This is a pure pytorch implementation for YOLOv1. Note that this is not an official implementation, only for me to learn details of Object Detection and entertainment.
In the process of my study, I found that it is not easy to build DarkNet environment, especailly on Windows, so I wrote this repo, YOLOv1 in pure pytorch. I'm going to write a simple and easy-to-use reproduction version of yolov1.
To skip the pre-train process, I directly use ResNet-18 model inside PyTorch (pre-trained on ImageNet-1000 dataset) without the last two classify layers instead of the first 20 layers in the original paper. See more details at yolo.model.ResNet18_YOLOv1.
Then I trained my model on the VOC-2007-trainval dataset and validated on the VOC-2007-test dataset. After about 2 days training, I got the best mAP of 58.2 at last.
There are some gaps with the original. The paper said that they got the best mAP of 63.4. But they got this data after training on the VOC-2007 & VOC-2012 trainval set and validated on the VOC-2012-test. It was the different training data and different model structure made the gaps.
Here are some visual data of training process.
| Loss | mAP |
|---|---|
.svg) |
.svg) |
The following sections are under preparation...
The part will guide you to detect a single image with my model.
conda create -n yolov1_pytorch python=3.7
conda activate yolov1_pytorch