Using YOLOv7 for crop and weed detection

Used dataset: The Crop and Weed dataset

Used YOLOv7 version: https://github.com/Chris-hughes10/Yolov7-training/issues

Installation

Clone the repository

You can download the source code including the models from the latest release.

# This prevents downloading the large model files; 
# optional, skip this line to include them
export GIT_LFS_SKIP_SMUDGE=1
# Clone the repository and the two submodules
git clone --recurse-submodules https://github.com/Daraan/CropAndWeedDetection.git

If you forgot the --recurse-submodules you can still download the submodules with:

# cd CropAndWeedDetection
git submodule init
git submodule update

Note: The currently linked Yolov7 variant is not compatible with half precision training. It is possible, however, I probably cannot assist you in this matter anymore.

---

Create a virtual environment

Choose a python version and set a location for the environment

python3.9 -m venv env
source env/bin/activate

Optional: Assure there is a pip in the environment. On my HPC-cluster this was wrong in some cases.

python3.9 -m pip install -U pip --no-cache-dir --force-reinstall  

Requirements

IMPORANT for WINDOWS users : Do not install the cropandweed-dataset and Yolov7-training via pip, use the cloned repositories provided through the submodule.

PyTorch-Accelerated: is integrated into the YOLOv7 code but not directly used. It is not so well maintained and might downgraded you to a PyTorch version < 2, this installation command prevents the downgrade:

pip install pytorch-accelerated==0.1.40 --no-dependencies  

---

Requirements (minimal):

The code was written with Python 3.9.

The CLI requirements where created by pipreqs and tested the last time this in April 2024.

pip install -r requirements_CLI.txt 

For the notebook or if you encounter problems you can try with a more restrictive installation, acquired through pip freeze:

pip install -r requirements_complete.txt 

Optional: not needed for this notebook, but optionally for supplementary code. There might be CUDA path problems therefore not putting it into requirements

pip install deepspeed 

Light version

For the .py files a lighter version

pip install -r pipreqs_requirements.txt 

Todos

• Create fork for half precision support of Yolov7 (#wontfix)