Facial Emotion Recognition

This package supports 1) Face extraction/alignment, 2) Training, 3) Inferring, 4) real-time webcam for Facial Emotion Recognition.

0. Dependencies

Install PyTorch.
Install OpenCV.
Install sklearn.
Install PIL, matplotlib, numpy, etc.

This package is checked in Conda + Python3.8 + Cuda 11.1 + PyTorch LTS (with CUDA 11.1)

1. Extract (crop) and align face from an image (MTCNN)

python extract_face.py

DIR_IN: raw image path (before face extraction). The code expects one-level hierarchy under DIR_IN (folders under DIR_IN).
DIR_OUT: face data path (after face extraction) with same one-level hierarchy.
NUM_IMAGES: number of images to extract faces under each folder (DIR_IN); 1) positive: from the first, 2) negative: from the last, 3) 0: all images

2. Train

python train.py

Modify settings via argument parsing.
This code results in f'{args.identifier}_model_best.pth.tar'

3. Infer

python infer.py

Use f'{args.identifier}_model_best.pth.tar' from python train.py as --model.

The user has three options for --image:

1. If --image is None, the code test all images under --data path and save the results under --results
2. If --image is fault_finder, the code makes csv file that includes how/what images were inferred incorrectly.
3. If --image is a specific path to an image, the code prints the probability and the class with the maximum probability on console.

4. Webcam

python webcam.py

model(*.pth.tar) must exist at --model.
Use --gpu, --fps for your convenience.
--resize-h is 0 by default, which does not resize the input from the real-time webcam at all.