RT-ODLab

PyTorch_YOLO_Tutorial

YOLO Tutorial

English | 简体中文

Introduction

Here is the source code for an introduction to YOLO. We adopted the core concepts of YOLOv1~v4, YOLOX and YOLOv7 for this project and made the necessary adjustments. By learning how to construct the well-known YOLO detector, we hope that newcomers can enter the field of object detection without any difficulty.

Book: The technical books that go along with this project's code is being reviewed, please be patient.

Requirements

• We recommend you to use Anaconda to create a conda environment:

  conda create -n yolo python=3.6
  

• Then, activate the environment:

  conda activate yolo
  

• Requirements:

  pip install -r requirements.txt 
  

My environment:

• PyTorch = 1.9.1
• Torchvision = 0.10.1

At least, please make sure your torch is version 1.x.

Training Configuration

Configuration
Per GPU Batch Size	16
Init Lr	0.01
Warmup Scheduler	Linear
Lr Scheduler	Linear
Optimizer	SGD
Multi Scale Train	True (320 ~ 640)

Due to my limited computing resources, I can not use a larger multi-scale range, such as 320-960.

Experiments

VOC

• Download VOC.

  cd <PyTorch_YOLO_Tutorial>
  cd dataset/scripts/
  sh VOC2007.sh
  sh VOC2012.sh
  

• Check VOC

  cd <PyTorch_YOLO_Tutorial>
  python dataset/voc.py
  

• Train on VOC

For example:

python train.py --cuda -d voc --root path/to/VOCdevkit -m yolov1 -bs 16 --max_epoch 150 --wp_epoch 1 --eval_epoch 10 --fp16 --ema --multi_scale

All models are trained with ImageNet pretrained weight (IP). All FLOPs are measured with a 640x640 image size on VOC2007 test. The FPS is measured with batch size 1 on 3090 GPU from the model inference to the NMS operation.

COCO

• Download COCO.

  cd <PyTorch_YOLO_Tutorial>
  cd dataset/scripts/
  sh COCO2017.sh
  

• Check COCO

  cd <PyTorch_YOLO_Tutorial>
  python dataset/coco.py
  

• Train on COCO

For example:

python train.py --cuda -d coco --root path/to/COCO -m yolov1 -bs 16 --max_epoch 150 --wp_epoch 1 --eval_epoch 10 --fp16 --ema --multi_scale

Due to my limited computing resources, I had to set the batch size to 16 or even smaller during training. I found that for small models such as *-Nano or *-Tiny, their performance seems less sensitive to batch size, such as the YOLOv5-N and S I reproduced, which are even slightly stronger than the official YOLOv5-N and S. However, for large models such as *-Large, their performance is significantly lower than the official performance, which seems to indicate that the large model is more sensitive to batch size.

I have provided a bash file train_ddp.sh that enables DDP training. I hope someone could use more GPUs to train the large models with a larger batch size, such as YOLOv5-L, YOLOX, and YOLOv7-L. If the performance trained with a larger batch size is higher, I would be grateful if you could share the trained model with me.

• Redesigned YOLOv1~v2:

Model	Backbone	Scale	IP	Epoch	APval 0.5	FPS3090 FP32-bs1	Weight
YOLOv1	ResNet-18	640	√	150	76.7		ckpt
YOLOv2	DarkNet-19	640	√	150	79.8		ckpt
YOLOv3	DarkNet-53	640	√	150	82.0		ckpt
YOLOv4	CSPDarkNet-53	640	√	150	83.6		ckpt
YOLOX-L	CSPDarkNet-L	640	√	150	84.6		ckpt
YOLOv7-Large	ELANNet-Large	640	√	150	86.0		ckpt

• YOLOv3:

Model	Backbone	Scale	Epoch	APval 0.5:0.95	APval 0.5	FLOPs (G)	Params (M)	Weight
YOLOv1	ResNet-18	640	150	27.9	47.5	37.8	21.3	ckpt
YOLOv2	DarkNet-19	640	150	32.7	50.9	53.9	30.9	ckpt

• YOLOv4:

Model	Backbone	Scale	Epoch	APval 0.5:0.95	APval 0.5	FLOPs (G)	Params (M)	Weight
YOLOv3-Tiny	DarkNet-Tiny	640	250	25.4	43.4	7.0	2.3	ckpt
YOLOv3	DarkNet-53	640	250	42.9	63.5	167.4	54.9	ckpt

• YOLOv5:

Model	Backbone	Scale	Epoch	APval 0.5:0.95	APval 0.5	FLOPs (G)	Params (M)	Weight
YOLOv4-Tiny	CSPDarkNet-Tiny	640	250	31.0	49.1	8.1	2.9	ckpt
YOLOv4	CSPDarkNet-53	640	250	46.6	65.8	162.7	61.5	ckpt

I attempted to reproduce the design philosophy of YOLOv5 but may have overlooked some details, leading to poor performance. However, I do not aim to fully replicate YOLOv5's performance, as it is too challenging and resource-intensive for me.

• YOLOX:

Model	Backbone	Scale	Epoch	APval 0.5:0.95	APval 0.5	FLOPs (G)	Params (M)	Weight
YOLOv5-N	CSPDarkNet-N	640	250	29.8	47.1	7.7	2.4	ckpt
YOLOv5-S	CSPDarkNet-S	640	250			27.1	9.0
YOLOv5-M	CSPDarkNet-M	640	250			74.3	25.4
YOLOv5-L	CSPDarkNet-L	640	250	46.7	65.5	155.6	54.2	ckpt

• YOLOv7:

Model	Backbone	Scale	Epoch	APval 0.5:0.95	APval 0.5	FLOPs (G)	Params (M)	Weight
YOLOX-L	CSPDarkNet-L	640	300	46.6	66.1	155.4	54.2	ckpt

While YOLOv7 incorporates several technical details, such as anchor box, SimOTA, AuxiliaryHead, and RepConv, I found it too challenging to fully reproduce. Instead, I created a simpler version of YOLOv7 using an anchor-free structure and SimOTA. As a result, my reproduction had poor performance due to the absence of the other technical details. However, since it was only intended as a tutorial, I am not too concerned about this gap.

