/hackathon2024_example_model

This repository demonstrates how to train a YOLOv8 model on plant dataset.

Primary LanguagePythonMIT LicenseMIT

ncpsi-hackathon-2024-example

This repository demonstrates how train a YOLO model and standardize predictions for leaderboard.

Pre-requisites :

  • Installing Conda :

    Install miniconda by following the official website.

  • Check Installation

    conda --version
>> conda 24.7.1

  • Create a conda environment

conda env create -f environment.yml
  • Start the environment
conda activate example_model
  • To exit the environment
conda deactivate

run the scripts in the conda environment

Training a yolov8 model

Here are a few steps to prepare your data for training a YOLO model.

  • Split your data into train , val and test.

YOLO expects the data to be split into small sub directories

    /datasets
    |   /images
    |    |   /train
    |    |   | img1.jpg
    |    |   | img2.jpg
    |    |   /val
    |    |   | imga.jpg
    |    |   | imgb.jpg
    |    |   /test
    |    |   | img00.jpg
    |    |   | img01.jpg
    |    /labels
    |    |   /train
    |    |   | img1.txt
    |    |   | img2.txt
    |    |   /val
    |    |   | imga.txt
    |    |   | imgb.txt
    |    |   /test
    |    |   | img00.txt
    |    |   | img01.txt
    |    plant_data.yaml

Note : YOLO requires the sub-folders to be named datasets, images and labels.

The plant_data.yaml file contains directory paths and class names. Example :

# set paths to train, val and test data.
train: images/train
val: images/val
test: images/test

# list class names
names:
  0: background
  1: Palmer amaranth
  2: Common ragweed
  3: Sicklepod
  • There are two scripts provided :
  1. yolo_train_data.py :

This file trains a YOLOv8 model, on the dataset provided in plant_data.yaml.

Import the model

model = YOLO('yolov8n.pt')
print(model.info())

Define some hyperparaters

epochs = 5
batch_size = 5
lr = 0.001

Train the model

model.train(data = plant_data_yml_fpath,
                epochs=epochs,
                batch=batch_size,
                lrf = lr,
                cache=False)

Save the model

model.save('yolov8_example.pt')

Read more about training a YOLO model in [1].

  1. output_yolo_predictions.py

This file creates a submission file <team-name>_predictions.csv.

Example team1_predictions.csv :

image_name,class,confidence,x,y,width,height
img1.jpg,7.0,0.7846097350120544,0.6045709252357483,0.8606472015380859,0.10770443081855774,0.27489542961120605
img2.jpg,0.0,0.6409725546836853,0.7783212661743164,0.8416421413421631,0.10649671405553818,0.12297263741493225
img3.jpg,0.0,0.5564624071121216,0.7094271183013916,0.4191618859767914,0.0850377157330513,0.15624332427978516

References :

  1. https://docs.ultralytics.com/modes/train/#multi-gpu-training
  2. Organize your data : https://docs.ultralytics.com/yolov5/tutorials/train_custom_data/#option-1-create-a-roboflow-dataset