age-gender

This repo contains code to train age / gender prediction and run inference on a flask server. The pytorch model training / testing was copied using this template.

Check out this colab for demo.

If you are only interested in model inference, go to this section.

Prerequisites

1. A unix or unix-like x86 machine
2. Docker. Don't be scared by docker. It's really easy and convenient
3. python 3.7 or higher. Running in a virtual environment (e.g., conda, virtualenv, etc.) is highly recommended so that you don't mess up with the system python.

Datasets

Click to expand!

I used Adience age and gender dataset. Download the data and place them at ./data/Adience/. After downloading them, your data directory should look something like this:

data
└── Adience
    ├── aligned
    ├── faces
    ├── fold_0_data.txt
    ├── fold_1_data.txt
    ├── fold_2_data.txt
    ├── fold_3_data.txt
    └── fold_4_data.txt

You can find the state of the art models at here for the age and here for the gender, respectively.

I also used the IMDB-WIKI dataset. This dataset is huge. It's got more than 500,000 faces with gender and age labeled. One weird thing is that this dataset doesn't have train / val / test splits. This dataset is pretty much only used to pre-train your model. I don't know why but that is what it is. People don't compare their scores against this dataset. So I'll do the same. I'll use this dataset to improve my training and report the final metrics on the Adience age and gender dataset. Anyways, download imdb_crop.tar and wiki_crop.tar, and place them at ./data/imdb_crop and ./data/wiki_crop, respectively.

Data pre-processing

Click to expand!

I advise you that you run all of below in a virutal python environment.

Pull and run the face-detection-recognition docker container.

For the CPU model

docker pull tae898/face-detection-recognition

For the GPU model

docker pull tae898/face-detection-recognition-cuda

The detailed insturctions can be found here.

Extract the arcface face embeddings.

It might take some time.

The port number 10002 is the port that the face-detection-recognition docker container listens to. Set cuda=True in the below code snippet, if you want to run on a NVIDIA GPU). The face embedding vectors are pickled. They are saved as <image-path>.pkl (e.g., landmark_aligned_face.2174.9523333835_c7887c3fde_o.jpg.RESIZED.pkl).

Resizing image to the same shape (e.g., resize=640 resizes every image to a black background square RGB image with the width and height being 640 pixels) dramatically increase the speed due to some mxnet stuff that I'm not a big fan of.

det_score is the confidence score on face detection. The faces whose confidence score is lower than this this threshold value will not be considered.

1. Adience age and gender dataset

   At the root of this repo, run the below command.

   python -c "from utils.scripts import extract_Adience_arcface; extract_Adience_arcface('aligned', docker_port=10002, cuda=False, resize=640)"

   The argument aligned means that we'll be using the aligned face images, not raw. The aligned images have the face of interest in the center, which makes it easier to find the face.

   python -c "from utils.scripts import get_Adience_clean; get_Adience_clean('aligned', resize=640, det_score=0.9)"

   This will write ./data/Adience/meta-data-aligned.json and ./data/Adience/data-aligned.npy

2. IMDB

   At the root of this repo, run the below command.

   python -c "from utils.scripts import extract_imdb_wiki_arcface; extract_imdb_wiki_arcface('imdb', docker_port=10002, cuda=False, resize=640)"

   python -c "from utils.scripts import get_imdb_wiki_clean; get_imdb_wiki_clean('imdb', resize=640, det_score=0.9)"

   This will write ./data/imdb_crop/meta-data.json and ./data/imdb_crop/data.npy

3. WIKI

   At the root of this repo, run the below command.

   python -c "from utils.scripts import extract_imdb_wiki_arcface; extract_imdb_wiki_arcface('wiki', docker_port=10002, cuda=False, resize=640)"

   python -c "from utils.scripts import get_imdb_wiki_clean; get_imdb_wiki_clean('wiki', resize=640, det_score=0.9)"

   This will write ./data/wiki_crop/meta-data.json and ./data/wiki_crop/data.npy

Dataset stats

1. Adience age and gender dataset

   This dataset has five folds. The performance metric is accuracy on five-fold cross validation.

   images before removal	fold 0	fold 1	fold 2	fold 3	fold 4
   19,370	4,484	3,730	3,894	3,446	3,816

   Removed data

   failed to process image	no age found	no gender found	no face detected	bad quality (det_score<0.9)	SUM
   0	748	1,170	322	75	2,315 (11.95 %)

   Genders

   female	male
   9,103	7,952

   Ages

   0 to 2	4 to 6	8 to 12	15 to 20	25 to 32	38 to 43	48 to 53	60 to 100
   1,363	2,087	2,226	1,761	5,162	2,719	907	830

2. IMDB age and gender dataset

   This dataset does not have train / val / test splits. Researchers normally use this dataset for pretraining.

   images before removal
   460,723

   Removed data

   failed to process image	no age found	no gender found	no face detected	more than one face	bad quality (det_score<0.9)	no embeddings	SUM
   22,200	690	8,453	21,441	47,278	3855	27	103,944 (22.56 %)

   Genders

   female	male
   153,316	203,463

   Ages

   Ages are fine-grained integers from 0 to 100. Check ./data/imdb_crop/meta-data.json for the details.

