Repo is not out-of-the box anymore, please specify versions

[gaborlegradi]

Hello, It is a really nice repo, however it is not out-of-the box anymore. I have been fighting version-hell for a while. I think that the problem is threefold:

• /requirements/*.txt do not contain specific versions.
• In /docs/getting_started.md versions mmdet==2.10.0 and mmcv==1.2.4 are given, however there is no hint for mmdet3d version.
• The best would be a full pip/conda freeze version snapshot when everything is installed according to INSTALLATION in getting_started.md.

[Rennylex]

That's a very good point! Did you get yours running? I'm stucked in the version-hell as well. What's the version working for you?

[gaborlegradi]

That's a very good point! Did you get yours running? I'm stucked in the version-hell as well. What's the version working for you?

Here are my notes, which I am not 100% sure that complete. I have abandoned playing with TransFusion, I will be back soon, then I awake again the repo from zero and I refresh here my notes.

Requirement: Conda or miniconda installed.

cd ~/install/ # Typically I use ~/install/ for storing install files.

wget http://developer.download.nvidia.com/compute/cuda/11.0.2/local_installers/cuda_11.0.2_450.51.05_linux.run

chmod 777 cuda_11.0.2_450.51.05_linux.run

sudo sh cuda_11.0.2_450.51.05_linux.run # accept, only toolkit to be installed

/usr/local/cuda/bin/nvcc --version # Build cuda_11.0_bu.TC445_37.28540450_0 # nvcc --version

#conda remove -n trfu --all -y #If env was created and deleted several times.

conda create -n trfu python=3.7 -y

conda activate trfu

conda install pytorch==1.7.0 torchvision==0.8.0 torchaudio==0.7.0 cudatoolkit=11.0 -c pytorch -y

#python
#import torch
#torch.cuda.is_available()
#exit()

pip install -U openmim # mim is the package management tool for the OpenMMLab projects, which makes it easy to install mmcv # https://mmcv.readthedocs.io/en/latest/get_started/installation.html

mim install "mmcv<=1.4.0"

mim install "mmdet<=2.10.0"

mim install mmcv-full==1.2.4

git clone https://github.com/open-mmlab/mmdetection3d.git

cd mmdetection3d

git checkout 9802d41

pip install -v -e . # or "python setup.py develop"

pip uninstall pycocotools -y

pip install mmpycocotools

pip uninstall mmpycocotools

pip install mmpycocotools

Then you can run demo/pcd_demo.py

[Rennylex]

Thanks a lot!!!! Have you got the Docker running btw?

[gaborlegradi]

Thanks a lot!!!! Have you got the Docker running btw?

I do not remember... I will fix that too sooner or later...

[Rennylex]

Thanks! Lemme know if you have any update. I'm gonna send you my notes too once I get mine working.

[changxu-zhang]

Hi,

I'm also trapped into version problem right now, tried all combinations between recommended mmcv (1.2.4 to 1.3.0) and mmdet (2.5.0 to 3.0.0) and it didn't work

[DanielDoerr]

Hey @changxu-zhang

I tried @gaborlegradi instruction but ran into cuda errors.
So I tried it with cuda 11.1 and Pytorch 1.8.0

wget https://developer.download.nvidia.com/compute/cuda/11.1.0/local_installers/cuda_11.1.0_455.23.05_linux.run

sudo sh cuda_11.1.0_455.23.05_linux.run --silent --toolkit --override

pip install torch==1.8.0+cu111 torchvision==0.9.0+cu111 torchaudio==0.8.0 -f https://download.pytorch.org/whl/torch_stable.html

pip install -U openmim

... and the further commands of the instruction

After that I had to copy/replace the following files to train with the nuscenes dataset.
From the folder TransFusion/mmdet3d into the folder mmdetection3d/mmdet3d:
./models/detectors
- transfusion.py
- mvx_two_stage.py
- init.py

./models/dense_heads
- transfusion_head.py
- init.py

./core/bbox/coders
- transfusion_bbox_coder.py
- init.py

./core/bbox/assigners.py
- hungarian_assigner.py
- init.py

./datasets
- nuscenes_dataset.py

./datasets/pipelines
- loading.py

I'm pretty new into this. Maybe there is a better workaround.

[changxu-zhang]

Hi @DanielDoerr ,

Could you please post your mmcv, mmdet and mmdet3d version?

Thanks

[DanielDoerr]

@changxu-zhang

mmcv 1.4.0
mmcv-full 1.2.4
mmdet 2.10.0
mmdet3d 0.11.0

What errors do you get during the installation?

I also have many installation attempts before that. The only way that worked for me was the instructions from above.
I installed it on Ubuntu 20.04 and also on WSL2.

Please make sure you install the correct CUDA version for your distribution.

[tlhopl]

hello @DanielDoerr @gaborlegradi
I have a problem while installing mmdet 2.10.0 to try @gaborlegradi instroductions..

mim install "mmdet<=2.10.0" or pip install mmpycocotools:

ERROR: Failed building wheel for mmpycocotools
I can't google abour this problem because the latest version of mmdet is no longer using mmpycocotools

I'm using CUDA 11.1 and pytorch 1.8.0, gcc 9.4.0 and ubuntu 20.04

how can I solve this problem to run Transfusion?