class Evaluate(Callback):
  def __init__(self,model, type):
      self.lowest = 1e10
      self.model = model
      self.type = type

  def on_epoch_end(self, epoch, logs=None):
    idxs = np.random.choice(range(len(newsdata)), 3)
    for idx in idxs:
      if self.type == 'coupletpoem':
        cptype = newsdata['input'].iloc[idx].split('&&')[1]
        print('格式：', cptype if cptype !='上联' else '对联')
        print('输入：', newsdata['input'].iloc[idx].split('&&')[0])
        try:
          if cptype == '上联':
            res = couplet_gen_sent(self.model, newsdata['input'].iloc[idx][:maxlen])
          elif cptype in ['五言绝句','七言绝句','五言律诗','七言律诗','五言绝句_藏头诗','七言绝句_藏头诗','五言律诗_藏头诗','七言律诗_藏头诗']:
            res = poem_gen_sent(self.model, newsdata['input'].iloc[idx][:maxlen])
          else:
            res = ci_gen_sent(self.model, newsdata['input'].iloc[idx][:maxlen])
          print('输出：', newsdata['output'].iloc[idx])
          print('预测：', res)
        except Exception as e:
          traceback.print_exc()
          print(e)

poemyun = {
  'a': '啊腌扒叭巴吧芭岜疤笆粑豝嚓跨叉杈差咖瓜抓胍榻喇哈阿花哗加茄迦痂枷耞珈袈嘉佳家傢葭豭咖夸姱啦妈摩嬷趴凹咱葩杉沙莎痧鲨纱砂他她它哇洼蛙娲虾丫呀鸦哑桠查楂喳呱欻旮笳拉垃吗蚂仨裟砂趿渣揸馇挝阿八捌擦插锸耷哒嗒鎝褡发筏夹嘎刮栝鸹拉邋抹掐袷葜撒杀刹铩煞刷趿塌溻禢踏挖呷瞎鸭压押扎匝咂拶吒哳浃撒答啊茶查搽嵖猹槎楂碴苴垞蛤华哗骅铧麻嘛蟆拿扒杷爬钯耙筢琶娃霞遐瑕暇牙伢芽岈玡蚜崖涯睚衙拔茇菝跋魃察茬檫达鞑沓怛妲笪靼答跶乏伐茷垡砝阀罚嘎滑划猾夹浃郏荚铗蛱恝戛颊旯拉匣狎挟柙侠峡狭硖辖黠杂砸扎札轧闸铡喋剳',
