mmsegmentation 项目排坑

任务

之前打算在segformer上做工作。但发现现在segformer已经被新版的mmsegmentation集成了。上次算是白忙活了，这次把新版的mmsegmentation跑通。

过程记录

创建conda环境

按照官网readme和此前经验进行：

conda create -n OPENMMLAB python=3.8 -y
conda activate OPENMMLAB
pip3 install empy rospkg pyyaml catkin_pkg
conda install pytorch torchvision -c pytorch
pip3 install -U openmim
mim install mmcv-full

中间提醒没有matplotlib，以及platformdirs版本不对，所以穿插进行了：

pip3 install matplotlib
pip install platformdirs==3.9.0

安装MMSegmentation

因为后面需要开发自己的模型，所以我选择源码安装：

git clone https://github.com/open-mmlab/mmsegmentation.git
cd mmsegmentation
pip3 install -v -e .

验证安装

下载配置文件和模型权重文件：

mim download mmsegmentation --config pspnet_r50-d8_512x1024_40k_cityscapes --dest .

报错一大堆：

不会用Model zoo，在上面找了半天没找到这个文件。搜了下，发现有个老哥跟我类似问题，但他就是执行我这条命令解决的：参考issue。所以没法直接用，但他说的要看config里有没有，提醒了我。我查了下自己的config：

find configs/pspnet/ -type f -name 'pspnet_r50-d8*cityscapes*'

发现：

configs/pspnet/pspnet_r50-d8_4xb2-40k_cityscapes-512x1024_dark-zurich-1920x1080.py
configs/pspnet/pspnet_r50-d8_4xb2-40k_cityscapes-512x1024_night-driving-1920x1080.py
configs/pspnet/pspnet_r50-d8_4xb2-40k_cityscapes-769x769.py
configs/pspnet/pspnet_r50-d8_4xb2-80k_cityscapes-512x1024.py
configs/pspnet/pspnet_r50-d8_4xb2-40k_cityscapes-512x1024.py
configs/pspnet/pspnet_r50-d8-rsb_4xb2-adamw-80k_cityscapes-512x1024.py
configs/pspnet/pspnet_r50-d8_4xb2-80k_cityscapes-512x1024_dark-zurich-1920x1080.py
configs/pspnet/pspnet_r50-d8_4xb2-80k_cityscapes-512x1024_night-driving-1920x1080.py
configs/pspnet/pspnet_r50-d8_4xb2-80k_cityscapes-769x769.py

也就是我pspnet_r50-d8_512x1024_40k_cityscapes下载不到，还可以下载别的这里有的。这时我看到了这个博客参考博客，发现下载的要和文件名匹配。
哦，那我把上述下载命令改为：

mim download mmsegmentation --config pspnet_r50-d8_4xb2-40k_cityscapes-512x1024 --dest .

果然成功了！
然后按照readme进行：

python demo/image_demo.py demo/demo.png pspnet_r50-d8_512x1024_40k_cityscapes.py pspnet_r50-d8_512x1024_40k_cityscapes_20200605_003338-2966598c.pth --device cpu --out-file result.jpg

报错：

ModuleNotFoundError: No module named 'mmengine'

执行：

pip3 install mmengine

再运行，发现糟糕了：

raceback (most recent call last):
  File "demo/image_demo.py", line 6, in <module>
    from mmseg.apis import inference_model, init_model, show_result_pyplot
  File "/home/lcy-magic/Segment_TEST/mmsegmentation/mmseg/__init__.py", line 61, in <module>
    assert (mmcv_min_version <= mmcv_version < mmcv_max_version), \
AssertionError: MMCV==1.7.2 is used but incompatible. Please install mmcv>=2.0.0rc4.

竟然版本不对！
看到当时他是这么提示我操作的：

于是我点开MMCV那个链接，看到这样的提示：

那就试试你吧：

mim install mmcv==2.0.0

结果报大错：

error: command '/usr/bin/gcc' failed with exit code 1
      [end of output]
  
  note: This error originates from a subprocess, and is likely not a problem with pip.
  ERROR: Failed building wheel for mmcv
  Running setup.py clean for mmcv
Failed to build mmcv
ERROR: Could not build wheels for mmcv, which is required to install pyproject.toml-based projects

我猜可能是cuda问题，检查下：

发现没有cuda。那就装。
查看torch版本：

pip3 show torch

发现是2.2.1，然后看pytorch官网，发现是11.8的cuda，于是：

conda install cudatoolkit=11.8 -y

但还是安装不了mmcv2.0.0。于是我放弃min安装。（上次就放弃了mim安装，用pip安装的）
参考官方指示：官方文档：

pip install mmcv==2.1.0 -f https://download.openmmlab.com/mmcv/dist/cu118/torch2.1/index.html

