Use pytorch - gpu training up a larger neural network (including convolution layer, bn, such as the layer model size 31 m), ready to put the model parameters are derived to embedded devices development, which bosses have a good way to consult