I am using ResNext architecture for classification. the training dataset contains approximately 31000 images distributed among 61 classes. And validation dataset contains appx. 4500 images. but unfortunately the training accuracy is not increasing over 13%, can any one show me what am I doing wrong here?

Code is here

Model architecture

Data Loader

Resnext research paper

With this data set 90% accuracy have achieved with inception resnet v2 model, and this in the research paper of resnext, it is claiming it should perform better than inception resnet v2 model. So I am expecting accuracy better than this.