Installing Tensorflow and CUDA 10 on Debian Buster the easy way

Installing Tensorflow and CUDA 10 on Debian Buster the easy way

There is plenty of conflicting advice on installing GPU accceleration (CUDA) on Debian machines.  If you haven't yet purchased a GPU, I'd recommend getting an AMD GPU and installing ROCm. It just works, and the driver is built into the linux kernel!

If however you have an old Nvidia GPU (I have a 1070Ti) there is a pretty easy approach is outlined here...

This link needs to be updated to use the latest tensorflow which depends on CUDA 10.

Step 1. Install Nvidia Drivers

This requires a specific version of the nvidia driver (418.56 at the time of writing). This is available in the stretch-backports repository. So first add a line for stretch-backports to /etc/apt/sources.list

Then clean out any other version of the driver (really you want to uninstall all possible nvidia software.
    aptitude purge nvidia-driver nvidia-dkms nvidia-alternative nvidia-legacy-check nvidia-opencl-common libcuda1 nvidia-smi nvidia-vdpau-driver nvidia-settings nvidia-vulkan-icd

 We need to install the older version of the nvidia-driver from backports.

    sudo apt -t stretch-backports install nvidia-driver

If this results in warnings, then you have some existing nvidia stuff installed. Use aptitude to search for it and purge any packages.

Step 2. Install CUDA 10.0

Download from the deb (local) installer for CUDA 10.0 from the Nvidia Site. You will also need to create a login to download cuDNN.
  • dpkg -i cuda-repo-ubuntu1804-10-0-local-10.0.130-410.48_1.0-1_amd64.deb
  • sudo apt-key adv --fetch-keys http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1704/x86_64/7fa2af80.pub
  • apt update
  • apt install cuda-libraries-10-0
  • apt install -t stretch-backports libcuda1
  • dpkg -i libcudnn7_7.6.0.64-1+cuda10.0_amd64.deb

Step 3. Install Tensorflow

  • sudo pip3 install --upgrade tensorflow-gpu
  • sudo pip3 install --upgrade keras
Test the tensorflow install using the following command:

    python3 -c "import tensorflow as tf; tf.enable_eager_execution(); print(tf.reduce_sum(tf.random_normal([1000, 1000])))" 

Comments

  1. Unknown30 May 2020 at 15:50

    stretch-backports on buster? is that right. I get the following error
    E: The value 'stretch-backports' is invalid for APT::Default-Release as such a release is not available in the sources

    any help would be appreciated.
    thank you

    ReplyDelete

Post a Comment

Popular posts from this blog

Wow, why is a Russian satellite signal stronger than all the GPS satellites in New Zealand?

Image
Something bright in the radio sky First noticed yesterday (24 Feb 2019) that a very bright object has appeared in the radio-sky at civilian GPS frequency (1.575 GHz).. This object is stationary, and it's sky-coordinates corresponds to the Russian LUCH 5A geostationary satellite. It hasn't moved in more than 24 hours. This object is so bright that all other GPS, Galileo and even the Japanese QZS satellites are drowned out. The image below is a snapshot all-sky image taken with the TART radio-telescope located in New Zealand.  Transient Array Radio Telescope all-sky image taken from near Dunedin New Zealand. Image Date: 2019-02-25 10:57am NZDT. The image is looking straight up, with the largest circle being the horizon. Due south from New Zealand is down. What is it the Luch 5A? The position corresponds to a Russian geostationary relay satellite, the LUCH 5A. My estimate is that the signal strength is 300 times stronger than a GPS satellite signal. This satellite
Read more

The three golden rules of SPICE in KiCAD

Image
There are three things that stop spice from working inside KiCAD. Here they are. 1. Use a .spiceinit file To allow ngspice to use PSPICE models, place a .spiceinit file in your home directory. With the contents: * user provided init file set ngbehavior=ps 2. You need a zero node Always place a node zero from the PSPICE library in the Schematic Editor. Normally you will connect this symbol to your ground. 3. Match Package Pins Always match the pins on the package to the SPICE model pins. This is done when you edit the spice model for a component. For transistors, SPICE expects pins 1,2,3 to be collector, base, emitter respectively. You should modify the pin order to match the pins in the pakage. For the BC807 the correct alternate numbering is 3,1,2 as pin 3 is the collector, pin 1 is the base, and pin 2 is the emitter. See Figure below For the 741 model, the pin order must match the package pins.
Read more