WSL2.md

Getting a linux virtual environment with GPU access on Windows 10

WSL2 with Ubuntu 18.04 and NVidia containers (tf 1.15)

• Follow these directions, everything before "Set up Python environment"
  • Choose Ubuntu 18.04
• Follow some sections in these directions
  • Do "Setting up CUDA Toolkit"
  • In "Running CUDA Applications", try to cd /usr/local/cuda/samples/0_Simple/matrixMulCUBLAS/, make, and ./matrixMulCUBLAS
  • Do "Setting up to Run Containers"
• Follow 1-4 in these directions
• Inside wsl do docker pull nvcr.io/nvidia/tensorflow:20.11-tf1-py3
  • run this container with docker run --gpus all --rm -it -v /mnt/d/Documents/:/projects nvcr.io/nvidia/tensorflow:20.11-tf1-py3 substituting /mnt/d/Documents/ for your documents folder
  • other containers are available here

Easy benchmark test of tf 1.15

git clone https://github.com/aime-team/tf1-benchmarks.git
cd tf1-benchmarks
python ./scripts/tf_cnn_benchmarks/tf_cnn_benchmarks.py --model resnet50 --num_gpus=1 --batch_size=1

Increase batch size slowly, out of memory error in this benchmark needs to be kill -9ed from another terminal (docker container list, docker exec -it CONTAINER_ID bash, ps -a, and kill -9 PID).

On GeForce GTX 960 4GB, max batchsize is 3 and gets 20.48 images/sec. This is 10x faster than benchmark running with directml tensorflow.