Necessary instructions：

• All models are trained with ImageNet pretrained weight (IP). All FLOPs are measured with a 640x640 image size on COCO val2017. The FPS is measured with batch size 1 on 3090 GPU from the model inference to the NMS operation.

• *The reproduced YOLOv5's head is the Decoupled Head, which is why the FLOPs and Params are higher than the official YOLOv5. Due to my limited computing resources, I can not align the training configuration with the official YOLOv5, so I cannot fully replicate the official performance. The YOLOv5 I reproduce is for learning purposes only.*

• Due to my limited computing resources, I had to abandon training on other YOLO detectors, including YOLOv7-Huge and YOLOv8-Nano~Large. If you are interested in these models and have trained them using the code from this project, I would greatly appreciate it if you could share the trained weight files with me.

Train

Single GPU

sh train.sh

You can change the configurations of train.sh, according to your own situation.

You also can add --vis_tgt to check the images and targets during the training stage. For example:

python train.py --cuda -d coco --root path/to/coco -m yolov1 --vis_tgt

Multi GPUs

sh train_ddp.sh

You can change the configurations of train_ddp.sh, according to your own situation.

In the event of a training interruption, you can pass --resume the latest training weight path (None by default) to resume training. For example:

python train.py \
        --cuda \
        -d coco \
        -m yolov1 \
        -bs 16 \
        --max_epoch 300 \
        --wp_epoch 3 \
        --eval_epoch 10 \
        --ema \
        --fp16 \
        --resume weights/coco/yolov1/yolov1_epoch_151_39.24.pth

Then, training will continue from 151 epoch.

Test

python test.py -d coco \
               --cuda \
               -m yolov1 \
               --img_size 640 \
               --weight path/to/weight \
               --root path/to/dataset/ \
               --show

For YOLOv7, since it uses the RepConv in PaFPN, you can add --fuse_repconv to fuse the RepConv block.

python test.py -d coco \
               --cuda \
               -m yolov7_large \
               --fuse_repconv \
               --img_size 640 \
               --weight path/to/weight \
               --root path/to/dataset/ \
               --show

Evaluation

python eval.py -d coco-val \
               --cuda \
               -m yolov1 \
               --img_size 640 \
               --weight path/to/weight \
               --root path/to/dataset/ \
               --show

Demo

I have provide some images in data/demo/images/, so you can run following command to run a demo:

python demo.py --mode image \
               --path_to_img data/demo/images/ \
               --cuda \
               --img_size 640 \
               -m yolov1 \
               --weight path/to/weight \
               --show

If you want run a demo of streaming video detection, you need to set --mode to video, and give the path to video --path_to_vid。

python demo.py --mode video \
               --path_to_img data/demo/videos/your_video \
               --cuda \
               --img_size 640 \
               -m yolov1 \
               --weight path/to/weight \
               --show \
               --gif

If you want run video detection with your camera, you need to set --mode to camera。

python demo.py --mode camera \
               --cuda \
               --img_size 640 \
               -m yolov1 \
               --weight path/to/weight \
               --show \
               --gif

Tracking

Our project also supports multi-object tracking tasks. We use the YOLO of this project as the detector, following the "tracking-by-detection" framework, and use the simple and efficient ByteTrack as the tracker.

• images tracking

  python track.py --mode image \
              --path_to_img path/to/images/ \
              --cuda \
              -size 640 \
              -dt yolov2 \
              -tk byte_tracker \
              --weight path/to/coco_pretrained/ \
              --show \
              --gif
  

• video tracking

  python track.py --mode video \
              --path_to_img path/to/video/ \
              --cuda \
              -size 640 \
              -dt yolov2 \
              -tk byte_tracker \
              --weight path/to/coco_pretrained/ \
              --show \
              --gif
  

• camera tracking

  python track.py --mode camera \
              --cuda \
              -size 640 \
              -dt yolov2 \
              -tk byte_tracker \
              --weight path/to/coco_pretrained/ \
              --show \
              --gif
  

Tracking visualization

• Detector: YOLOv2
• Tracker: ByteTracker

Command：

python track.py --mode video \
                --path_to_img ./dataset/demo/videos/000006.mp4 \
                --cuda \
                -size 640 \
                -dt yolov2 \
                -tk byte_tracker \
                --weight path/to/coco_pretrained/ \
                --show \
                --gif

Results:

© 2025 Gogs

ページ: 257ms テンプレート: 3ms

日本語

日本語 English 简体中文 繁體中文（香港） 繁體中文（臺灣） Deutsch français Nederlands latviešu русский español português do Brasil polski български italiano suomi Türkçe čeština српски svenska 한국어 galego українська English (United Kingdom) Magyar Slovenčina Indonesian Persian Vietnamese Português Монгол Română

Javascript Licenses WEBサイト

Model	Backbone	Scale	Epoch	APval 0.5:0.95	APval 0.5	FLOPs (G)	Params (M)	Weight
YOLOv7-T	ELANNet-Tiny	640	300	38.0	56.8	22.6	7.9	ckpt
YOLOv7-L	ELANNet-Large	640	300	48.0	67.5	144.6	44.0	ckpt