This repo contains the final project for our Reinforcement Learning course. The project contains an object detection framework using deep reinforcement learning agents. The core model is a DQN whose actions transform a bounding box until it tightly bounds the desired object. The code is inspired by this work.
Create a new conda environment and download all the required packages using the following line
conda env create --file=environment.yml
To use an existing conda environment and simply update its packages for this repo, run
conda env update -name <MYENV> --file environment.yml --prune
Please note that this package installs cudatoolkit version 11.1, which may not be compatible with your GPU drivers.
Download the data by running the prep_dataset.sh script provided
in the util folder. Run the script as follows
cd util
bash prep_dataset.sh <FOLDER_TO_STORE_DATA>
To use to package for training or visualization, refer to the scripts in scripts_shuhao/local.
Here are some examples; note that the pretrained model seems to produce the best bounding boxes, so these examples are mainly from the pretrained DQN.