成功安装。
此时运行测试：

python demo/image_demo.py demo/demo.png pspnet_r50-d8_512x1024_40k_cityscapes.py pspnet_r50-d8_512x1024_40k_cityscapes_20200605_003338-2966598c.pth --device cpu --out-file result.jpg

报错：

ModuleNotFoundError: No module named 'ftfy'
ModuleNotFoundError: No module named 'regex'

安装：

pip3 install ftfy
pip3 install regex

再运行还是报错：

ImportError: libc10_cuda.so: cannot open shared object file: No such file or directory

参考博客这篇博客说是cuda的问题。检查了下，还真是：

于是参照pytorch官网重新安装：

pip3 uninstall torch torchvision torchaudio
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

好长时间，终于装完了。再运行，还是报错了，呜呜：

ImportError: /home/lcy-magic/anaconda3/envs/OPENMMLAB/lib/python3.8/site-packages/mmcv/_ext.cpython-38-x86_64-linux-gnu.so: undefined symbol: _ZN2at4_ops10zeros_like4callERKNS_6TensorEN3c108optionalINS5_10ScalarTypeEEENS6_INS5_6LayoutEEENS6_INS5_6DeviceEEENS6_IbEENS6_INS5_12MemoryFormatEEE

再去看官网，发现那个mvcc对应的是torch2.1.X，还没有适配2.2.X。所以我打算再重根据pytorch官网装一次torch：

conda install pytorch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 pytorch-cuda=11.8 -c pytorch -c nvidia

结果还是不行，又找不到cuda了。我崩溃了。决定全部推倒重来，还用python3.8,torch2.1.0,cuda11.8,mmcv2.1.0。不行就再重来，用上个项目的torch和cuda。

忘了，中间conda install时候我遇到了网络问题。估计是我的代理问题，我把他关掉就好了。

重来

创建conda环境

conda remove -n OPENMMLAB --all -y
conda create -n OPENMMLAB python=3.8 -y
conda activate OPENMMLAB
pip3 install empy rospkg pyyaml catkin_pkg
conda install pytorch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 pytorch-cuda=11.8 -c pytorch -c nvidia
pip install -U openmim
pip3 install matplotlib
pip install platformdirs==3.9.0

检查：

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

结果：

2.1.0
11.8

没问题。然后装mmcv：

pip install mmcv==2.1.0 -f https://download.openmmlab.com/mmcv/dist/cu118/torch2.1/index.html

装其他：

pip install -v -e .
pip install ftfy regex

验证安装

执行：

python demo/image_demo.py demo/demo.png pspnet_r50-d8_512x1024_40k_cityscapes.py pspnet_r50-d8_512x1024_40k_cityscapes_20200605_003338-2966598c.pth --device cpu --out-file result.jpg

报错，但总归不是环境报错了：

Traceback (most recent call last):
  File "demo/image_demo.py", line 51, in <module>
    main()
  File "demo/image_demo.py", line 32, in main
    model = init_model(args.config, args.checkpoint, device=args.device)
  File "/home/lcy-magic/Segment_TEST/mmsegmentation/mmseg/apis/inference.py", line 41, in init_model
    config = Config.fromfile(config)
  File "/home/lcy-magic/anaconda3/envs/OPENMMLAB/lib/python3.8/site-packages/mmengine/config/config.py", line 458, in fromfile
    lazy_import is None and not Config._is_lazy_import(filename):
  File "/home/lcy-magic/anaconda3/envs/OPENMMLAB/lib/python3.8/site-packages/mmengine/config/config.py", line 1657, in _is_lazy_import
    with open(filename, encoding='utf-8') as f:
FileNotFoundError: [Errno 2] No such file or directory: 'pspnet_r50-d8_512x1024_40k_cityscapes.py'

改成之前下载的脚本和参数文件名：

python demo/image_demo.py demo/demo.png pspnet_r50-d8_4xb2-40k_cityscapes-512x1024.py pspnet_r50-d8_512x1024_40k_cityscapes_20200605_003338-2966598c.pth --device cpu --out-file result.jpg

警告：

/home/lcy-magic/Segment_TEST/mmsegmentation/mmseg/models/losses/cross_entropy_loss.py:250: UserWarning: Default ``avg_non_ignore`` is False, if you would like to ignore the certain label and average loss over non-ignore labels, which is the same with PyTorch official cross_entropy, set ``avg_non_ignore=True``.
  warnings.warn(
Loads checkpoint by local backend from path: pspnet_r50-d8_512x1024_40k_cityscapes_20200605_003338-2966598c.pth
/home/lcy-magic/anaconda3/envs/OPENMMLAB/lib/python3.8/site-packages/mmengine/visualization/visualizer.py:196: UserWarning: Failed to add <class 'mmengine.visualization.vis_backend.LocalVisBackend'>, please provide the `save_dir` argument.
  warnings.warn(f'Failed to add {vis_backend.__class__}, '

但是已经有结果了：

大功告成！