Transfer learning to the Scale Scalable Graph Network for advanced simulation

In recent years, the models of the graph nial network (GNN) show promising results in imenting complex socio-body programs. However, Dedicated Graph Network Simulator will be very expensive, as most models are locked in full guidance. The broad data generated from traditional simulators are required to train the model. It remains how the traveling learning can be added to improve the functioning of the model and efficiency of training. In this work, we introduce our integrity and transmits reading paradigm with the Graph Network Simulator. First, we proposed a graph EU-Net (Sguntt). By putting the depth of the first deeper depth (DFS) to enable the transfer reading between different SgunNut, we raise a set of map function to sync parameters between the specified model and target model. Additional Term of General It has been added to the loss to enhance similarities and weights for the best applicable model. Then we build a dataset to be simulators. Includes 20,000 physical measurements with 3D situations chosen randomly in an open dad (ABC) source. We show that in our proposed learning methods, a well-organized model that has a small part of the training information can achieve the best performance compared to the trained from scratch. On the 2D Plate of Visible, Our Improved Model in 1/19 training data can reach 11.05% development compared to the trained model from scratch.