Bayesian Dropout Layer in FairFace Network
Analyzed the addition of a Bayesian Dropout Layer to a pre-trained ResNet34 model for face recognition (FairFace). Converted the model weights and biases from a PyTorch neural network to Keras to leverage transfer learning. Found that Bayesian dropout resulted in lower accuracy vs. both a traditional dropout layer and the original model.
Features
- Pre-trained ResNet34 model for face recognition
- PyTorch to Keras model conversion
- Transfer learning
- Bayesian dropout layer analysis