zeenolife / age-estimation

age-estimation

Train age-estimation model inside a docker container

Data preparation

Please download the following datasets:

• IMDB faces only dataset - link
• UTKFace in-the-wild dataset all 3 parts - link
• APPA-REAL dataset - link
• SoF dataset - link. Please extract the SoF dataset manually because it has spaces in its name.

The datasets directory should have the following structure:

.
├── appa-real-release.zip   # APPA-REAL dataset
├── part1.tar.gz            # UTKFace part1
├── part2.tar.gz            # UTKFace part2
├── part3.tar.gz            # UTKFace part3
├── imdb_crop.tar           # IMDB dataset
├── original images.rar     # SoF dataset archived
└── original images         # SoF dataset unarchived

Docker

Build docker image

cd age-estimation
docker build -t age-estimation:v1

Run the docker image and link the datasets folder to /age-estimation/data

nvidia-docker run -p 8888:8888 -v /path/to/datasets_folder/:/age-estimation/data -d age-estimation:v1

This will automatically run Jupyter Notebook with 8888 port, so it can be accessed from the host. To get the token for jupyter notebook run the command docker ps and find the container name and docker logs container_name afterwards.

Training

The train_by_parts.ipynb does the data preparation and training.