YOLO-recreated

Recreated YOLO algorithm for my engineering thesis.

University: Warsaw University of Technology

Supervisor: dr. inż Krystian Radlak

How to run?

0. If conda is not installed install conda (for WSL: https://gist.github.com/kauffmanes/5e74916617f9993bc3479f401dfec7da)
1. Clone the repository and enter repository

git clone https://github.com/bindas1/YOLO-recreated.git
cd YOLO-recreated

2. Install conda environment and activate it:

conda env create -f ./requirements/conda-environment_linux.yaml
conda activate yolo

3. Add required packages

pip install -r ./requirements/requirements_linux.txt

4. Create wandb account Go to https://wandb.ai/authorize?signup=true, sign in using github, google or create account. After the process is complete You should receive your API key (it can be also accessed in settings for account)

Edit train.py Change line 14 entity from bindas1 to your username:

with wandb.init(project="YOLO-recreated", entity="bindas1", config=hyp):

6. Run

python main.py

When prompted choose 2 (if wandb account not created)

Copy the API from the sign up process is complete and paste to the terminal. (This process only needs to be done once)

After the login you should see that the data is being downloaded. If You want to make sure that the installation process is complete I suggest running the code with config["is_one_batch"] set to True in main.py.

You can access all the metrics in your profile, each run produces new instance on wandb page.

FAQ

If you receive the RuntimeErrror: CUDA out of memory You should change the batch_size in config in main.py to smaller number.