jorgeluceda / duckietown-exercise-b5

duckietown-exercise-b5

Steps

1. Install required python dependencies pip install opencv-python easygui

2. Clone the github repository: https://github.com/jorgeluceda/duckietown-exercise-b5

3. Uncompress the included data_4_classes.tar.gz file, which is our dataset tar -xf data_4_classes.tar.gz

4. Create blurry_folder and non_blurry_folder in the root of the project mkdir blurry_folder non_blurry_folder

5. Apply laplacian filter to our data set using python detect_blurry_img.py data_4_classes/frames/ blurry_folder non_blurry_folder --threshold 300, using threshold of 300 so that blurrier images are filtered.

6. Run custom script ./move_labels.sh, this filters the labels given and places the ones that our threshold allowed in our non_blurry_folder, this skips the need to use the given label_data.py script so that we do not need to label a large amount of images manually.

7. Create data_folder that is used in the next steps to train YOLO mkdir data_folder

8. Execute python create_datasets.py non_blurry_folder data_folder 80 to create the dataset and needed directories.

9. Create the two other files required for training with the bash commands ls data_folder/trainset/*.jpg > data_folder/train.txt and ls data_folder/validset/*.jpg > data_folder/valid.txt

10. Copy both the architecture_file and the pretrained weights from the cfg directory using the command cp cfg/yolov3-tiny.cfg architecture_file.cfg and cp cfg/yolov3-tiny-duckie-multi.weights pretrained_weights.weights

11. Create the class_name file to be used in the next step for the names touch class_names.names and add the following lines to it:

    bot
    duckie
    stop_sign
    road_sign

This contains all of our class names and is used in the names parameter in the data_file that we'll create next.

12. Create the data_file in the root with the following command: touch data_file and add the following lines to it:

    train = data_folder/train.txt
    valid = data_folder/valid.txt
    names = class_names.names
    backup = duckie_backup

This data_file specifies the data to use while training. This follows the YOLO Convention.

13. In our previously copied architecture_file.cfg, we need to modify lines 127 and 171 to filters=27 as we have 4 classes. The formula is filters=(classes + 5)*3. We also change lines 135 and 177 to classes=4 as we are training 4 classes.

14. To start the training, we need to first download cuda: wget -O cuda_9.2.88_396.26_linux.run -c https://developer.nvidia.com/compute/cuda/9.2/Prod/local_installers/cuda_9.2.88_396.26_linux (this file is > 1GB so be ready to wait a bit).

15. Now to install cuda we need to change the file permissions using the command chmod a+x cuda_9.2.88_396.26_linux.run then we need to execute the command ./cuda_9.2.88_396.26_linux.run --verbose --silent --toolkit --override to run the CUDA installer. Remember to run this as root otherwise it will not install. Might want to skip this step for a workshop or similar environment and have CUDA pre-installed as it takes a while to install.

16. If running ubuntu the easiest way to do this is to run the command sudo apt install nvidia-cuda-toolkit, saves doing the previous 2 steps.

17. Now run the command make to compile our program. We can now call the following command to start the training process: ./darknet detector train data_file architecture_file.cfg pretrained_weights.weights

18. Now that we have trained weights from our previous training process, we can run ./darknet detector test data_file architecture_file.cfg weights_file non_blurry_folder/frames/100_000005.jpg -thresh 0.7. The threshold parameter allows us to tell the predictor to only output bounding boxes when it is highly confident of its predictions.