  'o': '波播菠玻嶓搓磋蹉瑳多哆呙锅卜埚涡坡颇摸阿陂莎唆娑梭挲疙睃圪嗦嗍鲅蓑拖莴倭唷涡窝蜗踒阿婀痾哥歌戈呵科蝌柯疴苛珂窠轲颗屙菏棵髁的了么呢车奢赊畲遮仡猞拨鱿趵钵饽剥逴踔戳撮咄剟掇裰郭崞聒蝈豁劐擢捋泊泼钹说缩托脱喔拙捉苛桌倬涿焯作嘬鸽割搁喝磕瞌榼脖嵯痤瘥鹾罗萝啰逻脶猡锣椤箩骡螺谟馍馍摹模麽摩磨嬷蘑磨魔挪娜傩婆鄱繁皤驮佗陀坨驼柁砣鸵酡跎蛇鼍鹅蛾娥莪俄峨哦讹和禾何河荷阇膜婆皤沱渮哪挼孛荸伯驳帛瓝泊柏勃钹铂舶博鹁浡渤搏箔膊踣薄馞欂襮礴夺度铎踱怫佛掇咄裰剟国掴帼漍腘虢馘活嗑橐灼彴茁踔卓斫浊酌浞诼着啄琢椓蠋擢濯镯昨作笮阁葛蛤颌合涸盒膜拙棁捽貉曷盍阖壳德得额鹖则舴晢蜇革格鬲隔嗝槅膈滆塥鎘骼纥劾阂核翮壳咳颏舌则责择咋泽啧帻舴箦赜折哲辄蛰谪摺磔辙翟',
  'ie': '爹阶皆喈嗟街湝乜咩些靴耶倻椰偕掖瘪憋鳖跌颏疙节截疖角圪结接秸揭噘撅捏撇瞥切缺阙贴得怗贴帖楔歇蝎削薛噎曰约瘸斜邪偕谐鞋携爷耶茄伽鲑踅椰别蹩迭垤昳絰瓞谍堞耋揲喋牒叠碟蝶艓蹀孑节讦劫劼杰诘拮洁结桔桀捷偈玦觉决绝倔桷掘崛脚觖厥劂谲獗蕨橛噱爵蹶矍嚼爝攫钁协胁挟絜颉撷勰襭穴学噱',
  'ai': '开哎哀埃挨娭唉欸掰偲钗差揣呆该陔垓荄赅乖揩腮毸鳃筛酾衰摔苔台胎歪灾哉栽甾斋拍摘拆塞猜挨皑癌才财材裁侪柴豺还孩骸徊怀淮槐踝来莱崃徕涞埋霾俳排徘牌簰邰苔抬骀炱鲐白宅翟',
  'ei': '微欸陂杯卑背悲碑鹎衰崔催摧縗吹榱炊堆飞妃非菲啡騑绯馡蜚扉霏鲱归圭龟妫规邽扳闺硅黑嘿傀瘣瑰鲑灰诙虺挥咴恢袆珲豗晖辉翬麾徽亏刲岿勒悝盔窥胚呸绥衃醅虽荽睢濉忒推危威逶偎隈葳椳煨溦巍蝛薇追骓锥椎欸垂陲捶椎槌棰倕锤箽肥淝腓回茴洄蛔奎逵馗隗葵揆骙暌魁戣睽蝰累雷嫘缧擂檑礌镭羸罍虆玫枚眉莓脢梅嵋猸湄邳媒楣煤酶镅霉糜陪培赔没裵蕤绥隋随遂谁颓韦为圩贼违围帏沩桅唯帷惟薇维嵬巍潍闱',
  'ao': '凹熬包薄苞胞孢剥龅煲褒标飚彪骠镖瘭飙操糙藨镳瀌膘杓猋骉摽操抄怊钞超剿潦刀叨忉刁汈蛁雕貂叼碉凋鲷高皋羔槔睾膏篙糕蒿薅嚆交姣骄艽郊茭浇娇胶椒蛟焦蕉教跤僬鲛嶕礁噍鹪轇尻捞撩猫喵孬抛脬泡漂剽慓飘缥螵僄悄磽磽锹劁敲橇缲搔骚挑缫臊捎烧梢稍筲艄蛸叨涛绦掏滔韬弢饕慆佻祧肖枭枵哓骁嘐逍虓鸮消宵绡萧硝销削蠨蛸翛箫潇霄魈歊嚣幺约夭吆妖要喓腰邀遭糟钊招昭嘲啁着朝约剥削豪敖璈遨嗷廒獒熬隞嶅聱翱鳌鏖螯骜薄雹曹槽螬漕嘈晃巢朝嘲潮捡捯号嗥毫壕濠嚎蠔嚼劳崂痨牢捞唠醪聊辽疗僚漻寥嘹獠寮缭嫽燎憭髎毛矛茅牦旄酕锚髦蝥蟊苗描瞄饶挠饶蛲猱呶刨咆狍庖炮袍匏嫖朴瓢薸乔侨荞峤桥硚翘谯荍鞽憔樵瞧荛桡蛲饶娆桡蛲饶娆苕韶勺芍咷梼逃洮桃陶萄梼啕淘綯醄鼗条岹苕调笤齠蜩迢髫岧鲦崤淆洨爻尧肴淆轺峣陶姚窑谣摇徭遥猺瑶飖鳐轺凿着',
  'ou': '抽掐紬瘳丢都兜蔸勾佝沟枸钩缑篝巘鞲齁勼纠鸠究赳阄湫揪啾蝤轇芤抠眍溜熘搂喽哞妞区讴沤瓯欧殴呕鸥丘邱龟秋蚯湫楸鹙鳅鞦收搜嗖锼馊廋溲飕艘偷修脩休咻羞鸺貅馐髹优攸忧悠呦幽麀舟州谄侜周洲粥啁赒輈诪邹郰緅驺诹陬鲰粥儔帱畴筹酬俦踌惆绸稠禂仇愁雠侯喉猴篌瘊糇流留榴骝刘浏瘤琉硫旒鹠遛镏飗瘤鎏娄楼偻蒌喽耧蝼髅牟眸谋蛑缪鍪牛抔掊裒囚仇犰求毬虬泅俅璆酋逑球遒赇裘璆蝤柔揉糅煣蹂鞣头投骰尤犹疣鱿莸铀由邮油游猷繇蝣妯轴',
