This repository is to do facial expression prediction by fine-tuning ResNet-18 with FER-2013 Faces Database.
(Image source: www.kaggle.com)
Pytorch is used to create ResNet18 model. RandomCrop and HorizontalFlip are applied on data before training. The model is trained for 200 epochs with batch size 32.
Private Test Accuracy: 73%
Loss Graph:
Accuracy Graph:
Confusion Matrix:
Keras is used to create and train a CNN model with BatchNorm. Data is augmented using random rotation, width/height shift, zoom, horizontal flip before training. Model is trained for 50 epochs with batch size 64. More accuracy can be achieved by training for more epochs.
**Private Test Accuracy: 65.45% **
Loss & Accuracy Graph:
