# Pytorch-YOLOv3

**Repository Path**: firslov/YOLOv3-self

## Basic Information

- **Project Name**: Pytorch-YOLOv3
- **Description**: YoloV3 by Hao.
- **Primary Language**: Python
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2022-02-04
- **Last Updated**: 2023-01-29

## Categories & Tags

**Categories**: Uncategorized

**Tags**: Python, Deep-learning, Computer-vision

## README

# YoloV3_self

A YoloV3 model using pytorch.

> - Dataset: [十分类交通标志物数据集](https://aistudio.baidu.com/aistudio/datasetdetail/20495)
> - Backbone: darknet53
> - Loss function:
>   - BCE: x、y、confidence、class
>   - MSE: w、h
> - Anchors: COCO数据集9种尺寸聚类框
> - Features: gpu加速、学习率更新、断点续训、类内NMS...

## Loss function

![](fig/loss_func.jpg)

## Usage

### Install

```shell
git clone --depth=1 https://gitee.com/firslov/YOLOv3-self.git
```

### Config

| 配置项     | 默认值              |
| ---------- | ------------------- |
| 基本配置   | ./cfg/cfg.yaml      |
| 分类类别   | ./cfg/classes.names |
| 预训练权重 | ./weight/yo.pth     |
| 配色文件   | ./cfg/pallete       |

### Train

```shell
python yolo_train.py # 断点续训: cfg.yaml ==> resume=True
```

### Inference

| 配置项      | 功能               | 默认值           |
| ----------- | ------------------ | ---------------- |
| input_dir   | 检测目录或图片路径 | ./eval           |
| output_dir  | 推断图片存放路径   | ./output         |
| confidence  | 识别置信度阈值     | 0.5              |
| nms_thresh  | nms阈值            | 0.4              |
| weight_file | 预训练权重         | 默认Xavier初始化 |
| in_dim      | 输入图片边长       | 416              |

```shell
python yolo_detect.py
```

## Show

### Inference

![](fig/1.jpg)
![](fig/2.jpg)
![](fig/3.jpg)
![](fig/4.jpg)
![](fig/5.jpg)
![](fig/6.jpg)
![](fig/7.jpg)
![](fig/8.jpg)

### Train

- total loss

![](fig/loss_total.png)

- x loss

![](fig/loss_x.png)

- y loss

![](fig/loss_y.png)

- w loss

![](fig/loss_w.png)

- h loss

![](fig/loss_h.png)

- confidence loss

![](fig/loss_conf.png)

- class loss

![](fig/loss_cls.png)