  'an': '安氨唵桉庵谙鹌鞍盦扳班颁斑攽般搬瘢皤癍参骖餐觇搀幨襜川穿氽撺镩丹担单眈酖耽郸聃禅儋殚箪端帆番蕃幡藩翻干甘杆玕肝柑竿疳尴关观纶官冠矜倌棺瘝鳏顸酣憨鼾欢讙獾驩刊看勘龛堪戡宽髋颟囡番潘攀三叁山芟杉删衫姗珊栅舢扇跚煽潸膻闩拴栓酸坍贪摊滩瘫湍弯剜湾蜿豌糌簪占沾毡旃粘詹谵邅瞻专砖颛钻躜边砭萹笾编煸蝙鯿箯鞭参骖餐掂傎癫滇颠巅戋尖奸歼坚间肩艰监兼菅笺渐犍湔缄蒹煎缣鹣搛熸鞬鳒韉捐涓娟朘圈鹃镌蠲拈蔫片扁偏篇犏翩千仟阡芊扦迁佥钎牵铅悭谦签愆鹐骞搴磏諐褰圈悛棬弮天添黇仙先纤氙忺籼掀铦酰跹锨鲜暹骞轩宣谖萱揎喧瑄煖禤暄儇懁咽恹殷胭烟焉崦阉阏奄淹腌湮鄢嫣燕鸢眢鸳冤渊鹓箢寒残蚕惭单鋋馋谗婵禅孱缠蝉廛僝潺澶镡蟾镵巉躔传船遄椽攒凡矾烦墦蕃攀樊璠燔繁蘩邗汗邯含函琀焓晗涵韩还环桓圜阛寰缳鬟郇荁洹貆澴轘兰岚斓拦栏婪阑蓝谰澜褴篮斕镧峦娈孪挛鸾脔滦栾銮蛮谩蔓馒瞒鞔鳗鬘男南难喃楠爿胖般盘磻磐蹒蟠蚺然燃髯坛昙倓郯谈弹覃谭痰潭檀团漙抟咱丸纨完玩顽刓汍烷奁连镰怜帘莲涟联裢鲢廉濂鐮鬑磏眠绵棉年粘黏鲇便骈胼蹁钤前虔钱钳乾掮潜黔犍权全佺诠荃泉辁拳铨痊惓筌蜷醛鳈鬈颧田佃畋恬钿甜湉填阗闲贤弦咸挦涎娴衔舷痫鹇嫌玄悬旋漩璇延蜒严言芫妍岩炎沿铅研盐阎筵颜檐元园员沅垣湲袁原圆鼋援媛缘猿塬嫄源羱辕橼',
  'en': '奔贲锛玢宾彬傧斌滨缤槟濒豳参抻郴伧琛嗔汶瞋春椿蝽村皴踆惇吨墩礅敦蹲恩分芬吩纷玢菌氛棻雰根跟昏阍惛婚巾斤今紷金津衿矜筋禁襟军均龟君钧皲坤昆崑裈堃焜琨髡鹍锟鲲抡拎闷喷拼姘钦侵亲衾駸嵚囷逡森申伸身呻砷侁诜参绅珅莘娠深糁燊孙荪狲飧吞暾温瘟心芯骎辛忻昕欣炘锌新歆薪馨鑫勋埙熏薰薫獯曛醺窨因阴茵洇裀荫音姻氤殷堙喑闉愔禋晕缊氲煴赟贞针侦浈珍帧胗真桢砧祯蓁斟甄獉溱榛箴臻迍肫窀谆尊遵樽鳟岑涔臣尘辰沉忱陈宸晨谌纯莼唇淳鹑漘醇存蹲坟汾棼焚濆痕贲贲浑珲馄混哏魂邻林临淋琳粼磷潾嶙遴霖辚瞵鳞麟仑伦论抡沦纶轮门扪们民忞旻岷缗您盆湓贫频嫔颦芹芩矜秦琴覃禽勤懃擒噙螓裙群人壬仁任神什屯囤饨豚臀文纹炆闻蚊雯旬郇寻巡询洵荀荨峋恂鲟循吟垠龈狺訚崟银荧淫寅蟫鄞夤嚚霪云匀芸员沄纭昀畇筠耘筼麇',
  'ang': '肮邦帮梆浜仓苍沧鸧舱昌倡菖猖阊娼伥创疮窗胯当珰铛裆筜方坊芳枋邡钫冈岗刚矼肛纲钢缸釭罡堽光咣胱夯荒肓塃慌江将姜豇浆僵螀缰疆康慷糠囊匡劻诓恇筐牤乓雱滂膀枪锵羌戗戕将腔蜣铰镪嚷丧桑伤汤殇商觞墒熵双泷霜孀孀骦鹴汤铴耥嘡镗蹚汪乡芗相香厢湘缃箱襄骧镶秧央湍殃鸯鞅赃脏臧张章獐彰嫜璋樟蟑妆庄桩装卬昂藏长场苌肠尝常偿徜裳嫦床幢防妨坊忍肪鲂房行吭迒杭絎航颃皇黄凰隍喤煌遑徨湟惶粕锽潢璜蝗篁磺蟥簧鳇扛狂诳鵟郞狼郎琅榔桹琅廊嫏樃硠锒稂鎯螂良俍莨凉梁椋量粮粱邙芒忙杧盲氓茫硭铓牻嚢馕娘彷庞逄旁蒡膀磅螃强墙蔷嫱樯蘘灢禳瓤唐堂棠塘搪糖溏瑭樘膛赯螗螳鄌亡王详降绛庠祥翔扬阳羊玚飏炀杨旸佯疡徉洋',
