Closed liygcheng closed 4 years ago
@liygcheng Could you provide more details? For example, tell me what kinds of datasets and networks you used.
dataset: CUB_200_2011 arch: ResNet50 loss: label smooth cross entropy
@liygcheng We also have tried to train models with CutMix on CUB-200 dataset, but we observed that CutMix improved the performance well. Maybe you should tune other hyper-parameters, such as learning rate, etc., to get optimal performance.
When I employ CutMix, the results get worse instead. What is important to keep in mind when I use CutMix?