Not stable training in CNN

Question

I am currently doing a image classfication which takes up to 81000 data to train the model(CNN). It has a very not stable val_training accuracy and val_accuracy between each epo