Age esteemation & Face Detection

1 Description

1.1 Aim

The aim of the project is to provide service runed in Docker with API FastAPI interface which predicts age with custom finetuned neural-network model over the existent fine face extractor.

1.2 Specifications

Date of creation: 7 September 2023

App:

Component	Details
Docs API	Swagger
UI	is out of task

ML models and frameroks

Component	Details
Extraction framework	Deepface
Selected extraction backend	mtcnn
Age model	custom
Age backbone	Facenet
Age dataset	UTKFace_Dataset cropped

2. Instructions

2.1 App start

1.A) Create docker image

docker build -f ./docker/Dockerfile -t local/nvidia_conda:face_detection_age .

1.B) Load custom weights for age module https://disk.yandex.ru/d/oC-5YQYaHAS-ag and place it in ${project_folder}/age_module/weights

2. Run from the project folder with desired GPU amount in interactive mode

cur_folder=$(realpath ./);
sudo docker run --gpus all --rm -it  \
-v $cur_folder:/home \
-p 8000:2020 \
local/nvidia_conda:face_detection_age

3. Start FastAPI

cd ./home/age_module; uvicorn app.app:app --reload --port 2020 --host 0.0.0.0

Done. The app is ready to operate now.

Swagger UI is available for above settings:

http://127.0.0.1:8000/docs#/

Optional: try on example image

curl -X 'POST' \
  'http://127.0.0.1:8000/api/predict_photo' \
  -H 'accept: application/json' \
  -H 'Content-Type: multipart/form-data' \
  -F 'file=@./age_module/tests/data/116_1_0_20170120134921760.jpg.chip.jpg;type=image/jpeg'

2.2 Reproducing and fine tuning on custom datasets

Age training dataset "UTKFace Cropped" migrated to www.kaggle.com/datasets/abhikjha/utk-face-cropped

3 Further possible improvments

Balance load