I’m trying to build a Wide ResNet (WRN - 16 - 4) on CIFAR10 from scratch. I encounter over fitting problem with training accuracy at 91% and validation accuracy at 25%. The following are measures I tried but couldn’t get a solution,
- Drop out - 40%
- Data augmentation and normalization
- Xavier Uniform weight initialization
- Batch size
Kindly help me with ideas I can try to solve this issue.