This repository contains code and resources for training YOLOv8 models for object detection with a focus on evaluating the environmental impact of the training process. The project utilizes different configurations of datasets, including augmented and concatenated versions, to assess model performance and carbon emissions.
To set up the environment, install the required Python packages listed in requirements.txt:
pip install -r requirements.txt
The main objectives of this project are:
- Data Augmentation: Apply various data augmentation techniques to enhance the training dataset.
- Train YOLOv8 Models: Train object detection models using various dataset configurations.
- Track Emissions: Monitor and report the carbon emissions produced during training.
- Analyze Results: Evaluate model performance based on accuracy metrics and environmental impact.
The dataset is organized within the Larch_Daset folder. Two subsets are present: one with annotations (folder names ending with '7') and another without annotations (folder names ending with '9'). In both cases, bounding box annotations and tree types are included. Unfortunately, this dataset cannot be pushed on git because of its size 3.6 GB, so the following part is about this. However the dataset used for training is on the git in the Data so you should not need to do the following step anymore. If so go on the train tab.
To obtain the data:
- Click on this link to download the Larch Dataset.
- Extract the directory.
- Open the extracted directory.
- Copy all the folders (
Ctrl+A,Ctrl+C). - Go back to your local repo, create a
Larch_Dataset/folder. - Paste everything into your
Larch_Dataset/folder.
You should have an architecture like this:
Larch_Dataset
├── Bebehojd_20190527
├── Bebehojd_20190819
├── Ekbacka_20190527
├── Ekbacka_20190819
├── Jallasvag_20190527
├── Jallasvag_20190819
├── Kampe_20190527
├── Kampe_20190819
├── Nordkap_20190527
└── Nordkap_20190819
From this point on, you can preprocess the data by running the file preprocessing.py. This script will create a Data/ folder and divide the data into three parts: Train, Test, and Validation. Simultaneously, the script transforms the label files from XML to text files using the YOLO format, and resize the images to the YOLO format as well which is (640,640). For more information about the YOLO format, refer to this link.
Upon completing this step, you should still have your original Larch_Dataset/ alongside the newly created Data/. If desired, you can remove the Larch_Dataset/ directory to free up space. Your Data/ directory should have the following structure:
Data
├── test
│ ├── images
│ └── labels
├── train
│ ├── images
│ └── labels
└── valid
├── images
└── labels
For the annotated subset:
- Bounding Box Annotations: Present for all images.
- Damage Annotations:
- H: Healthy
- LD: Light Damage
- HD: High Damage
- Other: Unspecified
The tree type is also annotated:
- Larch: Pertaining to the trees of interest.
- Other: Denoting trees that are not relevant to the project.
To enhance the training dataset, you need to run the datamodule.py script. This script performs data augmentation and concatenates datasets to improve model training. Execute the following command to run the data augmentation and concatenation process:
python datamodule.py
This script generates augmented datasets using various techniques, concatenates them with the original datasets, and saves the results to the appropriate directories.
Before to train you need to import the yolov8 repository. You need to be in your project repo and clone the yolov8 repo with this command:
git clone [email protected]:ultralytics/yolov8.git
From here you can try to train the model. I provide here an example of command to pass through the terminal but you can change every parameter you want following the ones in the train.py file of the yolov8 repo. Also you need to be in your repo in the yolov8 repo to execute the following command:
python train.py --data ../data.yaml --weights yolov8n.pt --img 640 --epochs 10 --batch-size 16
You can change the batch size to 8 if you are struggling running the training. You can also increase or decrease the amount of epochs.
Our training process tracks the carbon emissions using the EmissionsTracker. This feature monitors CO2 emissions produced during training and converts it to energy consumption (kWh) using the conversion factor CO2_TO_KWH.
To test the models on your images, use Streamlit to run the demo:
streamlit run front.py
Run the analysis.py script to generate plots and rankings:
python analysis.py