Facial Expression Recognition
Transfer learning and a network built from first principles on one task.
Built facial expression classifiers using Xception transfer learning and a convolutional network implemented from scratch. Training experiments used custom callbacks and compared the two approaches on the same data.
The project was a way to understand what transfer learning provides and what is hidden when a model is used only as a ready made component.
Two approaches
The Xception path used existing visual features and fine tuning. The second path built the convolutional network directly, making the layers, training behavior, and limitations easier to inspect.
The comparison
Custom callbacks and a repeatable experiment setup made it possible to compare the approaches instead of relying on one successful training run.
selected tools
- Python
- TensorFlow
- Keras
- Xception
- CNNs