mmdetection installation errors: possible fixes

[anushka-kore]

I was facing a ton of issues getting the mmdetection toolbox working even after following the installation guide, especially conflicts with certain packages (setuptools) and predependencies (mmcv). Any beginner (like me) installing the toolbox might encounter these common errors-

• AttributeError: module 'pkgutil' has no attribute 'ImpImporter'. Did you mean: 'zipimporter'?
• ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts. openxlab 0.1.2 requires setuptools~=60.2.0, but you have setuptools 76.0.0 which is incompatible.
• This error originates from a subprocess, and is likely not a problem with pip
• ModuleNotFoundError: No module named 'mmcv._ext'

Having gone through a lot of discussions and issues in the repo and some tweaking, I finally found a resolution to my issues. I'm noting the process I followed here, which would hopefully help anyone facing the same problems with installation.

The steps-

1. Make a new directory for your project using python 3.8 and navigate there, say, mkdir -p mmdet_project && cd mmdet_project. If you don't have python 3.8, I would recommend installing it first and making sure it is loaded.
2. Check your CUDA version. That can be done by nvcc --version or nvidia-smi if you have a NVIDA GPU. If CUDA is installed, you’ll see an output like:
   nvcc: NVIDIA (R) Cuda compiler release 12.4, V11.8.89
   This means you have CUDA 12.4.
3. Install torch based on the CUDA version. Check this site for the official installation link: https://pytorch.org/get-started/locally/. If you have CUDA 12.4, the prompt should be- pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124.
4. Proceed with the rest of the commands as given in the mmdetection installation guide-
   pip install -U openmim
mim install mmengine
mim install mmcv-full
   Note that the last prompt installs mmcv-full-1.7.2 which is going to give us an assertion error later because of mmdetection's mmcv version requirement.
5. Proceed with cloning the library-
   git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection
pip install -v -e .
   Note that this installation is still under your mmdet_project directory that you had initially created.
6. To verify whether mmdetection is installed correctly, the installation guide provides some sample codes to run an inference demo-
   mim download mmdet --config rtmdet_tiny_8xb32-300e_coco --dest .
python demo/image_demo.py demo/demo.jpg rtmdet_tiny_8xb32-300e_coco.py --weights rtmdet_tiny_8xb32-300e_coco_20220902_112414-78e30dcc.pth --device cpu
   But running this will give us the assertion error as previously anticipated. You might see something like-
   AssertionError: MMCV==1.7.2 is used but incompatible. Please install mmcv>=2.0.0rc4, <2.2.0. Now the installation guide tells you to do mim install "mmcv==<version>" but do NOT do this. Instead, install using pip- pip install "mmcv>=2.0.0rc4,<2.2.0" to overcome the assertion error and issues with mim installation. This will install mmcv-2.1.0.
7. Check your installation using the sample codes in the previous step again. They should run successfully and you will see a new image demo.jpg in your ./outputs/vis folder, where bounding boxes are plotted on cars, benches, etc.

Now the mmdetection toolbox is ready to be used. Hope this helps someone.