aws-darknet-yolo-docker

Nvidia based docker image with Darknet to train a yolov3-tiny network on your custom images

Works seemlesly with AWS EC2 and S3

Steps:

1. Create a S3 bucket on AWS
2. Upload images, train.txt and test.txt lists to the S3 bucket. Expected structure of the bucket

your-s3-bucket/
└── data/
    ├── train.txt
    ├── test.txt
    ├── images/
    |   ├── image-name-1.jpg
    |   ├── image-name-2.jpg
    |   └── ...
    └── labels/
        ├── image-name-1.txt
        ├── image-name-2.txt
        └── ...

train.txt and test.txt should contain image relative paths to the data/ folder in S3, ex.

data/images/image-name-1.jpg
data/images/image-name-2.jpg

3. Modify coco.names, config.data and yolov3-tiny.cfg as described by https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects

4. Set up your EC2 instance

   1. Create a custom role in IAM settings for EC2, which allows access to your S3 bucket.
   2. Create a GPU instance on AWS which already has NVIDIA and CUDA pre-installed. Search for "Deep Learning" when choosing your instance AMI
   3. Ensure you set your created role as an IAM role of the instance
   4. SSH into the instance. More info
   5. clone this repo

   git clone https://github.com/deividasskiparis/aws-darknet-yolo-docker.git aws-darknet-yolo-docker && cd aws-darknet-yolo-docker
   

5. Build docker image

docker build --build-arg S3=your-s3-bucket-name -t train-yolo .

6. Run image

docker run -d --runtime=nvidia --name=train-yolo train-yolo

NOTE: If running locally instead on EC2, you will have to provide AWS credentials to your docker container for a role, that has access to your S3 bucket

docker run -d --runtime=nvidia --name=train-yolo -e "AWS_ACCESS_KEY_ID=*********" -e "AWS_SECRET_ACCESS_KEY=*********" train-yolo

7. After the training is finished, the trained weights will be uploaded to you S3 bucket /models directory

8. Terminate your EC2 after the training is finished to avoid charges