  'eng': '庚更横盛正应乘胜兴行刑莹茔宁凭暝钉称令并伻崩听祊绷嘣冰兵槟屏柽琤称蛏铛赪撑噌瞠灯登噔蹬镫丁仃叮玎盯钉疔酊靪丰风封枫疯峰烽葑锋蜂更庚耕赓鹒羹亨哼精茎惊京经睛泾荆菁旌晶粳兢鲸坑吭硁铿蒙抨怦砰烹嘭乒俜娉青轻氢倾卿圊清蜻鲭扔僧升生声牲笙甥鼪厅汀听翁嗡兴星狚惺腥应英莺婴撄嘤罂缨璎樱鹦媖瑛膺鹰曾增憎缯罾正争筝蒸征怔挣峥狰钲症烝睁铮稳东冲憧充忡翀舂惮艟匆苁囱枞葱骢璁聪熜冬咚鸫工弓公功攻供肱宫恭蚣躬龚觥哼轰哄訇烘薨埛駉扃空倥崆屄箜忪松凇菘嵩恫通嗵瘑凶兄芎匈汹恟胸佣痈拥邕鄘雍墉慵庸镛壅臃鳙中忪忠终钟盅衷螽宗综棕踪鬃层曾嶒成酲丞呈枨诚承城宬乘盛程惩裎塍澄橙冯逢缝恒姮桁珩横衡蘅楞棱伶灵苓蛉囹泠玲令瓴铃鸰凌陵聆菱棂暹舲翎羚绫棱零龄鲮酃氓虻萌蒙盟甍瞢幪濛曚矇朦艨檬名茗明鸣冥铭洺蓂溟暝瞑螟能棕拧咛狞柠凝芃仍朋膨堋澎彭棚蓬硼鹏篷鬅平冯评坪苹凭枰洴帡屏瓶萍勍情晴檠擎黥绳渑疼腾誊螣藤廷亭停庭蜓婷霆行形邢陉型荥盈萤莹营萦楹滢蝇潆贏赢瀛迎虫重崇从丛尝悰琮弘红吰闳宏泓荭虹闳洪翃鸿黉龙栊茏咙泷珑眬胧昽聋笼隆癃窿农侬哝傢浓脓秾邛穷茕穹藭筇琼蛩跫戎茸荣绒容崂蓉溶瑢榕融嵘同彤侗苘峒桐砼垌佟烔鲷峂樟僮铜童潼瞳朣曈艟雄熊喁颙',
  'di': '氐低羝堤提蹂揉沤几讥叽饥玑机乩肌矶鸡奇屐剞笄姬基期赍犄嵇畸跻箕稽齑丽畿羁咪眯妮丕邳裼批纰坯披砒沏妻栖萋期攲梯蹊欺兮西溪希茜郗稀熙牺唏悕晞傒豨僖嘻奚嬉熹樨羲蹊粞犀曦醯巇鼳伊铱医衣依袆咿猗漪噫繄居车苴拘驹俱狙罝疽据琚趄睢裾区岖佉驱祛蛆躯焌趋黢呈圩盱须虚需嘘墟胥湑谞訏迂纡淤逼嘀滴圾芨唧积击缉激唧襀劈噼霹七柒凄欹戚缉喴漆剔踢夕吸汐昔析矽穸息悉蜥晰淅惜翕蜤锡晰熄噏膝螅歙螅窸蟋一壹揖曲屈掬鞠鋦离蛐欻戌厘狸离骊梨犁鹂喱蓠漓缡璃嫠孷藜黎鲡罹篱黧蠡弥迷眯猕谜糜麋靡蘼醾尼泥坭呢妮輗怩倪霓猊鲵麑皮陂疲枇芘狓毗蚍陴埤啤琵脾裨蜱罴貔鼙齐祈圻芪岐荠祁其奇跂祈祇俟耆颀脐旂萁畦跛崎淇颈骑琪琦棋蛴祺錡綦旗蕲鳍麒鬐荑绨提啼鹈騠缇稊题醍蹄仪圮夷痍匜迤饴怡宜荑贻沂诒眙簃姨胰扅蛇移遗颐椸疑嶷彝儿而驴闾榈劬渠蕖瞿蘧氍癯衢蘧鸲徐于予妤玙余馀欤盂臾鱼竽舁俞谀娱萸雩渔隅揄喁畲逾腴渝愉瑜榆虞愚觎舆窬髃荸鼻锹迪笛的荻敌涤微嫡翟镝及伋吉岌汲级极即舍诘亟革芨急疾棘殛蕺集蒺楫辑嵴踖瘠藉籍习席觋袭媳嶍隰檄熄锡局桔菊焗跼橘曲',
  'i': '哧蚩鸱絺眵笞瓻摛嗤痴媸螭魑呲差疵跐骴尸师诗思鸤坻絁狮葹施訾酾司丝私咝鸶仂斯蛳幼飔厮罳澌撕嘶之卮知支芝吱枝肢栀胝祗脂蜘仔吱孜咨姿兹赀资訾泑嗞缁辎赠粢孳滋趑觜锱龇菑鲻吃失虱湿只织池弛驰迟踟坁持匙漦墀篪词茈茨祠瓷辞慈磁雌慈糍鹚时拾十石实识食蚀仁浉执直侄值职填植殖絷跍摭踯',
  'u': '逋蜅初粗都阇嘟夫肤玞鄜孵敷估姑咕沽孤轱畈鸪罛菇菰蛄辜酤呱觚箍乎呼毂糊刳矻枯骷撸噜铺痡殳书倏姝抒纾枢妷殊梳舒摴樗摅毹输疏蔬苏稣酥栈乌圬邬污呜於钨巫朱洙侏诛茱珠株诸铢猪蛛槠潴橥租菹出督忽胡惚唿淴哭窟仆菩不鹄扑噗叔修辞菽淑窣突秃凸葖屋诬刍除厨锄滁蜍橱篨躇蹰雏徂殂凫扶孚罘苻莩俘浮蚨桴符涪蜉鞭郛狐弧縠和壶葫猢瑚糊蝴糊醐湖瓠鹕卢芦庐垆炉泸栌轳胪鸬颅舻鲈模奴孥驽匍莆脯葡蒲酺如茹儒薷嚅濡孺襦蠕图荼徒途涂菟屠酴无毋芜吾吴唔梧蜈鼯醭毒独读渎椟牍犊髑顿弗佛艴茀幅袱刜拂彿氟绋茀怫芾伏洑茯绂祓服菔芙匐福蝠辐幞襆囫斛觳鹘醭璞濮孰赎塾熟秫俗竹术竺逐烛舳瘃躅足卒崒族镞'
}

