Deep Learning Software Setup: CUDA 10 + Ubuntu 18.04 (Part 2 of a $4000 RTX 2080Ti (MSI) DL box series)

Part 2 of getting our First Deep Learning Build

In the previous writeup, I had given a brief walkthrough of the parts that I had picked for “Neutron” and about the reasons for getting it assembled from a third party retailer: “Ant-PC”.

In this blog, I’ll share the step by step instructions that for setting up software on an Nvidia-based “Deep Learning Box”.

Overview:

For storage, I have 2 Drives:

• Samsung 970 Pro NVMe M.2 512GB
• 2TB HDD

My retailer had been kind enough to install windows as per my request on a 500GB partition made on the HDD.

Ubuntu 18.04 Installation

For Neutron, I was going to have Ubuntu 18.04 installed on the M.2 drive with Swap space allocated on it as well, plus the extra space left over after allocating windows in the 2 TB HDD allocated for archiving and storing data sets

Pre-Setup

• Disable Fast Startup
  (Using Search, Open Power Options) Power Options > System Settings > Choose what the power buttons do and uncheck the Turn on fast startup box.

Source

• Disable Secure Boot
  - Search Advanced Startup Options
  - Click on Restart Now
  - Select Troubleshoot>Advanced Options>UEFI Firmware Option Settings>Restart
  - Next BIOS Settings will open, this will vary depending on your setup
  - Navigate to Boot Tab
  - Disable Secure Boot
  - You may need to Turn Legacy Support On/Off

Source

Create Installer

• Download the Latest Image of Ubuntu 18.04 from Here.
• You will need a Memory Stick of ≥4GB.
• If you’re using Windows, use This Tool for a quick setup.
• If you have access to an Ubuntu System, use the inbuilt Startup Disk Creator.
• Restart and Enter BIOS and Enable Boot from USB.

Installing Ubuntu

• You should be able to boot into Ubuntu
• For sanity checks-select “Try Ubuntu without installing”
  - Check if your network cards are working okay.
  - Check the sound cards for quick relief.
• Note: If either of the cards isn’t being detected by Ubuntu, it might have to do with the kernel version-check your device names and see if those are supported. If not, you may need to get an external network/wifi card and an external sound card.
• Select the Language

Source

• Ignore the Update while Installing (It slows down the installation if your internet is slow)
• Select Language.

Source

• Select Something Else
• From here, select your drive. I picked the 512 GB M.2
• Click on “+”
  Next you need to allocate space to your OS and to SWAP.
• Give Most of the space to /
  (I prefer allocating one single space to / and /home. Reason: Some Installations, especially CUDA might cross the size allocated to /)

Source

• Give 2x The Space of your RAM to SWAP.
  Personally, I have 64GB RAM and have allocated 32GB just to give me extra room to work with Huge Datasets.

Source

• Setup your login credentials

Source

• Finally Click Install

Source

• Finally, Reboot.

That’s it, you’re done.

In the next sub-part, we’ll look at CUDA 10 Installation

CUDA 10 Installation

Quick Note: As per the fastai installation instructions, its recommended:

If you use NVIDIA driver 410+, you most likely want to install the cuda100 pytorch variant, via:

conda install -c pytorch pytorch cuda100

Below are the instructions for installing CUDA using the .run file provided by Nvidia. Reason: It may have some issues with Tensorflow since the cuda100 variant is just for PyTorch, I’ll update this post later after testing on that more.

Now, on to the installation:

Update and Upgrade

$ sudo apt-get update
$ sudo apt-get upgrade

Install Nvidia Drivers

• Download the drivers from here.
• Next, give run permissions to the .run file

chmod +x NVIDIA-Linux-x86_64-410.93.run

• Reboot and Check if Drivers are installed correctly by running:

nvidia-smi

• If you get an output with details of your GPU, everything is set up.

Pitfall:

• In case your Ubuntu gets stuck in a “boot loop”:
  - Press Ctrl+Alt+F3 (Note: on 16, its ctrl+alt+f1 to get tty1)
  - Purge the drivers

sudo apt-get purge nvidia-*

• Reboot

Install Conda

• To install the Latest version of Conda, we’ll use a handy Script by Kaggle Master Mikel

curl https://conda.ml | bash

Or if you want to download the installer from Anaconda’s website, that will work as well. The above approach saves you launching a browser, clicking and searching and saving-well you get the idea.

Install CUDA

• Download CUDA from here.
• I’m Setting up CUDA 10.0
• Download the runfile (local)

• Open Terminal and Change directory to Downloads.
• Change Executable Permissions for the Installer.
• Run it.

$ cd Downloads
$ chmod +x ./cuda_10.0.130_410.48_linux.run
$ sudo ./cuda_10.0.130_410.48_linux.run

• Press Space to Scroll Down and accept the terms and conditions
• Typeyes to the Prompt “Install with an unsupported configuration”
• Type no when prompted “Install NVIDIA Accelerated Graphics Driver for Linux-x86_64 xyz.xyz ”.
• Keep the defaults.
• Reboot
• After Reboot, check CUDA Version

nvcc -V

This should prompt the Installed CUDA Version

• In case of a Boot Loop problem, purge the drivers and try again.

Set Paths

export PATH="/home/User/anaconda3/bin:/usr/local/cuda/bin:$PATH"
export LD_LIBRARY_PATH="/usr/local/cuda/lib64:$LD_LIBRARY_PATH"
source ~/.bashrc

Install Cudnn

• Download the Cudnn version supported by your installed CUDA Version from Here (you will need an Nvidia Account for this)
• Once downloaded, we are going to unpack the archive and move it the contents into the directory where we installed CUDA 10.0:

tar -zxvf cudnn-_.0-linux-x64-v_.tgz
sudo cp -P cuda/lib64/* /usr/local/cuda-10.0/lib64/
sudo cp  cuda/include/* /usr/local/cuda-10.0/include
sudo chmod a+r /usr/local/cuda-10.0/include/cudnn.h

Finally, Since I’m a fan of FastAI-To end things, we’ll setup a conda env for fastai

Setup FastAI

• Create a conda env

$ conda create -n fastai 
$ conda activate fastai

• Sharing the instructions from the official repo.

$ conda install -c pytorch -c fastai fastai

• Follow some further instructions for performance gains

$ conda uninstall --force jpeg libtiff -y
$ conda install -c conda-forge libjpeg-turbo
$ CC="cc -mavx2" pip install --no-cache-dir -U --force-reinstall --no-binary :all: --compile pillow-simd

Setup (Other/Any) Environment

• Pip install or conda install using the requirements.txt or req.yml files.

pip install -r <Link>

Quick Test

• Checking if CUDA devices are accessible inside a Library.
• PyTorch has a function to Check CUDA device(s)

torch.cuda.is_available()

This should return a True.

Quick Test

To Warm up our GPU, we can try the sample script from the docs:

path = untar_data(URLs.MNIST_SAMPLE)
data = ImageDataBunch.from_folder(path)
learn = create_cnn(data, models.resnet18, metrics=accuracy)
learn.fit(1)

Just run these few lines of code in an nb.

On my GPU-this takes about 2 seconds to run.

That’s it. You’re done!

If you found this article to be useful and would like to stay in touch, you can find me on Twitter here..