The minimal training set to train a working convolutional neural network (CNN) is explored in detail. The model under consideration is the frustrated ,[object Object],- ,[object Object], Ising model on the square lattice. Here ,[object Object], and ,[object Object], are the nearest and next-to-nearest neighboring couplings, respectively. We train the CNN using the configurations of ,[object Object], and employ the resulting CNN to study the phase transition of ,[object Object],. We find that this transfer learning is successful. In particular, only configurations of two temperatures, one below and one above the critical temperature ,[object Object], of ,[object Object],, are needed to accurately determine the ,[object Object], of ,[object Object],. However, it may be subtle to use this strategy for the training. Specifically, for the model under consideration, due to the inefficiency of the single spin-flip algorithm used in sampling the configurations in the low-temperature region, the two temperatures associated with the training set should not be too far away from the ,[object Object], of ,[object Object],. Otherwise, the performance of the obtained CNN is not of high quality, and hence cannot determine the ,[object Object], of ,[object Object], accurately. For the model under consideration, we also uncover the condition for training a successful CNN when only configurations of two temperatures are considered as the training set.