poemyuns = [poemyun[key] for key in poemyun]
poemyunsall = ''.join(poemyuns)
poemobj = {
  '五言绝句': [10,22],
  '五言律诗': [10,22,34,46],
  '七言绝句': [14,30],
  '七言律诗': [14,30,46,62]
}

poemobjreword = {
  '五言绝句': [],
  '五言律诗': [11,35],
  '七言绝句': [],
  '七言律诗': [15,47]
}

def poem_gen_sent(self, model, tokenizer, s, topk=2):
    """beam search解码
    每次只保留topk个最优候选结果；如果topk=1，那么就是贪心搜索
    """
    #print('topk: ',topk)
    if '&&' not in s:
      s = s + '&&七言律诗'
      print(s)
    token_ids, segment_ids = tokenizer.encode(s[:maxlen])
    target_ids = [[] for _ in range(topk)]  # 候选答案id
    target_scores = [0] * topk  # 候选答案分数

    poemtype = s.split('&&')[1].split('_')[0]
    #print(poemtype,poemobj[poemtype])

    comma, _ = tokenizer.encode('，。？！、；：')
    comma = comma[1:-1]
    yunlist = {}
    for i in range(maxlen):  # 强制要求输出不超过maxlen字
        _target_ids = [token_ids + t for t in target_ids]
        _segment_ids = [segment_ids + [1] * len(t) for t in target_ids]
        _probas = model.predict([_target_ids, _segment_ids
                                 ])[:, -1, 3:]  # 直接忽略[PAD], [UNK], [CLS]
        _log_probas = np.log(_probas + 1e-6)  # 取对数，方便计算
        _arg = _log_probas.argsort(axis=1)[:, -32:]

        for j in range(topk):
          if i == poemobj[poemtype][1]:
            yunword = tokenizer.decode([target_ids[j][poemobj[poemtype][0]]])
            for item in poemyuns:
              if yunword in item:
                yunlist[j] = item
                break
          for k in _arg[j]:
            # 之前已经出现过引入惩罚
            if k + 3 not in comma:
              if len(poemobjreword[poemtype]) == 2 and i > poemobjreword[poemtype][0] and i < poemobjreword[poemtype][1] and k + 3 in target_ids[j][:-1]:
              # 律诗第二联和第三联可以出现连着两个字
              #print('重复引入惩罚:', k+3, tokenizer.decode([k+3]), _log_probas[j][k], _log_probas[j][k] * 100)
                _log_probas[j][k] = _log_probas[j][k] * 10000
              elif k + 3 in target_ids[j]:
                _log_probas[j][k] = _log_probas[j][k] * 10000
            if i == poemobj[poemtype][0] and tokenizer.decode([k+3]) in poemyunsall:
              # 第一个押韵的字在平水韵表里面奖励
              _log_probas[j][k] = _log_probas[j][k] / 100
            elif i in poemobj[poemtype][1:] and j in yunlist and tokenizer.decode([k+3]) in yunlist[j]:
              # 后面押韵的字要跟第一个押韵的字在同一个韵里面
              _log_probas[j][k] = _log_probas[j][k] / 100

            # 第一个字和标题第一个字重复惩罚
            if i == 0 and tokenizer.decode([k+3]) == s[0] and '藏头诗' not in s:
              _log_probas[j][k] = _log_probas[j][k] * 10000
            elif tokenizer.decode([k+3]) in s and '藏头诗' not in s:
              _log_probas[j][k] = _log_probas[j][k] * 2


