Device to device (D2D) communication underlaying LTE can be used to distribute traffic loads of eNBs. However, a conventional D2D link is controlled by an eNB, and it still remains burdens to the eNB. We propose a completely distributed power allocation method for D2D communication underlaying LTE using deep learning. In the proposed scheme, a D2D transmitter can decide the transmit power without any help from other nodes, such as an eNB or another D2D device. Also, the power set, which is delivered from each D2D node independently, can optimize the overall cell throughput. We suggest a distirbuted deep learning architecture in which the devices are trained as a group, but operate independently. The deep learning can optimize total cell throughput while keeping constraints such as interference to eNB. The proposed scheme, which is implemented model using Tensorflow, can provide same throughput with the conventional method even it operates completely on distributed manner.