Skip to content

Latest commit

 

History

History
57 lines (42 loc) · 2.91 KB

README_CUDA.md

File metadata and controls

57 lines (42 loc) · 2.91 KB

CUDA Installation for use with this project.

  • Be sure to have the latest Nvidia Grapic Driver and we need to determine your hardware CUDA Version with the following command, anotate the exact version for next steps:

    • On Windows, :

      c:\Program Files\NVIDIA Corporation\NVSMI\nvidia-smi.exe

    • On Linux, run:

      nvidia-smi

  • After finding the correct CUDA Version for your device in the output of the previous command, please download and install the Cuda Toolkit for your EXACT CUDA Version from: CUDA Toolkit Archive

  • Go to TensorFlow, CUDA and cuDNN Compatibility and search for the following for your current CUDA Version and anotate the versions for the next steps:

    • The Tensorflow Version
    • The Python Version
    • The CUDNN Version

  • Search, download and install the correct CuDNN Version from the CuDNN Archive by using the CuDNN installation instructions

  • Restart your CLI, console or terminal, so the enviroment variables set by CuDNN installation are loaded

  • After restarting your console to load the environment variables, activate your conda environment again: bash conda activate feature-extractor-env

  • (Optionally) Update to the required Python Version for your CUDA Version in your conda environment:

    conda install python=<REQUIRED_PYTHON_VERSION>
python --version

  • Modify the requirements.txt file to show tensorflow-gpu==<REQUIRED_TENSORFLOW_VERSION_HERE> instead of just tensorflow, and if using tensorflow-gpu version more than 2.0, remove the keras line, since tensorflow-gpu > 2.0, already includes keras-gpu. Save the changes.

  • Install the modified requirements.txt, this time with tensorflow-gpu (Keras-gpu included) instead of just keras (you may need to fix some package versions in the readme for the requirements of your current tensorflow-gpu version, if some error appears):

    pip uninstall -y numpy scipy pandas tensorflow keras
pip install -r requirements.txt --no-cache-dir

  • Since tensorflow-gpu version 2.0, the keras-gpu package comes included and do not need separate installation, for previous versions, install the keras package with: pip install keras

  • To test if Keras is using the GPU:

    python
from keras import backend as K
K.tensorflow_backend._get_available_gpus()
exit()

  • If the previous test is passed, the GPU can be used, and no other changes in this repo code are required since it detects if a gpu is available automatically for training and evalation of trained models.