3.1.2. Pre-Training

This work uses publicly available pre-trained model of the base model (i.e., transfer learning-based model) instead of directly pre-training it. Specifically, the SE-ResNet-50 pre-trained model, which is published on the official website of the VGGFace2 database [55], was trained with the VGGFace2 dataset [22] for the task of recognizing 8631 faces. As a results, the network learned rich feature represent for face factors.
