####1.Background
VGG 16 architecture is used to classify 6 emotions.
-
Paper on VGG architecture: Very Deep Convolutional Networks for Large-Scale Image Recognition
-
Paper on which this project is based on: Recognizing Facial Expressions Using Deep Learning
-
Other references: Convolutional Neural Network
-
Techniques used in training:
-
Class balancing using Resampling of Dataset
Reference: 8 Tactics to Combat Imbalanced Classes in Your Machine Learning Dataset
-
Data augmentation
Reference:Data augmentation
-
pip install -r requirements.txt
- Use train_in_colab.ipynb to train the model in google colab
- Use model_from_weights.py to create model from weights
- Use realtime.py for realtime emotion recognition
- vgg.py is architecture definition using Keras
Training:
- Loss : 0.9238
- Accuracy: 0.6496
Validation:
- Loss: 0.8733
- Accuracy: 0.6800
Test:
- Loss: 0.95643
- Accuracy: 0.6461