val_angle_error is too high

[Bothgone]

i use your net to train my own dataset,but the val_angle_error is too high around 90%. What can i do to reduce my val_angle_error