Hello, sorry to bother you, there are some problems that always bother me, I mean, after I train my model,I find that I need to sovle variance,so I want to solve the variance through parameter adjustment or regularization, should I only use it on the dev set Retrain? Thanks very much.
After you have optimized the parameters from train dev set, you go back again and train on train set and then check performance again on the dev set. This process is repeated till satisfactory performance is achieved.
Thanks very much,but I think that if I use L2 or drop out to reduce variance ,it is no problem to go back again and train on training set and then check performance again on the dev set as you say.but if I just adjust my parameter, such as lambda in L2.Is it necessary to do that in the training set or I just need to do that on dev set ?Thanks very much
Its necessary to train again so the model fits and then test again on dev set.