Authors: ShaoAn Lu

Description: An ensemble approach using two ResNet50 (imagenet pre-trained) networks in which one is case sensitive and another is not. The predictions are first generated by the non-sensitive model and refined by the case-sensitive model. We first trained our model using cross entropy with class weights and heuristically switched to label smoothing cross entropy at some point. Data augmentations are also applied: random cropping/rotating/lightening/zooming, random affine transformation, and random cutout. Other tricks including discriminative layer training and progressively input resolution increasing are also introduced.

Name: Shao-An Lu (呂紹安)
Organization: awesome techno