SOLVED: mmdet setup installation for macOS (and other platforms)

[tam-trance]

I'm sharing what worked for me on my Macbook M2 Max
I encountered a lot of the same problems that others had posted, was about to give up, and fortunately finally found something working for me. So here it is step by step 🙂

Goal: New user to openmmlab. Want to install mmdet / mmdetection.
DON'T: Follow instructions on mmdetection page. https://mmdetection.readthedocs.io/en/latest/get_started.html
DO: Follow the installation instructions per PACKAGE individually, in the following order: torch, mmengine, mmcv, then lastly mmdet

I'm not a big conda user, but for this I reluctantly accepted lol.
So:

Create the virtual env with python-3.8

conda create --name openmmlab python=3.8 -y
conda activate openmmlab

Install torch and compatible torchvision packages
conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 -c pytorch
Find your necessary torch version here, according to your platform
https://mmcv.readthedocs.io/en/latest/get_started/installation.html

Then find the appropriate torch package versions here
https://pytorch.org/get-started/previous-versions/

I decide to skip installing their package manager openmim for simplicity's sake
and thus go with pip

pip install mmengine
pip install mmcv==2.0.0rc4 -f https://download.openmmlab.com/mmcv/dist/cpu/torch1.13/index.html

Note: above command was retrieved from above picture

Now you're ready to install mmdet! I do the git version so that I can use pip

git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection
pip install -v -e .

Verify the installation

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

And it worked for me! Trying the tracking installation next. Good luck!

[tam-trance]

Note: Even with MPS on my Macbook M2 Max, looks like it will still only work with CPU

When I tried with mps device
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 mps
It required this
export PYTORCH_ENABLE_MPS_FALLBACK=1

But that led to a worse more complicated error
'anec.reshape' op failed: input tensor dimensions are not supported on ANEs.
and
RuntimeError: nms_impl: implementation for device mps:0 not found.

So macbook CPU it is 🤷‍♀️

[tam-trance]

For Tracking Installation (https://mmdetection.readthedocs.io/en/latest/get_started.html), at the final step
python demo/mot_demo.py demo/demo_mot.mp4 bytetrack_yolox_x_8xb4-amp-80e_crowdhuman-mot17halftrain_test-mot17halfval.py --checkpoint bytetrack_yolox_x_crowdhuman_mot17-private-half_20211218_205500-1985c9f0.pth --out mot.mp4
I get this error 🙃
AssertionError: MMCV==2.0.0rc4 is used but incompatible. Please install mmcv>=2.0.0, <2.4.0.

But I literally can't change the mmcv version, as based in the above image.

FML time to get CUDA 🥲

[tam-trance]

Another very interesting data point:
Before starting the steps for Tracking Installation, this command was working!

cd mmdetection
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 after attempting the steps for Tracking Installation, I get this error!
AssertionError: MMCV==2.0.0rc4 is used but incompatible. Please install mmcv>=2.0.0, <2.4.0.

So... don't follow the steps for the Tracking Installation 👀