I\'m implementing a 7-class classification task with normalised features and one-hot encoded labels. However, the training and validation accuracies have been extremely poor