        _topk_arg = _log_probas.argsort(axis=1)[:, -topk:]  # 每一项选出topk
        _candidate_ids, _candidate_scores = [], []
        for j, (ids, sco) in enumerate(zip(target_ids, target_scores)):
            # 预测第一个字的时候，输入的topk事实上都是同一个，
            # 所以只需要看第一个，不需要遍历后面的。
            if i == 0 and j > 0:
                continue
            for k in _topk_arg[j]:
                _candidate_ids.append(ids + [k + 3])
                _candidate_scores.append(sco + _log_probas[j][k])
        _topk_arg = np.argsort(_candidate_scores)[-topk:]  # 从中选出新的topk
        target_ids = [_candidate_ids[k] for k in _topk_arg]
        target_scores = [_candidate_scores[k] for k in _topk_arg]
        best_one = np.argmax(target_scores)
        if target_ids[best_one][-1] == 3:
          return tokenizer.decode(target_ids[best_one])
    # 如果max_output_len字都找不到结束符，直接返回
    return tokenizer.decode(target_ids[np.argmax(target_scores)])

def couplet_gen_sent(self, model, tokenizer, s, topk=2):
    """beam search解码
    每次只保留topk个最优候选结果；如果topk=1，那么就是贪心搜索
    """
    s = s.replace(' ','，').replace('。','')
    token_ids, segment_ids = tokenizer.encode(s)
    target_ids = [[] for _ in range(topk)]  # 候选答案id
    target_scores = [0] * topk  # 候选答案分数

    comma, _ = tokenizer.encode('，。？！、；：')
    comma = comma[1:-1]
    for i in range(len(s.split('&&')[0])):  # 强制要求输出不超过title_maxlen字
        _target_ids = [token_ids + t for t in target_ids]
        _segment_ids = [segment_ids + [1] * len(t) for t in target_ids]
        _probas = model.predict([_target_ids, _segment_ids
                                 ])[:, -1, 4:]  # 直接忽略[PAD], [UNK], [CLS]
        _log_probas = np.log(_probas + 1e-6)  # 取对数，方便计算
        _arg = _log_probas.argsort(axis=1)[:, -64:]

        for j in range(topk):
          if token_ids[i+1] in token_ids[1:i+1]:
            # 第i个值的输入之前出现过，奖励
            idxs = [ii for ii,x in enumerate(token_ids[1:i+1]) if x==token_ids[i+1]]
            target_idxs = [target_ids[j][ii] for ii in idxs]
          else:
            idxs = []
            target_idxs = []

          for k in _arg[j]:
            if k + 4 not in comma and k + 4 in token_ids:
              # 当前不为标点，如果在输入里面引入惩罚
              _log_probas[j][k] = (_log_probas[j][k]-1) * 1000
            elif k + 4 in comma and token_ids[i+1] not in comma:
              # 当前为标点，但是输入不是，引入惩罚
              _log_probas[j][k] = (_log_probas[j][k]-1) * 1000
            elif k + 4 in comma and token_ids[i+1] in comma and k+4 == token_ids[i+1]:
              # 当前为标点，预测也为标点，奖励
              _log_probas[j][k] = _log_probas[j][k] / 10
            elif k + 4 in target_idxs:
              _log_probas[j][k] = _log_probas[j][k] / 1000
            elif k + 4 in target_ids[j]:
              # 如果第i个值的输入之前没有出现过，惩罚
              _log_probas[j][k] = (_log_probas[j][k]-1) * 1000

        _topk_arg = _log_probas.argsort(axis=1)[:, -topk:]  # 每一项选出topk
        _candidate_ids, _candidate_scores = [], []
        for j, (ids, sco) in enumerate(zip(target_ids, target_scores)):
            # 预测第一个字的时候，输入的topk事实上都是同一个，
            # 所以只需要看第一个，不需要遍历后面的。
            if i == 0 and j > 0:
                continue
            for k in _topk_arg[j]:
                _candidate_ids.append(ids + [k + 4])
                _candidate_scores.append(sco + _log_probas[j][k])

        _arg = np.argsort(_candidate_scores)
        _topk_arg = np.argsort(_candidate_scores)[-topk:]  # 从中选出新的topk
        target_ids = [_candidate_ids[k] for k in _topk_arg]
        target_scores = [_candidate_scores[k] for k in _topk_arg]
    return tokenizer.decode(target_ids[np.argmax(target_scores)])

class BertUtil:
  def __init__(self):
    self.modelobj = {}
    self.tokenizerobj = {}
    print('bert util init succeed!')
    
  def init_seq2seq_token(self, premodel, type, level='seq2seq'):
    token_dict = {}
    keep_words = []
    seq2seq_config = seq2seq_config_base %(premodel, level, type)
    config_path, checkpoint_path, dict_path, do_lower_case = getconfig(premodel)

