The　previous　scheme　used　imitation　learning　by　classification　of　branching　or　pruning,　combined　with　Branch　and　Bound　(B&B)　algorithm　to　solve　physical-layer-only　joint　optimal　subcarrier　and　power　allocation　problem　in　the　device-to-device　communications.　In　this　paper,　we　propose　joint　source　encoding　rate　control　and　machine　learning　assisted　cross-layer　joint　optimal　subcarrier/　power　allocation.　The　proposed　scheme　has　source　encoder　rate　control　and　can　adaptively　adjust　video　rate　to　increase　the　video　quality,　peak　signal　to　noise　ratio　(PSNR).　The　previous　physical-layer-only　scheme　did　not　use　the　content-based　video　rate　adaption.　Furthermore,　the　proposed　scheme　uses　the　objective　function　of　PSNR　directly　and　allocates　the　subcarrier　and　power　considering　the　different　rate-distortion　function　of　users＇　videos.　The　previous　physical-layer-only　scheme　could　only　treat　the　users＇　video　equally.　Under　the　new　minimum　PSNR　constraint　of　the　cellular　user　(CU),　we　derive　a　new　objective　function　that　is　independent　of　the　transmission　power　of　the　CU　to　simplify　the　optimization　problem　formulation.　The　previous　scheme　considered　the　physical-layer-only　objective　function　and　constraints.　Finally,　in　addition　to　imitation　learning,　the　proposed　scheme　adopts　ensemble　learning　with　downsampling　the　majority　set　{prune}　to　alleviate　the　class　imbalance　problem　and　improves　performance.　The　simulation　results　show　that　in　the　scenario　where　the　number　of　CUs　is　5,　the　number　of　subcarriers　is　equal　to　the　number　of　CUs,　the　bandwidth　is　15k　Hz,　and　the　number　of　D2D　pairs　is　2,　the　PSNR　of　the　previous　physical-layer-only　scheme　is　31.03　dB,　while　our　proposed　cross-layer　allocation　scheme　is　35.67　dB,　a　4.64　dB　gain.　The　trained　model　trained　at　5　CUs　can　generalize　without　re-training　to　10　CUs　with　only　5.91%　gap　to　the　optimal　PSNR　and　20.43　times　speed　(95%　execution　time　reduction)　when　compared　to　the　globally　joint　optimal　subcarrier/power　allocation　B&B　algorithm.