Got different result at the second time to train.

[kuangzy2011]

The first time to train, sparse_categorical_accuracy is about 0.9xxxx, but the second time or continous serval times will be
sparse_categorical_accuracy: 0.0894

Is there something wrong with the model?