eval reach a good performance but saved model cannot predict as good as eval

[SUMMER1234]

[SUMMER1234]

my classification reach performance{'accuracy': 0.8878125, 'loss': 0.33625996, 'precision': 0.8802214, 'recall': 0.9352941, 'global_step': 1000} but using same evaluation data on pb model only reach about 0.6 accuracy