    _token_dict = load_vocab(dict_path)  # 读取词典
    _tokenizer = Tokenizer(_token_dict, do_lower_case=do_lower_case)  # 建立临时分词器

    if type in ['cmrc','infoexts2s','kbqa','mathp']:
      newdict = False
    else:
      newdict = True

    if newdict == True:
      if os.path.exists(seq2seq_config):
        tokens = json.load(open(seq2seq_config))
        print('load config %s succeed!' %(seq2seq_config))
      else:
        exit('invalid config %s' %(seq2seq_config))


      for t in ['[PAD]', '[UNK]', '[CLS]', '[SEP]']:
        token_dict[t] = len(token_dict)
        keep_words.append(_token_dict[t])

      for t in tokens:
        if t in _token_dict and t not in token_dict:
          token_dict[t] = len(token_dict)
          keep_words.append(_token_dict[t])
    else:
      for t in ['[PAD]', '[UNK]', '[CLS]', '[SEP]']:
        token_dict[t] = len(token_dict)
        keep_words.append(_token_dict[t])

      for t, _ in sorted(_token_dict.items(), key=lambda s: s[1]):
        if t not in token_dict:
          if len(t) == 3 and (Tokenizer._is_cjk_character(t[-1])
                              or Tokenizer._is_punctuation(t[-1])):
            continue
          token_dict[t] = len(token_dict)
          keep_words.append(_token_dict[t])

    tokenizer = Tokenizer(token_dict, do_lower_case=do_lower_case)  # 建立分词器
    return tokenizer, keep_words
    
  def init_seq2seq(self, premodel='roberta', type='coupletpoem'):
    level = 'seq2seq'
    modeltype = modeltype_base %(level, type)
    self.tokenizerobj[modeltype], keep_words = self.init_seq2seq_token(premodel, type)
    modelfile = modelfile_h5_base %(premodel, level, type)
    if os.path.exists(modelfile):
      #self.modelobj[modeltype].load_weights(modelfile)
      self.modelobj[modeltype] = load_model(modelfile)
      print('load model %s succeed!' %(modelfile))
    else:
      exit("invalid model %s!" %(modelfile))
      
  def predict_seq2seq(self, type='medicalqa', text='', topk=1, printflag=True):
    modeltype = modeltype_base %('seq2seq', type)
    if printflag:
      print(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()), type, 'topk:', topk)
    if type == 'coupletpoem':
      text = text.replace('##','&&')
      if '&&' not in text:
        text = text + '&&七言律诗'
      cptype = text.split('&&')[1]
      if cptype == '上联':
        text = text.replace(' ','，').replace('。','')
        if topk <= 3:
          result = self.couplet_gen_sent(self.modelobj[modeltype], self.tokenizerobj[modeltype], text, topk=topk)
        else:
          result = self.couplet_gen_sent_random(self.modelobj[modeltype], self.tokenizerobj[modeltype], text, topk=topk)
        if printflag:
          print('上联：%s' %(text.split('&&')[0]))
          print('下联：%s' %(result))
      elif cptype in ['五言绝句','七言绝句','五言律诗','七言律诗','五言绝句_藏头诗','七言绝句_藏头诗','五言律诗_藏头诗','七言律诗_藏头诗']:
        if topk <= 3:
          result = self.poem_gen_sent(self.modelobj[modeltype], self.tokenizerobj[modeltype], text, topk=topk)
        else:
          result = self.poem_gen_sent_random(self.modelobj[modeltype], self.tokenizerobj[modeltype], text, topk=topk)
        result = result.replace('。','。\n')
        if printflag:
          print(text.split('&&')[0], text.split('&&')[1])
          print(result)
      else:
        if topk <= 3:
          result = self.ci_gen_sent(self.modelobj[modeltype], self.tokenizerobj[modeltype], text, topk=topk)
        else:
          result = self.ci_gen_sent_random(self.modelobj[modeltype], self.tokenizerobj[modeltype], text, topk=topk)
        result = result.replace('。','。\n')
        if printflag:
          print(text.split('&&')[0], text.split('&&')[1])
          print(result)
    else:
      result = self.gen_sent(self.modelobj[modeltype], self.tokenizerobj[modeltype], text, topk=topk)
    return result

#!/usr/bin/env python3
#-*- coding:utf-8 -*-

import os,sys
import json
import datetime
import time
#from io import BytesIO
from flask import Flask, request
#from gevent.wsgi import WSGIServer

#from gevent.pywsgi import WSGIServer
#from multiprocessing import cpu_count, Process
#from impala.dbapi import connect
import tensorflow as tf
from keras.backend.tensorflow_backend import set_session

os.environ['CUDA_VISIBLE_DEVICES']='0' # 指定GPU 0
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.08 
config.gpu_options.allow_growth = True
set_session(tf.Session(config=config))

print('server start...')


from bertutil import BertUtil

#flask app
app = Flask(__name__)

# BERTUTIL
bertutil = BertUtil()
bertutil.init_seq2seq(premodel='roberta',type='coupletpoem')
print('-----------BERT SEQ2SEQ初始化完毕------------')

