 0.001  max_mem: 4141M
...

Explanation

See [[FCSGG Repo Explanation]]

Posted 2025-03-31Updated 2025-05-23Note2 minutes read (About 236 words)

DINOv2 Repository Application

Repository:

Installation

official:

1	conda env create -f conda.yaml

不建议使用official的conda.yaml, 使用更改后的conda_cyl.yaml。

pip install torch==2.0.0 torchvision==0.15.1 torchaudio==2.0.1

conda install nvidia/label/cuda-11.7.0::cuda-toolkit -c nvidia/label/cuda-11.7.0
conda install cudatoolkit
conda install -c conda-forge gcc=11.2.0
conda install -c conda-forge gxx=11.2.0

conda env config vars set LD_LIBRARY_PATH="/home/cyl/miniconda3/envs/dinov2/lib/"
conda env config vars set CPATH="/home/cyl/miniconda3/envs/dinov2/include/"
conda env config vars set CUDA_HOME="/home/cyl/miniconda3/envs/dinov2/"

export CC=$CONDA_PREFIX/bin/gcc
export CXX=$CONDA_PREFIX/bin/g++

# check with `which g++`

conda env update -f conda_cyl.yaml

pip3 install -U xformers==0.0.18

conda env config vars set PYTHONPATH="/home/cyl/Reconst/dinov2/"

Demo 🐱

官方提供了 depth estimation 和 segmentation 的 notebook，可以找时间理解一下

Train

使用的数据集为Imagenet-mini

imagenet-mini
├── labels.txt
├── train
└── val

Note: 需要额外添加一个label.txt

使用脚本生产数据集的meta data:

Posted 2024-12-14Updated 2025-05-23Note3 minutes read (About 485 words)

Cuda+Torch

nvidia-smi返回的是driver所能支持的最新的cuda版本
系统安装的cuda版本可以随意，torch会优先使用虚拟环境中安装的cuda版本

Conda管理Cuda

安装指定版本cuda-toolkit

1	conda install nvidia/label/cuda-12.4.0::cuda-toolkit -c nvidia/label/cuda-12.4.0

安装最新版本

1	conda install cuda-toolkit

某些仓库需要指定cuda路径才能编译包

1
2
3

conda env config vars set LD_LIBRARY_PATH="/home/cyl/miniconda3/envs/gsam/lib/python3.10/site-packages/nvidia/cuda_runtime/lib/:$LD_LIBRARY_PATH"
conda env config vars set CPATH="/home/cyl/miniconda3/envs/gsam/lib/python3.10/site-packages/nvidia/cuda_runtime/include/:$CPATH"
conda env config vars set CUDA_HOME="/home/cyl/miniconda3/envs/gsam/"

Note: 注意改变了库路径之后nvim中的lsp会报错，建议之后改回去

1
2
3

conda env config vars set LD_LIBRARY_PATH=""
conda env config vars set CPATH=""
conda env config vars set CUDA_HOME=""

Note: To find the correct path for CUDA_HOME use which nvcc. In my case, output of the command was:

1 2	>>> which nvcc /home/user/miniconda3/envs/py12/bin/nvcc

Therefore, I set the CUDA_HOME as /home/user/miniconda3/envs/py12/.

Note: To find the correct path for LD_LIBRARY_PATH use find ~ -name cuda_runtime_api.h. In my case, output of the command was:

>>> find ~ -name cuda_runtime_api.h
...
/home/user/miniconda3/envs/py12/targets/x86_64-linux/include/cuda_runtime_api.h
...

So I set the LD_LIBRARY_PATH as /home/user/miniconda3/envs/py12/targets/x86_64-linux/lib/ and CPATH as /home/user/miniconda3/envs/py12/targets/x86_64-linux/include/. If you have multiple CUDA installations, the output of find ~ -name cuda_runtime_api.h will display multiple paths. Make sure to choose the path that corresponds to the environment you have created.

ref:https://github.com/IDEA-Research/GroundingDINO/issues/355

Note: Always reboot the computer after the cuda is upgraded

Note: 在更改LD_LIBRARY_PATH后可能会导致neovim的pyright无法运行，所以建议在编译完成后设回该变量

1	conda env config vars set LD_LIBRARY_PATH=""

`cudatoolkit`和`cuda-toolkit`

这两个可以同时安装
如果不安装cudatoolkit可能会在编译时出现ld: cannot find -lcudart: No such file or directory collect2: error: ld returned 1 exit status 报错

使用以下指令获取版本信息

1	python -c 'import torch;print(torch.__version__);print(torch.version.cuda)'

1 2	2.0.0+cu117 11.7

Installation

Environment Preparation

Downloads

Issues

1

2

Training

Explanation

Installation

Demo 🐱

Train

Conda管理Cuda

`cudatoolkit`和`cuda-toolkit`

Archives

Recents

Tags

Installation

Environment Preparation

Downloads

Issues

1

2

Training

Explanation

Installation

Demo 🐱

Train

Conda管理Cuda

cudatoolkit和cuda-toolkit

Archives

Recents

Tags

`cudatoolkit`和`cuda-toolkit`