DeepFaceR

This is an R wrapper for a Python package deepface (https://github.com/serengil/deepface)

Steps

1. Installation

R

This wrapper relies on an R package "reticulate":

install.package("reticulate")
library("reticulate")

Use use_python() to enable a specific Python version or use_condaenv to enable a specific Python version in Conda. Make sure the following packages are installed in your Python:

os
urllib
numpy
pandas
deepface
cv2
PIL
mtcnn
matplotlib

If not, use py_install() (or conda_install if using Conda environment) in R to install packages. For example,

py_install('cv2')

If neither Conda or Python is installed yet, download Anaconda.org (https://www.anaconda.com/) and run the following code:

python.pkgs <- c('os', 'urllib', 'cv2', 'numpy', 'pandas', 'deepface', 'PIL', 'mtcnn', 'matplotlib')
conda_create('DeepFaceR')
for (pkg in python.pkgs) conda_install('DeepFaceR', pkg)
use_condaenv('DeepFaceR')

This code creates a new Conda environment called DeepFaceR and installs all necessary Python packages.

2. Download DeepFaceR.R

Download DeepFaceR.R to your working directory and call the program in R using

source('DeepFaceR.R')

3. Run

Ethnorace prediction can be performed using the function DeepFaceR. For example,

df = DeepFaceR(imgfile = 'FULL_PATH_TO_YOUR_IMAGE_FILE')

WHAT IT DOES: The function performs the following tasks:

It detects the number of faces in a given image using MTCNN
If more than one face is detected, it returns an error ERROR: More than one face detected and the code stops.
If only one face is detected, it proceeds to the following steps:
- It pre-processes the image by detecting and aligning the face using DeepFace.detectFace.
- It then outputs the pre-processed image as a temporary file.
- It then predicts an individual's ethnorace using DeepFace.analyze.

OUTPUT: The function outputs a dataframe containing probabilities assigned to each of the 6 ethnorace category, name of the image file, and name of the backend detector employed.

PARAMETERS:

imgfile = full path to your image file (e.g., 'images/image1.jpeg')
outfile = full path of the processed image file that are to be saved temporarily (default: saves as 'temp.jpg' to the working directory)
backend_detector = face detector backend choice from 'retinaface', 'mtcnn', 'opencv', 'ssd' or 'dlib' during the pre-processing stage (default: 'opencv'). See https://github.com/serengil/deepface/blob/master/deepface for details.
skip_multiface = if True it skips the detection of multiple faces using MTCNN. Set this to False if you already know that there's only one face in the image and want to speed the process (default: False)

4. Example

library(reticulate)
source('DeepFaceR.R')

df <- plyr::ldply(list.files('images', full.names = T, recursive = T), function(x){
  df = DeepFaceR(imgfile = x, skip_multiface = FALSE)
  return(df)
})

print(df)

index	asian	indian	black	white	middle eastern	latino hispanic	fn	detector
race	6.56082403	5.34949462	62.598781276	3.887645	3.017876	18.58537	images/image2.jpeg	opencv
race	20.92332840	1.12569947	0.433347281	62.203264	4.683015	10.63134	images/image3.jpeg	opencv
race	0.03598446	0.02714235	0.001058319	80.777507	3.225864	15.93245	images/image4.jpeg	opencv

DianaDaInLee / DeepFaceR