-
Notifications
You must be signed in to change notification settings - Fork 14
Linux & Docker Setup Instructions
arobie edited this page Jan 16, 2020
·
13 revisions
This page describes setting up APT to train and track on local GPU(s) using Docker on Linux.
- Our current pre-built APT Docker container uses CUDA 10.1. Nvidia GPUs with Compute Capability >= 3.0 are supported. CUDA GPUs_supported
- CUDA 10.1 requires CUDA Driver versions>=418.39 (https://docs.nvidia.com/deploy/cuda-compatibility/index.html#binary-compatibility__table-toolkit-driver).
- The CUDA Driver will need to be present/installed on your machine. If it is not yet installed, you can download a driver at either i) https://developer.nvidia.com/cuda-downloads or ii) https://www.nvidia.com/Download/index.aspx. The first link downloads the entire "CUDA Toolkit", which as a convenience includes the CUDA driver. Make sure that a recent-enough version of the driver (per above) is installed.
-
APT requires Docker version 19.03 or later. Install Docker by following the instructions at https://docs.docker.com/install/linux/docker-ce/ubuntu.
- If you have an older version of Docker (pre-19.03), make sure to uninstall the old version first as described in the instructions.
- Run the linux post-installation steps at https://docs.docker.com/install/linux/linux-postinstall, under the section "Manage Docker as a non-root user," so that you can run docker without sudo.
- Install nvidia-docker at https://github.com/NVIDIA/nvidia-docker.
- Pull the latest APT docker image by running
docker pull bransonlabapt/apt_docker:latest
in a linux terminal. This will download the pre-built APT docker container.- In some cases, you may need to pull an older docker image eg for compatibility with your hardware.
Here are some test commands you can try to check that everything is working as expected:
- In a linux terminal:
docker run hello-world
. This tests your Docker installation. - In a linux terminal:
docker run --gpus all nvidia/cuda:9.0-base nvidia-smi
. This tests your Nvidia-Docker installation.
- Start APT as usual with
StartAPT
orlObj = StartAPT;
(not inside docker). - Under Track> GPU/Backend Configuration, select Docker.
- Defaults to local host, if using remote host you can set using Track> GPU/Backend Configuration,(Docker) set remote host ...
- Test your Docker installation by selecting Track> GPU/Backend Configuration> Test backend configuration. This will perform some tests to ensure that your Docker installation will work with APT.
- If all tests pass, then you are good to go with training and tracking in APT with your local GPU(s)!