flownet2-pytorch issues
GitHub：https://github.com/NVIDIA/flownet2-pytorch
———————————————————————————————————————————————
pytorch>=0.4.1
1.bash ./install.sh报错：nvcc fatal : Unsupported gpu architecture ‘compute_70’
- cuda版本要求9.0
1> pytorch安装对应版本
conda install -n py36t4 pytorch=0.4.1 torchvision cudatoolkit=9.0 -c pytorch
2>查看nvcc版本
nvcc --version
3>修改当前cuda（服务器）vim ~/.bash.profile
# .bash_profile
2
3 # Get the aliases and functions
4 if [ -f ~/.bashrc ]; then
5 . ~/.bashrc
6 fi
7
8 # User specific environment and startup programs
9
10 PATH=/usr/local/cuda2/bin:$HOME/.local/bin:$HOME/bin:$PATH
11 export PATH
12
13 LD_LIBRARY_PATH=/usr/local/cuda2/lib64:$LD_LIBRARY_PATH
14 export LD_LIBRARY_PATH
15
16 INCLUDE_PATH=/usr/local/cuda2/include:$INCLUD_PATH
17 export INCLUDE_PATH
退出当前环境后执行source ~/.bash_profile
logout后重新启动
vim ~/.bashrc
# .bashrc
2
3 # Source global definitions
4 if [ -f /etc/bashrc ]; then
5 . /etc/bashrc
6 fi
7
8 # Uncomment the following line if you don't like systemctl's auto-paging feature:
9 # export SYSTEMD_PAGER=
10
11 # >>> conda initialize >>>
12 # !! Contents within this block are managed by 'conda init' !!
13 __conda_setup="$('/.../anaconda3/bin/conda' 'shell.bash' 'hook' 2> /dev/null)"
14 if [ $? -eq 0 ]; then
15 eval "$__conda_setup"
16 else
17 if [ -f "/.../anaconda3/etc/profile.d/conda.sh" ]; then
18 . "/.../anaconda3/etc/profile.d/conda.sh"
19 else
20 export PATH="/.../anaconda3/bin:$PATH"
21 fi
22 fi
23 unset __conda_setup
24 # <<< conda initialize <<<
25
4>修改三个文件夹下的setup.py
nvcc_args = [
'-gencode', 'arch=compute_35,code=sm_35',
#'-gencode', 'arch=compute_50,code=sm_50',
#'-gencode', 'arch=compute_52,code=sm_52',
#'-gencode', 'arch=compute_60,code=sm_60',
#'-gencode', 'arch=compute_61,code=sm_61',
#'-gencode', 'arch=compute_70,code=sm_70',
#'-gencode', 'arch=compute_70,code=compute_70'
]
2.报错
correlation_cuda_kernel.cu(130): error: identifier "__syncwarp" is undefined
correlation_cuda_kernel.cu(19): error: identifier "__shfl_down_sync" is undefined
- cuda链接库问题
3.Inference报错
segmentation fault(core dumped)
- 版本对应问题 未解决
pytorch<=0.4.0
1.Inference报错：
ImportError:.../.../lib/libstdc++.so.6:version `GLIBCXX_3.4.20' not found
- 解决：
conda install libgcc=5.2.0
继续报错：
ImportError:.../.../resempled2d_cuda-0.0.0-py3.6-linux-x86_64.egg/resampled2d_cuda.cpython-36m-x86_64-linux-gnu.so:undefined symbol: _ZN2at19UndefinedTensorImpl10_singletonE
- 原因：git未切换分支
- 解决：
git branch -a
git pull origin python36-PyTorch0.4
git commit -a
git checkout -b flow40 origin/python36-PyTorch0.4
git branch -a
2.bash ./install.sh报错：
error: ‘for’ loop initial declarations are only allowed in C99 mode
command 'gcc' failed with exit status 1
解决： export CFLAGS="-std=c99"
3.Inference测试时报错
命令为 python main.py --inference --model FlowNet2 --save_flow --inference_dataset ImagesFromFolder --inference_dataset_root /path/to/image/folder --resume ./FlowNet2_checkpoint.pth.tar
（需将原命令处的MpiSintelClean替换为ImagesFromFolder,同时注意datasets.py中对应类的参数itex值是否与图片格式相符,否则会报错IndexError:list index out of range）
报错：
error in Correlation_forward_cuda_kernel: invalid device functionncing
解决：
Turns out in the 3 make.sh, the architecture is hardcoded to -arch=sm_52 and in my case it needs to be -arch=sm_35.
//
//
command ‘gcc’ failed with exit status 1
export CXXFLAGS="-std=c++11"
export CFLAGS="-std=c99"