In dynamic manipulation, robots can manipulate objects without grasping by utilizing inertia effect. However, the trajectory planning for dynamic manipulation is a difficult issue due to dynamic constraint. Trajectory deformation considering dynamic constraint after original trajectories are generated is necessary for the issue. To realize such deformation methods, we introduce on sequence-to-sequence (seq2seq) models, which can convert a time series to another time series. This paper proposes a trajectory deformation method with seq2seq models deforming trajectories to satisfy dynamic constraint. Users can obtain trajectories for dynamic manipulation by designing outlines of motion and inputting them to the proposed seq2seq model. In addition, this paper proposes a learning curriculum that does not need labeled dataset. Only mathematical representation of constraint and unlabeled trajectories are necessary. We implement a seq2seq model by the proposed method to a robot turning over pancakes and confirm the validity by a simulation and an experiment.