Within DNN training process, the test set of data set is very uneven, appear in the test set acc curve drawing over into acc curve starting from 0.99 1, then drop some, about 0.8 to flatten out, loss curve of test set is normal, don't know what circumstance, I think maybe because model caused by unbalanced data sets are in favour of the negative at the beginning, I didn't use training set down sampling of the balance of Yin and Yang data set, it's figure: