Shulou(Shulou.com)06/03 Report--

In this issue, the editor will bring you about whether the PyTorch version of YOLOv4 is suitable for custom datasets. The article is rich in content and analyzes and describes for you from a professional point of view. I hope you can get something after reading this article.

It has been more than 5 months since the launch of YOLO v4. The YOLO framework uses C as the underlying code, which is a bit unfriendly for researchers who are accustomed to Python. As a result, there are many replicated versions of YOLO based on various deep learning frameworks. Recently, some researchers have updated the PyTorch-based YOLOv4 on GitHub.

Since the release of YOLOv4 in April this year, the most frequently asked question about this target detection framework is: "do you have any students who can reproduce YOLOv4? can you communicate with each other?" Because the original YOLO uses C language for programming, this alone deters many students. There are many reproduction versions based on TF/Keras and Caffe on the Internet, but many projects only give the code and do not give the training results of the model on the COCO and PASCAL VOC data sets.

Recently, some researchers have opened up an open source project on GitHub: a YOLOv4 replica based on the PyTorch deep learning framework, which is based on the implementation AlexeyAB/darknet given by the YOLOv4 author and runs on PASCAL VOC, COCO, and custom datasets.

Project address: https://github.com/argusswift/YOLOv4-PyTorch

In addition, the project adds some useful attention methods to the backbone and implements mobilenetv2-YOLOV4 and mobilenetv3-YOLOV4.

Attentive YOLOv4

The project adds some attention methods to the backbone network, such as SEnet, CBAM.

SEnet (CVPR 2017)

CBAM (CVPR 2018)

Mobilenet YOLOv4

The study also implemented mobilenetv2-YOLOV4 and mobilenetv3-YOLOV4 (just change the MODEL_TYPE in config/yolov4_config.py).

The following table shows the performance results of mobilenetv2-YOLOV4:

Now let's look at the details and requirements of the project.

Environmental requirements

Nvida GeForce RTX 2080TI

CUDA10.0

CUDNN7.0

Windows or linux system

Python 3.6

Characteristics

DO-Conv (https://arxiv.org/abs/2006.12030) (torch > = 1.2)

Attention

Fp_16 training

Mish

Custom data

Data Augment (RandomHorizontalFlip, RandomCrop, RandomAffine, Resize)

Multi-scale Training (320 to 640)

Focal loss

CIOU

Label smooth

Mixup

Cosine lr

Install dependencies

Run the script to install the dependency. You need to provide the conda installation path (for example, ~ / anaconda3) and the name of the conda environment you created (YOLOv4-PyTorch here).

It is important to note that the installation script has been tested on Ubuntu 18.04 and Window 10 systems. If there is a problem, please check the detailed installation instructions: https://github.com/argusswift/YOLOv4-PyTorch/blob/master/INSTALL.md.

Preparatory work

1. Git copy YOLOv4 library

The first step in preparation is to copy the YOLOv4.

Then update "PROJECT_PATH" in the configuration file "config/yolov4_config.py".

two。 Data set preparation

The project prepares Pascal VOC and MSCOCO 2017 datasets. The PascalVOC dataset includes VOC 2012_trainval, VOC2007_ trainval and VOC2007_test,MSCOCO 2017 datasets including train2017_img, train2017_ann, val2017_img, val2017_ann, test2017_img, and test2017_list.

PascalVOC dataset download command:

MSCOCO 2017 dataset download command:

After the dataset has been downloaded, you need to do the following:

Place the dataset in the directory and update the DATA_PATH parameter in config/yolov4_config.py.

(for COCO datasets) use coco_to_voc.py to convert COCO data types to VOC data types.

Convert data format: use utils/voc.py or utils/coco.py to convert pascal voc * .xml format (or COCO * .json format) to * .txt format (Image_path xmin0,ymin0,xmax0,ymax0,class0 xmin1,ymin1,xmax1,ymax1,class1...).

3. Download weight file

1) darknet pre-training weight: yolov4 (https://drive.google.com/file/d/1cewMfusmPjYWbrnuJRuKhPMwRe_b9PaT/view).

2) Mobilenet pre-training weight:

Mobilenetv2: (https://pan.baidu.com/share/init?surl=sjixK2L9L0YgQnvfDuVTJQ, extraction code: args)

Mobilenetv3: (https://pan.baidu.com/share/init?surl=75wKejULuM0ZD05b9iSftg, extraction code: args).

3) create a weight folder under the root directory, and place the downloaded weight file in the weight / directory.

4) set MODEL_TYPE in config/yolov4_config.py during training.

4. Convert to a custom dataset (training based on a custom dataset)

1) put the pictures of the custom dataset in the JPEGImages folder and the comment files in the Annotations folder.

2) use the xml_to_txt.py file to write the list of training and test files to ImageSets/Main/*.txt.

3) convert data format: use utils/voc.py or utils/coco.py to convert pascal voc * .xml format (or COCO * .json format) to * .txt format (Image_path xmin0,ymin0,xmax0,ymax0,class0 xmin1,ymin1,xmax1,ymax1,class1...).

Training

Run the following command to start the training, see config / yolov4_config.py for details. DATA_TYPE should be set to VOC or COCO during training.

It also supports resume training, adds-- resume, and automatically loads last.pt using the following command.

Detection

Modify the detection image path: DATA_TEST=/path/to/your/test_data# your own images.

The results can be viewed in output /, as shown below:

Evaluation (Pascal VOC dataset)

Modify the path of the evaluation dataset: DATA_PATH=/path/to/your/test_data # your own images

Evaluation (COCO dataset)

Modify the path of the evaluation dataset: DATA_PATH=/path/to/your/test_data # your own images

Visual heat map

Set showatt=Ture in val_voc.py and the network can output heat maps.

You can view the heat map in output /, as shown below:

The above is whether the PyTorch version of YOLOv4 shared by the editor is suitable for custom datasets. If you happen to have similar doubts, you might as well refer to the above analysis to understand. If you want to know more about it, you are welcome to follow the industry information channel.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.