How to train using coco_100img.data

[Kev1nZheng]

Thanks for your work. I want to train a small dataset like coco-1000img.data and I see you posed Visualize Results for coco-100img.data using python3 train.py --data data/coco_100img.data. But do I need to change other hyparamiters? like LR, Optimizer,etc.

[glenn-jocher]

I'm going to close this as it is a duplicate of #192

But to answer your question the tutorial results are created simply by running train.py and pointing it to the relevant *.data file. To be extra clear I've added a section on reproducing the tutorial results to the end of it. Please review the custom data tutorial thoroughly: https://docs.ultralytics.com/yolov5/tutorials/train_custom_data

Reproduce Our Results

To reproduce the exact training results shown in this tutorial, simply run the following code. This trains each of the 3 scenarios plotted above to 273 epochs, saves each results*.txt file separately, and plots them together as results.png. It all takes less than 30 minutes on a GCP VM V100 instance created by our GCP Quickstart Guide.

git pull  # download latest updates
rm results.txt  # remove existing results
python3 train.py --data data\coco_1img.data && mv results.txt results_1img.txt
python3 train.py --data data\coco_10img.data && mv results.txt results_10img.txt
python3 train.py --data data\coco_100img.data && mv results.txt results_100img.txt
python3 -c "from utils import utils; utils.plot_results()"