Method and device of selecting training parameters for training a model are disclosed. The method includes: (1) setting a precision requirement for the model, and a first parameter value interval defined by an upper limit and a lower limit; (2) obtaining a first value point and a second value point within the first parameter value interval; (3) obtaining and comparing respective first and second error rates by respectively setting the training parameter at the first and second value points for the model; (4) updating three values out of the upper limit, the lower limit, the first value point and the second value point; (5) repeating steps (3) and (4), until the precision requirement is net by the respective first and second value points; and (6) obtaining the optimal value of the training parameter.