I tried with the usual integer labeling in the assignment together with sparse_categorical, and I also tried vectorizing the labels as one-hot encoded using utils.to_categorical together with categorical_crossentropy as loss function; in both cases val_accuracy is bigger than accuracy (training) , isn’t that weird? does anyone know why this is happening? It happens from the first epoch and continues till the end!
Please post some numbers for your results. How much bigger?