Sleep　is　crucial　for　human　health.　Automatic　sleep　stage　classification　based　on　polysomnogram　(PSG)　is　meaningful　for　the　diagnosis　of　sleep　disorders,　which　has　attracted　extensive　attention　in　recent　years.　Most　existing　methods　could　not　fully　consider　the　different　transitions　of　sleep　stages　and　fit　the　visual　inspection　of　sleep　experts　simultaneously.　To　this　end,　we　propose　a　temporal　multi-scale　hybrid　attention　network,　namely　TMHAN,　to　automatically　achieve　sleep　staging.　The　temporal　multi-scale　mechanism　incorporates　short-term　abrupt　and　long-term　periodic　transitions　of　the　successive　PSG　epochs.　Furthermore,　the　hybrid　attention　mechanism　includes　1-D　local　attention,　2-D　global　attention,　and　2-D　contextual　sparse　multi-head　self-attention　for　three　kinds　of　sequence-level　representations.　The　concatenated　representation　is　subsequently　fed　into　a　softmax　layer　to　train　an　end-to-end　model.　Experimental　results　on　two　benchmark　sleep　datasets　show　that　TMHAN　obtains　the　best　performance　compared　with　several　baselines,　demonstrating　the　effectiveness　of　our　model.　In　general,　our　work　not　only　provides　good　classification　performance,　but　also　fits　the　actual　sleep　staging　processes,　which　makes　contribution　for　the　combination　of　deep　learning　and　sleep　medicine.