Gradient Estimation Methods of Approximate Multipliers for High-Accuracy Retraining of Deep Learning Models