Training yolo on custum dataset

This tutorail focuses on the training yolov4 on custum dataset on google colab in order to benefite from free GPU. For this purpose, we choose the retrain yolov4 to detect vehicle registration plate which is a class of dataset present on Open Images Dataset.

1. Data preparation

Open Images Dataset is an online open source dataset provided by google that consist of ~9 millions images annotated labels, object bounding boxes, object segmentation masks and more. Interestingly, is containts up to 600 objects classes which is quite interesting for developping different type of computer vision models.

1.1 Download the interested dataset

Differents modules were proposed including oidv6 and OIDv4_ToolKit to interact with Open Images Dataset and download single or multiple dataset classes. For this tutorial, we incorporated the main.py file and modules/ from OIDv4_ToolKit as it was enough to download the needed dataset. The data class we are interested in is known on Open Images Datasetas Vehicle registration plate.

To download, move into the main directory where there is the main.py and run the following line of code.

python main.py downloader --classes "Vehicle registration plate" --type_csv train --limit 1000

This will create OID directory where where the downloaded dataset is store. Next, run again the same line of code while stting the argument --type_csv to validation and --limit to 100. This will download 100 images of the the validation set.

 python3 main.py downloader --classes "Vehicle registration plate" --type_csv validation --limit 100

For more detail regarding this code we suggest to refer to OIDv4_ToolKit repository.

1.2 adapt the annotations to the yolo formate

At this step, the first action consist of editing the classes.txt to make sure it contains the right name of the data class we are interested in. In this case it should contain Vehicle registration plate.

Next, run the folowing line of code.

python format_annotations.py 

1.3 Prepare zip files

The training is to be conducted on google colab is practical to compress the training and test sets as well as the corresponding annotations into a obj.zip and test.zip files respectively.

Manual mode

Both trainset and testset to zip correspond to the output generated after executing python format_annotations.py. The process is as simple as navigating manuallyt into OID/Dataset/train and OID/Dataset/validation subdirs. In principle, one is expected to find a single class dir name Vehicle registration plate en each of both train and validation subdirs. Hence, remane Vehicle registration plate from train subdir to obj/ and remane Vehicle registration plate from validation subdir to test/. Then navigate into both obj/ and test/ and delete the Label/ subdir. Next, manually zip both obj/ and test/ directories to obj.zip and test.zip file.

Preferable mode

To compress both train and test sets, simply run the following line of code.

     python prepare_zip_file.py

This will generate both obj.zip and test.zip files into train_yolo/ root directory.

2. Training procedure on google colab

First, open your google drive session. Create in My Drive/ a directrory name train_yolo and upload into it both obj.zip and test.zip files. The name of train_yolo is not of much important here. It can be any other name as fare as one remenber it when refering into during training.

Now, open your google colab session and upload the train_yolo.ipynb notebook file and start following the instructions.

Structure of the repository

.
├── classes.txt
├── format_annotations.py
├── main.py
├── modules
│   ├── bounding_boxes.py
│   ├── csv_downloader.py
│   ├── downloader.py
│   ├── image_level.py
│   ├── parser.py
│   ├── show.py
│   └── utils.py
├── OID
│   ├── csv_folder
│   │   ├── class-descriptions-boxable.csv
│   │   ├── train-annotations-bbox.csv
│   │   └── validation-annotations-bbox.csv
│   └── Dataset
├── Pipfile
├── prepare_zip_file.py
├── README.md
└── train_yolo.ipynb