In general, as the amount of training data is increased, a deep learning model gains a higher training accuracy. To assign labels to training data for use in supervised learning, human resources are required, which incur temporal and economic costs. Therefore, if a sufficient amount of training data cannot be constructed owing to existing cost constraints, it becomes necessary to select the training data that can maximize the accuracy of the deep learning model with only a limited amount of training data. However, although conventional studies on such training data selections take into consideration the training data labeling cost, the selection cost required in the training data selection process is not taken into consideration, which is a problem. Therefore, with the consideration of the selection cost constraint in addition to the data labeling cost constraint, we introduce a training data selection problem and propose novel algorithms to solve it. The advantage of the proposed algorithms is that they can be applied to any network model or data model of deep learning. The performance was verified through experiments using various network models and data.