How to train the deep learning model efficiently?

It can be adjusted in many directions: 1 The loss function is changed to the cross entropy loss function, because the later training error is small, and the gradient becomes smaller and smaller when the mean square error is used as the loss function, so the training process will become smoother and smoother. 2. Using random gradient descent method to prevent local convergence; 3. Reduce the learning rate.