This project focuses on training the YOLOX model for the task of object detection of people and forklifts, and additionally for the task of magnetic tape detection. The original README file can be found here.
git clone git@github.com:manojlovska/YOLOX-DAIS.gitconda create -n env_name python=3.8.5
conda activate env_namecd YOLOX-DAIS
pip install -v -e .
pip install -r requirements.txtcd datasets/
# TODOThe link for downloading the pretrained YOLOX-Darknet53 model is provided here.
python tools/train.py -f train_custom_data.py -d 1 -b 16 --fp16 -c /path/to/your/yolox_darknet.pth- -f: configuration file of the experiment
- -d: number of GPU devices
- -b: batch size
- --fp16: mixed precision training
- -c: .pth file of the pretrained model
(Optional) If you want to visualize the training in real-time, use the following command:
python tools/train.py -f train_custom_data.py -d 1 -b 16 --fp16 -c /path/to/your/yolox_darknet.pth --logger wandb wandb-project <project-name>NOTE: You have to change the name of the project in the train_custom_data.py as well:
os.environ['WANDB_PROJECT'] = 'project-name'
run = wandb.init(project='project-name')For training the magnetic tape detection head, we use the best trained model for object detection and freeze its weights. Use the following command:
python tools/train.py -f train_yolino_freeze_backbone.py -d 1 -b 16 -c /path/to/your/best_ckpt.pth --logger wandb wandb-project <project-name>Again, logging to Weights and Biases is optional, and you have to change the project name in the train_yolino_freeze_backbone.py script as well as before.
For training the model for magnetic tape detection only, without freezing the backbone, please use the following command:
python tools/train.py -f train_yolino.py -d 1 -b 8 --fp16 -c /path/to/your/yolox_darknet.pth --logger wandb wandb-project <project_name>- In train_yolino.py change the basic_lr_per_img parameter
self.basic_lr_per_img = 0.01 / 64.0 => self.basic_lr_per_img = wandb.config.lr /
88C9
64.0- In /yolox/models/yolino_head.py change the p parameter
p = 0.5 => p = wandb.config.loss_param- In train_yolino.py, function get_dataset() change the sweeps parameter
return DAISDataset(
data_dir=self.data_dir,
json_file=self.train_ann,
img_size=self.input_size,
mag_tape=self.mag_tape,
preproc=TrainTransformYOLinO(sweeps=True),
cache=cache,
cache_type=cache_type,
)wandb sweep --project <propject-name> wandb_sweeps.yamlwandb agent <sweep-ID>Change the parameters as mentioned in the previous chapter.
sudo singularity build container_name.sif Singularity.defOn the login node initialize the wandb sweep.
wandb sweep wandb_sweeps.yamlsbatch sweeps_sbatch.sh wandb agent <sweep-ID>NOTE: In the sweeps_sbatch.sh script adapt the directives according to your specific requirements. Also, be careful to bind the correct path to the yolox module.
singularity exec --bind /path/to/your/YOLOX-DAIS:/workspace/yolox --nv container_name.sif "$@"