Skip to content

Latest commit

 

History

History
50 lines (36 loc) · 4 KB

install_windows_nvidia.md

File metadata and controls

50 lines (36 loc) · 4 KB
Raw

MARS installation instructions

Environment: Windows

GPU: NVIDIA

Setting up your GPU for TensorFlow

MARS currently runs on TensorFlow version 1.15. You won't need to install TensorFlow yourself, as installation is handled by the conda environment below-- but you will need to install CUDA version 10.0 or later and cuDNN version 7.4 or later to allow TensorFlow to interact with your GPU. You will also want to make sure your NVIDIA graphic drivers are up to date.

First, install CUDA by following the NVIDIA CUDA installation guide, the core of which is to:

  1. Go to https://developer.nvidia.com/cuda-downloads, and enter your platform information.
  2. Download and run the provided *.exe file.
  3. Open a terminal by searching "cmd" in your start menu, and enter nvcc --version to confirm that CUDA has installed correctly.

Next, install cuDNN by following the NVIDIA cuDNN installation guide, the core of which is to:

  1. Go to https://developer.nvidia.com/cudnn, select "Download cuDNN", and sign into/sign up for a Developer account.
  2. In the "Downloads" menu on the developer site, select cuDNN version 7.4 or later. Make sure you select a cuDNN version that is compatible with your CUDA version! To check your CUDA version, launch a terminal and type nvcc --version.
  3. Follow instructions here to add cuDNN to your CUDA directory.

Setting up the MARS conda environment

  1. Install two prerequisites:
  • Install miniconda by downloading the Python 3.x version appropriate to your system from this page.
  • Install git from this page.

  1. Check your install by opening a new Anaconda Prompt (launched by searching "Anaconda" in the start menu) and typing where python. This should return one or more paths to python executables- make sure this includes something with miniconda3, like C:\Users\me\miniconda3\python.exe.

  2. Clone or download the contents of this GitHub repository. For this guide we'll assume you downloaded everything to C:\users\me\MARS.

    cd C:\users\me
git clone --recurse-submodules https://github.com/neuroethology/MARS

    (or read about other ways to clone/download repos here).

Note: the --recurse-submodules also ensures that the contents of the neuroethology/Util repository are also cloned when you clone MARS. Util contains code that is common to multiple projects on Neuroethology. If you didn't include this tag in the initial clone, you can clone the Util repo later by calling git submodule update --init --recursive from within the MARS directory.

  1. Build the MARS conda environment by entering the following in an Anaconda Prompt (launched by searching "Anaconda" in the start menu): 
    cd C:\users\me\MARS
conda env create -f MARS_environment_windows.yml

Once the build has finished, you can activate the MARS environment by calling conda activate mars (or system activate mars) from the Anaconda Prompt.

Downloading the trained MARS models

Before you can run MARS, you need to download the trained neural networks and classifiers MARS uses for pose estimation and behavior classification.

Models can be downloaded from https://data.caltech.edu/records/1655. After downloading, unzip the models folder into the MARS/mars_v1_8 directory. The contents of MARS/mars_v1_8/models should now be three directories called classifier, detection, and pose.

Now you're ready to detect some behaviors!