I've struggled quite a bit trying to install the right CUDA version and some of the dependencies that are required by Tensorflow to work with GPU. As a result, I decided to make a recipe for my future self and others. Most of the steps that I've mentioned below are from https://deeplabcut.github.io/DeepLabCut/docs/recipes/installTips.html This is an attempt to make it more general for varied Ubuntu versions and GPUs.

Installing CUDA for the GPU

  • sudo apt install gcc
  • Go to the Link: https://developer.nvidia.com/cuda-downloads
    • Select Ubuntu
    • You can check your architecture type with command arch on your terminal
      • Select the architecture type(mine is x86_64)
    • Select your Linux Destro Type
      • Followed by Version
    • Lastly, Installer Type, Since I'm installing it locally so the install type is deb(local) Doing this will provide you with a list of commands, For example: After following these commands you'll observe 2 cuda folders in /usr/local/ Optional commands:
sudo add-apt-repository ppa:graphics-drivers/ppa
sudo apt update
sudo ubuntu-drivers autoinstall
  • And reboot
  • Verify the installation by
gcc --version
nvcc --version
nvidia-smi
  • If any of these commands return an error, you can top it up with sudo apt install nvidia-cuda-toolkit gcc-9

⚠️ Caveat: apt-installing nvidia-cuda-toolkit installs another cuda version on the device, as a result you might see 3 cuda folder in /usr/local/

Installing Anaconda

Installing DeepLabCut

  • sudo apt install libcanberra-gtk-module libcanberra-gtk3-module
  • git clone https://github.com/DeepLabCut/DeepLabCut.git
  • Open DeepLabCut/conda-environments/DEEPLABCUT.yaml
  • Change the following:
	-pip:
		- "deeplabcut[gui]"

         to:

	-pip:
		- deeplabcut
  • Run conda env create -f DEEPLABCUT.yaml
  • conda activate DEEPLABCUT
  • After you are done building the environment and installing deeplabcut within the environment, the next obstacle that we might face is GUI not working. DeepLabCut uses wxPython for its GUI, but wxPython doesn't have a general wheel for all Linux destros, so we need to install a specific wheel, which you can look up on https://extras.wxpython.org/wxPython4/extras/linux/gtk3, since I'm using Ubuntu 20.04, I'll run pip install -f https://extras.wxpython.org/wxPython4/extras/linux/gtk3/ubuntu-20.04 wxPython
  • conda install -c conda-forge wxpython
  • To verify the installation run python -m deeplabcut which should launch the DLC GUI.

Last Obstacle

  • Although you have installed CUDA but you might not be able to use GPU. Let's check it once before stating anything concretely.
  • In the ipythonconsole:
import tensorflow as tf
tf.test.is_gpu_available()

If this returns True your setup is COMPLETE!
Elseyou just need to wait a lil longer and we'll surely address this

  • You might be getting an error like this:
     ImportError: libcudart.so.8.0: cannot open shared object file: No such file or directory
  • You need to install cudnn files from https://developer.nvidia.com/rdp/cudnn-download
    • You might have to create an account and participate in a survey if you haven't already done that.
    • Post downloading it, you can install the software just by double clicking on it. -software install will pop up and it will look something like this:
    • To verify the aforementioned installation look for libcudnn.so.8.3.1 or similar named file in /usr/lib/x86_64-linux-gnu.
  • Now cd into all the cuda*/lib64 folder in /usr/local/
    • For example: /usr/local/cuda-11/lib64
  • And enter the following command
sudo ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.8.3.1 libcudnn.so.8
sudo ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.8.3.1 libcudnn.so
  • You are good to go!

Thank you reading! I hope it was useful.