How to build a Docker container with tensorflow, Spyder IDE, jupyter etc. With GPU support !
Based on docker image tensorflow/tensorflow:latest-gpu-jupyter.
Strongly inspired by caliari-italo and dalthviz on issue 17542. And spyder workflow for linux tests. Thank you guys !
This is for Linux only. Tested on Ubuntu 22.04.
Follow tensorflow docker requirements.
Since we want to use GPU, also follow the NVIDIA container toolkit installation guide.
To do so, you will have to follow the CUDA installation guide for Linux.
To test the requirements, you should be able to run :
sudo docker run --runtime=nvidia --gpus all -it --rm tensorflow/tensorflow:latest-gpu-jupyter nvidia-smi
which should return something like :
Sun Sep 1 08:44:19 2024
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 560.35.03 Driver Version: 560.35.03 CUDA Version: 12.6 |
|-----------------------------------------+------------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA GeForce RTX 3060 Off | 00000000:07:00.0 On | N/A |
| 0% 50C P8 19W / 170W | 512MiB / 12288MiB | 2% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+
+-----------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=========================================================================================|
| No running processes found |
+-----------------------------------------------------------------------------------------+
Where your GPU card is visible (NVIDIA GeForce RTX 3060 in my case, please feel free to send me a newer one 😉).
Remark : this may work currently but may fail after some hours / reboots. See troubleshooting section.
You should also be able to run :
sudo docker run --runtime=nvidia --gpus all -it --rm tensorflow/tensorflow:latest-gpu-jupyter python -c "from tensorflow.python.client import device_lib ; import tensorflow as tf ; print('devices found:\n',tf.config.list_physical_devices('GPU'))"
which should output something like :
devices found:
[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]
(with a lot of warnings in my case, such as Unable to register cuDNN factor etc.)
If so, you should have the requirements properly working.
NOTE : yes we need to sudo the docker commands. I have not configure rootless mode so far.
sudo docker build -t tensorflow_spyder .
First activate X forwarding between your local and the container
xhost +local:docker
Then start a container with all these arguments
sudo docker run --runtime=nvidia --gpus all -it -e DISPLAY=$DISPLAY --net=host -v /tmp/.X11-unix:/tmp/.X11-unix -v $PWD:/home --rm tensorflow_spyder spyder
To start your container and Spyder IDE.
You can also start a bash (instead of spyder) and interact as you want within your container. But personnaly, I wasn't able to run spyder & (it stucks at "Loading Code Analysis...")
Since this container is based on docker image tensorflow/tensorflow:latest-gpu-jupyter, you can also refere to tensorflow documentation concerning docker.
For instance, to start a container with jupyter running :
sudo docker run --runtime=nvidia --gpus all -it -e DISPLAY=$DISPLAY --net=host -v /tmp/.X11-unix:/tmp/.X11-unix -v $PWD:/home -p 8888:8888 --rm tensorflow_spyder
Eventhough you maybe able to currently run nvidia-smi in you container, you may get error "Failed to initialize NVML: Unknown Error" after some hours / reboots.
I found the solution on stackoverflow : Failed to initialize NVML: Unknown Error in Docker after Few hours. I sudo edited /etc/nvidia-container-runtime/config.toml, to change no-cgroups = false