AI建模師-素養(yǎng)手冊(7)-把專家直覺納入AI的實(shí)踐篇

目錄1.前言2.范例（一）：工廠機(jī)器的維修判斷3.范例（二）：醫(yī)院護(hù)理師排班4.護(hù)理師排班的實(shí)踐代碼、"__1—月I」言-在上一集(第6集)里，舉了兩個(gè)專家直覺的范例，一個(gè)是工廠里機(jī)器的維修判斷。另一個(gè)范例是醫(yī)院護(hù)理師的排班。-不過，在上一集里，您只看到ExceI畫面的操作，以及所輸入的數(shù)據(jù)和輸出的結(jié)果。-也許您會(huì)很好奇，這Excel背后的Python程序是如何實(shí)現(xiàn)的昵？-那么，本集就來演示一下這幕后的程序碼。其中使用了TensorFIow框架(Framework)。***本文摘自高煥堂的下列書籍******以及北京【電子世界雜志】連載專欄***ChatGPT的啟示-ChatGPT的能力很驚人，但它仍是縱橫江湖的野貓，而非真正貼心的〈家貓〉。-ChatGPT的表現(xiàn)讓人驚艷，但它仍是位創(chuàng)新組合食材的炒飯快手，還需搭配您自己的素材，才真正創(chuàng)新大廚師。-在ChatGPT上想搭配您自己的食材，可行途徑之一是：您自己建立中小格局的AI模型，輸入您的素材，您自已訓(xùn)練該模型，訓(xùn)練出〈潛藏空間向量〉，然后將它（向量），融合進(jìn)去ChatGPT的潛藏空間里。-所以，逐漸地家家戶戶都將需要〈AI建模師〉來建模、訓(xùn)練，然后融合成有咼度智慧的AI家貓。-許多人對于機(jī)器學(xué)習(xí)，常常只關(guān)注于〈訓(xùn)練數(shù)據(jù)＞與＜算法＞，然后就是〈輸出結(jié)果〉；而沒意會(huì)到：數(shù)據(jù)只是〈沙子＞，經(jīng)由模型的淬煉才成為〈金子〉，再經(jīng)由模型鑄造才做出漂亮〈金飾〉。-其中，〈金子〉才是關(guān)鍵性的素材。不同素材的創(chuàng)新組合，并貼心依據(jù)個(gè)人的心意(Attension)而調(diào)整和修飾，才是今天ChatGPT的真諦，才是GreatPoint(GPT)。-金子在哪里，就藏在人們無法理解的桃花源里，就是潛藏空間(LatentSpace)，又稱為：隱藏空間或隱空間。-金子在哪里，就藏在人們無法理解的桃花源里，就是潛藏空間(LatentSpace)，又稱為：隱藏空間或隱空間。I范例：專家直覺-在一個(gè)工廠里，有一部機(jī)器天天運(yùn)作中，它會(huì)處于3種狀態(tài)之―，分別以RGB顏色代表之。如下圖：()范例：專家直覺-每天中午12:00記錄其當(dāng)天狀態(tài)。當(dāng)其狀態(tài)為順時(shí)鐘、或反時(shí)鐘變化，屬于正常變化。如下圖：正常變化）正常變化)詳細(xì)說明：一部機(jī)器會(huì)處于3種狀態(tài)，分別以RGB顏色代表之。每天