# seq2seq请求
@app.route("/zqcloudapi/v1.0/nlp/bertseq2seq", methods=['GET','POST'])
def bertseq2seq():
  start =time.time()
  data = request.values['data'] if 'data' in request.values else '虎啸青山抒壮志&&上联'
  type = request.values['type'] if 'type' in request.values else 'coupletpoem'
  topk = int(request.values['topk']) if 'topk' in request.values else 1
  r = bertutil.predict_seq2seq(type=type, text=data, topk=topk)
  end = time.time()
  res={}
  res['result'] = r
  res['timeused'] = int(1000 * (end - start))
  print('使用时间：%sms\n---------' %(res['timeused']))
  return json.dumps(res)

# 跨域支持
def after_request(resp):
    resp.headers['Access-Control-Allow-Origin'] = '*'
    return resp

@app.after_request
def cors(environ):
  environ.headers['Access-Control-Allow-Origin']='*'
  environ.headers['Access-Control-Allow-Method']='*'
  environ.headers['Access-Control-Allow-Headers']='x-requested-with,content-type'
  return environ

def start_web_server(host='0.0.0.0', port=11456):
  print('listen %s:%s' %(host, port))
  app.run(host, port)

if __name__ == "__main__":
  start_web_server()

netstat -nlp|grep 11456

fetch(url, {
    headers: {
      'Accept': 'text/plain',
      'Content-Type': 'application/x-www-form-urlencoded; charset=UTF-8'
    },
    method: 'post',
    body: param
  })
    .then(res => {
      return res.text();
    })
    .then(res => {
      callback(res);
    })
    .catch(err => {
      errcallback(err);
    });

autossh -M 5686 -fCNR *:8888:localhost:11456 user@xxx.xxx.xxx.xxx

if topk <= 3:
  result = self.poem_gen_sent(text, topk=topk)
else:
  result = self.poem_gen_sent_random(text, topk=topk)

gunicorn -w $1 -b 0.0.0.0:11456 server_poem:app

sh server_poem.sh 9

upstream poemapi.com {
    server 127.0.0.1:8888 weight=1;
    server 127.0.0.1:8889 weight=4;
    server 127.0.0.1:8890 weight=2;
}

server {
    listen 12575;
    server_name _;
    location / {
        proxy_pass http://poemapi.com/zqcloudapi/;
        proxy_set_header Host $host;
        proxy_set_header X-Real-IP $remote_addr;
        proxy_set_header REMOTE-HOST $remote_addr;
        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
    }
}

export const getServerData = (url, method, object, period, callback, errcallback) => {
  url = Config.host + url;
  const key = md5(url + method + JSON.stringify(object));
  // 判断本地localstorage数据是否有效
  const data = getDataFromLocalStorage(key, period);
  if (!data) {
    getData(url, method, object, text => {
      const resultobj = JSON.parse(text);
      // console.log(resultobj);
      if (resultobj.errcode === 0) {
        if (period > 0) {
          resultobj.result.time = new Date().getTime(); // 添加当前时间
          localStorageUtil.save(key, resultobj.result);
        }
        callback(resultobj.result);
      } else {
        // console.log('业务逻辑异常:', resultobj.errmsg);
        errcallback('业务逻辑异常: ' + resultobj.errmsg);
      }
    }, err => {
      errcallback(err);
    });
  }
  else {
    callback(data);
  }
};

npm install pinyin

export const checkKeys = (text) => {
  let flag = true;
  // 根据白名单判断
  if (keyobj['whitekeys'].indexOf(text) >= 0) {
    return true;
  }
  // 剔除非法字符
  text = text.replace(/[\s+\-\.\。\，\；\?\？\！\!\…\+\!\@\#\$\%\^\&\*()\（\）\￥\[\]]/g,"");
  if (words.indexOf(text) >=0) {
    return false;
  }

  // 转换成拼音并判断，如果文本中包含核心敏感词的拼音，就返回false
  let py = pinyin(text, {
    style: pinyin.STYLE_NORMAL
  });
  py = py.join('');

  for (let i = 0; i<pinyins.length; i++) {
    if (py.indexOf(pinyins[i])>-1) {
      console.log(text, py, pinyins[i]);
      return false;
    }
  }

  py = pinyin(text);
  py = py.join('');
  console.log(py);

  for (let i = 0; i<pinyins2.length; i++) {
    if (py.indexOf(pinyins2[i])>-1) {
      console.log(text, py, pinyins2[i]);
      return false;
    }
  }

  // 根据敏感词判断，只要输入文本包含敏感词就返回false
  for (let i = 0; i < keyobj['keys'].length; i++) {
    if (text.indexOf(keyobj['keys'][i])>=0) {
      console.log(text, keyobj['keys'][i]);
      flag = false;
      return false;
    }
  }

  // 根据核心敏感词判断，只要输入出现敏感词这几个字，不论什么顺序，中间夹了多少个字，都返回false
  for (let i = 0; i < keyobj['subkeys'].length; i++) {
    let j = 0;
    let item = keyobj['subkeys'][i];
    let newtext = text;
    for (j = 0; j < item.length; j ++) {
      if (newtext.indexOf(item[j]) < 0){
        break;
      }
      newtext = newtext.replace(item[j],'');
    }
    if (j == item.length) {
      return false;
    }
  }
  return flag;
};