Skip to content

Latest commit

 

History

History
55 lines (39 loc) · 2.17 KB

setup.md

File metadata and controls

55 lines (39 loc) · 2.17 KB

Run & Test

Structure of the project :

  • train.py
    • to launch a training set based on parameters
  • dataset folder
    • Where 2 versions of the dataset is present, including image and labels
    • one dataset is the grayscaled version of the first one, to consider performance differences
  • test folder
    • Used to run scripts after the training
      • video.py to launch the webcam and hold signs in front to try any fine-tuned model
      • test.py to test the fine-tuned model on a whole folder (dataset_test) of images

Prerequisites

Start by cloning the project on your machine

git clone https://github.com/PingoLeon/Yolo-Road.git

You will need a recent version of Python (like 3.12) with multiple dependencies :

  • Ultralytics

  • Torch

  • CUDA toolkit (If using Nvidia GPU)

    • try nvcc --version in a terminal to ensure that CUDA toolkit is installed
    • also try running torch.cuda.is_available() to ensure it is available, and thus ensure maximum possible speed on your Nvidia GPU
    >>> import torch
>>> torch.cuda.is_available()
True

Versions used for this project :

  • Ultralytics 8.3.13
  • Torch 2.4.1+cu118 (checkout this page to get the install command corresponding your system requirements)
  • CUDA 11.8 (download here)
pip install ultralytics torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118 opencv-python pyyaml

Additionals requirements :

Warning

You will need to run the scripts from the original repo folder reference, else the paths will be messed up

Once you're done with all the requirements, just hit python train.py to start the training. Training parameters are in the beginning of the train.py file.

For testing, place your test images in test/test_images/dataset_test folder to be able to run the test/test.py code. you have to put the name of the run version yourself to choose which model you would like to test