Skip to content

Latest commit

 

History

History
65 lines (51 loc) · 3.02 KB

yolov5_wsl2_setup.md

File metadata and controls

65 lines (51 loc) · 3.02 KB
Raw

Setting up Yolov5 on WSL2 for GPU Usage

Updated 9/23/2022

Sytem Specs

GPU: NVIDIA 3080

NVDIA Driver: 516.94

CUDA: 11.2

PyTorch Libariries: torch 1.9.0+cu111, torchvision 0.10.0+cu111

Update NVIDIA Drivers

Update your NVIDIA drivers to the latest available version that will work with CUDA 11.2, so any driver with 11.X compatibility.

Install WSL2

Follow the steps listed here https://learn.microsoft.com/en-us/windows/ai/directml/gpu-cuda-in-wsl, then here https://docs.nvidia.com/cuda/wsl-user-guide/index.html

- Windows Version 21H2 or higher required for CUDA compatibility.
    - Run the `winver` application in Windows to see current version.
- Test that CUDA is working with `nvidia-smi` command. Note that CUDA Version lists compatiblity, not the actual installed version.

$ nvidia-smi
Fri Sep 23 11:11:54 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 515.65.01    Driver Version: 516.94       CUDA Version: 11.7     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  NVIDIA GeForce ...  On   | 00000000:06:00.0  On |                  N/A |
|  0%   48C    P8    35W / 320W |   2072MiB / 10240MiB |     26%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|  No running processes found                                                 |

Get Yolov5 Repository

Clone the yolov5 repository to your WSL instance and follow the instructions in the README to install dependencies. $ git clone https://github.com/ultralytics/yolov5.git

OPTIONAL: I used virtualenv to prepare a fresh pip environment for all the dependencies.

$ python3 -m venv /path/to/new/virtual/environment
$ source activate /path/to/new/virtual/environment/bin/activate

Find working PyTorch

My initial pytorch version installed through requirements.txt was not compatible with CUDA 11.2.

I found a working version from this github issue comment: pytorch/pytorch#50032 (comment)

Using that, running the below installed a working version

pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html

Run Training