HYU-AUE8088, Understanding and Utilizing Deep Learning

PA #2. Object Detection

Important Files

├── README.md
├── requirements.txt
├── datasets
│   └── nuscenes/ (see below explanation)
├── data
│   ├── ...
│   └── nuscenes.yaml
├── models
│   ├── ...
│   ├── yolo.py
│   └── yolo5n_nuscenes.yaml
├── utils
│   ├── ...
│   ├── dataloaders.py
│   └── loss.py
├── detect.py
├── debug.ipynb
└── train_simple.py

Preparation

• Prepare dataset (4.3GB, resized images with bbox labels, front camera only)

  $ wget https://hyu-aue8088.s3.ap-northeast-2.amazonaws.com/nuscenes_det2d.tar.gz
  $ tar xzvf nuscenes_det2d.tar.gz

• Create python virtual environment

  $ python3 -m venv venv/aue8088-pa2
  $ source venv/aue8088-pa2/bin/activate

• Check whether the virtual environment set properly : The result should end with venv/aue8088-pa2/bin/python.

  $ which python

• Clone base code repository (replace ircvlab to your account if you forked the repository)

  $ git clone https://github.com/ircvlab/aue8088-pa2

• [!] Create a symbolic link for nuscenes dataset

  • Assume the below folder structure

    ├── nuscenes_det2d
    ├── aue8088-pa2
    │   ├── data/
    │   ├── models/
    │   ├── train_simple.py
    │   ├── ...
    │   └── README.md (this file)

  • Follow below commands

    $ cd aue8088-pa2
    $ mkdir datasets
    $ ln -s $(realpath ../nuscenes_det2d) datasets/nuscenes
    $

• Install required packages

  $ pip install -r requirements.txt

Train

• Command

  $ python train_simple.py \
    --img 416 \
    --batch-size 64 \
    --epochs 40 \
    --data data/nuscenes.yaml \
    --cfg models/yolov5n_nuscenes.yaml \
    --weights yolov5n.pt \
    --workers 16 \
    --name yolov5n