Yolov5 advanced 5 GPU environment setup

I'm familiar yolov5 Of CPU Environmental Science , The next step is to build GPU The environment . from yolov2 tensorflow keras darknet, The original GPU Building is very cumbersome , It's not just about cuda and cudnn Corresponding , It should be completely consistent with the driver version , More in windows Environment sets a large number of environment variables and paths , This is a very imperfect solution . Now? yolov5 pytorch This problem no longer exists , It can be said that it can be completed with one click in the activation environment GPU Construction of environment .
First of all, check Your own graphics card supports cuda edition , Because the test I do with my notebook now , The graphics card tested is MX230, Don't laugh , The training speed has been significantly improved .
Right click on the desktop nvidia panel see cuda

Now my reality is 11.7, In fact, before I upgraded the driver, I was 10.0. If you want to support the new Cuda edition , Please go nvidia Download the new driver on the official website , Don't worry that this driver is backward compatible , You just need to install Support below cuda11.7 Of pytorch, In fact, when I posted this article ,pytorch Only to 11.3.
After upgrading the driver , Follow the previous steps to create an isolated environment , in the future cuda cudnn And so on , There is no need to modify the environment variables , I am very tired of the process of changing environment variables .

If there are still problems in building the environment and creating separate folders , You can read the first article yolov5 Environment quickly created articles
https://blog.csdn.net/buaaweibin/article/details/125356544?spm=1001.2014.3001.5502

Open below pytorch Website

You can choose 11.0 edition , But the front cpu It was said during the test 11.0 Compatibility is not good , There will be a path error prompt after the training . I suggest 1.8.0 edition , Of course 9 10 Version can also test the following .
Go to the previous version page .
You can see , There are basically two methods for online installation
One is conda One is pip


conda Method support to cuda 11.3 pip Support to 11.1 The two methods pytorch Same version , Corresponding torchvision torchaudio The version of is also slightly different . But there is no problem installing according to the corresponding version .
It's starting to work conda For installation cuda 11.3

conda install pytorch==1.10.1 torchvision==0.11.2 torchaudio==0.10.1 cudatoolkit=11.3 -c pytorch -c conda-forge

The results are installed in pytorch1.10 when （ Or other versions ） Very slow , Need about 2 In hours or so .

I used pip edition

pip install torch==1.8.0+cu111 torchvision==0.9.0+cu111 torchaudio==0.8.0 -f https://download.pytorch.org/whl/torch_stable.html

from pytorch Official website installation , Very fast ,2 Finish loading in minutes , After that, you may be prompted that some dependencies are not installed , It doesn't matter. We can test it first pytorch and GPU

It shocked me ,gpu The environment has been built , In fact, it can be found during installation , Installed version pytorch It's with cuda and cudnn Of .

The following routine is the same as before , hold yolov5-master decompression （ The previous image has been unzipped ）, Then install dependencies in the current directory

pip install -r requirements.txt

After installation , The environment is set up , Let's test first coco128 Data set training , Get into yolov5-master Folder , perform

python train.py --img 640 --batch 16 --epochs 5 --data ./data/coco128.yaml --cfg ./models/yolov5s.yaml --weights yolov5s.pt

There is an error , The page file is too small

because cpu There are no mistakes in training , It can be associated with GPU The existing is too small , Can reduce batch solve

python train.py --img 640 --batch 2 --epochs 5 --data ./data/coco128.yaml --cfg ./models/yolov5s.yaml --weights yolov5s.pt

Start training , Correctly identify GPU

During the training, you can see gpu Video memory usage , monitor cpu You can find cpu Very low occupancy

At the same time, the speed has increased a lot , All in all, only 4 minute .CPU Training should be 20 Minutes or more .
Use the low-end graphics card to complete gpu I didn't have much hope for it , The following research will rely on GTX1070 8G and RTX2060s 8G Graphics card to complete , We'll see .