The loss did not converge,can you help me?

[yang0817manman]

     hello,thanks for sharing your code.
     Can you tell me there is wheather goog strategy for trainning large scale dataset. I have trained sphereface-20 with ms-celeb-1m.Then I finetune spherefacenet20_ms-celeb-1m.caffemodel with VGGFace2 datasets. At the begin of training,the loss is 12 ,iterater 20000 times,loss decrease to 9,after that the loss did chage.I try setting lambda_min = 10 or decreate lr=0.001 but it did not converge. Should I set the hyperparameters and is it really possible to converge? Can you help me?

Using the settings in my prototxt:
type:SINGLE
base: 5
gamma: 0
power: 1
lambda_min: 5