This is the repository for camera related algorithm models that are used in Elevator Ads Platform.
The following core dependencies need to be installed manually or using Docker:
- CUDA 9
- CuDNN 7
- Python 3.5+
- OpenCV
- Pytorch >= 0.3.0
The following core depenencies can be installed through pip3 install -r requirements.txt:
- Chainer
- Mxnet
To use Docker for development, install the following dependencies on host:
- Nvidia Driver
- Docker
- Nvidia-Docker 2
You can also use this script to setup the host machine automatically.
Start the Docker container using:
sudo docker run -ti --runtime=nvidia --privileged -e="DISPLAY" -e="QT_X11_NO_MITSHM=1" -v="/tmp/.X11-unix:/tmp/.X11-unix:rw" --ipc=host -p 0.0.0.0:6006:6006 -p 8888:8888 -v /dev/video0:/dev/video0 --name eap-models deepgaze/eap-models-dev bash
You may also want to add extra -v options to map codes/IDE/data into docker container.
To restart the container, simply run:
sudo docker start -i eap-models
To open multiple docker terminal to the same container, simply run:
sudo docker exec -ti eap-models bash
- Download GazeCapture dataset
- Untar all tar files in the dataset
- Download pretrained models
bash facedet/script/download_models.sh
- Download pretrained models
bash faceattr/script/download_models.sh
Assuming the GazeCapture dataset is located at ~/fast-storage/GazeCapture, start training with
python3 train_gaze.py --root_path ~/fast-storage/GazeCapture --result_path results --dataset gazecapture --model resnet --model_depth 18 --batch_size 1024 --pretrain --log_dir results --n_epochs 50 --lr 2e-5 --n_thread 12 --checkpoint 5
Check gaze/opts.py for more training options.