In object manipulation involving contact, the object may be hidden from the camera, and haptics are useful to compensate for this information. In this study, we propose a deep learning-based method for generating robot motions using tactile data. We introduced attention mechanism for image feature extraction, softmax transformation for motion generation, and convolutional neural network for processing tactile sensor data. We tested the effectiveness of the proposed method on the unzip task of an flexible bag. We confirmed that the proposed method can realize the motion generation according to the deformation of the zipper while reducing the load on the zipper, and achieved a success rate of 90 percent.