3. WIKI age and gender dataset

   This dataset does not have train / val / test splits. Researchers normally use this dataset for pretraining.

   images before removal
   62,328

   Removed data

   failed to process image	no age found	no gender found	no face detected	more than one face	bad quality (det_score<0.9)	no embeddings	SUM
   10,909	1,781	2,485	3,074	2,179	428	0	20,856 (33.46 %)

   Genders

   female	male
   9,912	31,560

   Ages

   Ages are fine-grained integers from 0 to 100. Check ./data/wiki_crop/meta-data.json for the details.

Training

Click to expand!

Model

The model is basically an MLP. There are two variants considered. One is a plain MLP and the other is MLP with IC layers. It's emperically shown that the latter is better than the plain MLP.

Training steps

There are three training steps involved.

1. Hyperparameter search using Ray Tune

   This searches dropout rate, number of residuals per block, number of blocks in the network, batch size, peak learning rate, weight decay rate, and gamma of exponential learning rate decay. Configure the values in hp-tuning.json and run python hp-tuning.py.

2. Pre-training on the IMDB and WIKI dataset.

   We'll use the optimal hyperparameters found in the step 1 to pre-train the model. Configure the values in train.json and run python train.py.

3. Five random seeds on 5-fold cross-validation on the Adience dataset.

   Since the reported metrics (i.e., accuracy) is 5-fold cross-validation, we will do the same here. In order to get the least biased numbers, we run this five times each with a different seed. This means that we are training in total of 25 times and report the average of the 25 numbers. Configure the values in cross-val.json and run python cross-val.py.

Evaluation results

Click to see the detailed results.

Qualitative analysis

Check ./test-images to see the model inference results on some stock images.

Deployment

Click to expand!

We provide the gender and the age models, which are trained on IMDB, WIKI, and Adience datasets. The gender model is a binary classification and the age model is a 101-class (from 0 to 100 years old) classification. They are MLPs with dropout, batch norm, and residual connections. They can be found at ./models/gender.pth and ./models/age.pth, respectively. Both are light-weight. Running on a CPU is enough.

app.py is a flask server app that receives accepts 512-dimensional arcface embeddings and returns estimated genders and ages. You can also run this on a docker container.

Check out this demo video.

Run it as a docker container (recommended).

• Pull and run on CPU

  1. Pull the image from docker hub and run the container.

     docker run -it --rm -p 10003:10003 tae898/age-gender

  2. Build it (optional)

     For whatever reason if you want to build it from scratch,

     docker build -t age-gender .
     docker run -it --rm -p 10003:10003 age-gender

• Pull and run on GPU

  1. Pull the image from docker hub and run the container.

     docker run -it --rm -p 10003:10003 --gpus all tae898/age-gender-cuda

  2. Build it (optional)

     If you want to build this container from scratch for whatever reason, you can do so.

     Make sure your current directory is the root directory of this repo.

     docker build -f Dockerfile-cuda -t age-gender-cuda .
     docker run -it --rm -p 10003:10003 --gpus all age-gender-cuda

Run directly.

1. Install the required python packages.

   pip install -r requirements.txt

2. Run app.py

   python app.py

Running a client

First install the requirements by running pip install requirements-client.txt, and then run the two containers:

1. docker run -it --rm -p 10002:10002 tae898/face-detection-recognition for CPU or docker run --gpus all -it --rm -p 10002:10002 tae898/face-detection-recognition-cuda for cuda.
2. docker run -it --rm -p 10003:10003 tae898/age-gender for CPU or docker run -it --rm -p 10003:10003 --gpus all tae898/age-gender-cuda for cuda.

Now that the two containers are running, you can run client.py. There are two options to run the client.

usage: client.py [-h] [--url-face URL_FACE] [--url-age-gender URL_AGE_GENDER]
                 [--image-path IMAGE_PATH] [--camera-id CAMERA_ID]
                 [--mode MODE]

1. If you have an image stored in disk and want to run the models on this image, then do something like:

   python client.py --mode image --image-path test-images/gettyimages-1067881118-2048x2048.jpg

2. If you want to run the models on your webcam video, then do something like:

   python client.py --mode webcam

Troubleshooting

The best way to find and solve your problems is to see in the github issue tab. If you can't find what you want, feel free to raise an issue. We are pretty responsive.

Contributing

Contributions are what make the open source community such an amazing place to be learn, inspire, and create. Any contributions you make are greatly appreciated.

1. Fork the Project
2. Create your Feature Branch (git checkout -b feature/AmazingFeature)
3. Run make style && make quality in the root repo directory, to ensure code quality.
4. Commit your Changes (git commit -m 'Add some AmazingFeature')
5. Push to the Branch (git push origin feature/AmazingFeature)
6. Open a Pull Request

Cite our paper

@misc{kim2021generalizing,
      title={Generalizing MLPs With Dropouts, Batch Normalization, and Skip Connections},
      author={Taewoon Kim},
      year={2021},
      eprint={2108.08186},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Authors

• Taewoon Kim