中午12:00記錄其當(dāng)天狀態(tài)。當(dāng)其狀態(tài)為順時(shí)鐘、或反時(shí)鐘變化，屬于正常變化；否則為異常變化（跳機(jī)）。如果出現(xiàn)〈連續(xù)異常變化〉就必須停機(jī)檢修。異常變化O范例：專家直覺-否其中值得留意的是，依據(jù)工廠的機(jī)器管理準(zhǔn)則：如果出現(xiàn)〈連續(xù)異常變化（跳機(jī)）〉就必須停機(jī)檢修。-現(xiàn)在，我們就來看看過去一周（工作6天）以來，這部機(jī)器狀態(tài)紀(jì)錄數(shù)據(jù)，如下：I范例：專家直覺-有一位負(fù)責(zé)檢視機(jī)器狀態(tài)的老師傅來了，他一眼就能看出了這部機(jī)器，在過去一周（工作6天）里并沒有出現(xiàn)〈連續(xù)異常變化（跳機(jī)）〉的現(xiàn)象。?所以不必須停機(jī)檢修。-那么AI是否也能瞬間看出來呢？范例：Al來學(xué)習(xí)專家直覺-茲把這些數(shù)據(jù)呈現(xiàn)于Exce1表格里，如下圖：I范例：Al來學(xué)習(xí)專家直覺-運(yùn)用專家直覺，把它表達(dá)于卷積核里：給予的百例：Al來學(xué)習(xí)專家直覺-請按下〈卷積〉，就拿K0[]和K1[]卷積核來對X[]進(jìn)行卷積運(yùn)算，得到Y(jié)0[]:范例：Al來學(xué)習(xí)專家直覺-從上圖里的Y0[]就可以看出來了：有一個(gè)值達(dá)到510，代表發(fā)現(xiàn)一次異常（跳機(jī)）現(xiàn)象，從紅色狀態(tài)跳到藍(lán)色。-同理，從Y1[]可以看出來：有一個(gè)值是達(dá)到510，代表發(fā)現(xiàn)一次異常（跳機(jī)）現(xiàn)象，從藍(lán)色跳到紅色。范例：Al來學(xué)習(xí)專家直覺?接下來，請按下〈相加〉。把K0[]所提取的特征（即Y0）與K1[]所提取的特征（即Y1），合并起來。例如，把Y0[]和Y1[]的對應(yīng)元素進(jìn)行V兩兩相加〉計(jì)算，而得到z[]°?從Z[]可以看出來：在本周里總共跳機(jī)2次。I范例：Al來學(xué)習(xí)專家直覺-人類專家一眼就看得出來：本周沒有發(fā)生〈連續(xù)兩天跳機(jī)＞的現(xiàn)象。?那么，AI是否也能一眼看出來呢？-答案是：可以的。?剛才由兩個(gè)卷積核：K0[]和kl[]去進(jìn)行卷積運(yùn)算（自動(dòng)提取特征），分別看到了一次跳機(jī)現(xiàn)象。但是如何看出來是否V連續(xù)跳機(jī)〉呢？答案是：再進(jìn)行一次特征提?。ň矸e）就可以看出來了。范例：專家直覺?再一次使用卷積核，如下圖：I范例：Al來學(xué)習(xí)專家直覺?接下來，請按下〈卷積〉。就拿KZ[]卷積核來對Z[]進(jìn)行卷積運(yùn)算，得到Y(jié)Z[]，就可以看出來了。范例：Al來學(xué)習(xí)專家直覺?請看看Python程序，來實(shí)踐上述的情境。范例：Al來學(xué)習(xí)專家直覺kO=np.array([1,kl=npfarrayt[0,yO=conv(x,kOf3)piint("\nYD=",np,round(yO,2))yl=Conv(x,k1,3)print("\nYl=",np,round(yl,2))z=np.zeros((yO.size),dtype='int32()far1inrange(yO.Size):z[i]=y0[i]+yl[i]printC'VnZ=”，z)kz=np.array([1,1])yz=conv(z,kz,1)print("\nYZ=",yz)#............ContInued--------------------x=np.array([255,0,0,0,0.255,0,255,。,255,0,0,0,255,01)0,0.255.專家直接給卷積核KO和Kl___________丿-卷積核的W，直接從專家心中來。^1/]]1OQOD-1O1OO_______________范例：Al來學(xué)習(xí)專家直覺?卷積核的w，直接從專家心中來。#............ContInued.............................x=np.array([255,0,0,0,0,255,0,255,丄0,0,255.255,0,0,0,255,0])k0=np.array([1,0,□,0,Q,1])kl=np.array([0?0,I,1,0,0])yO=conv(x,

kOf3)print("\nY0=H,np,round(yO,2))yl=Conv(x,k1,3)print("\nYl=",np,round(yl,2))z

=npnzeros((yO,size),dtype='int32')fciriinrange(yD.SIze):z[i]=y0[i]+yl[i]printfAnZ=”，z)kz=np.array([1,1])yz=conv(z,

kz,1)print("\nYZ=\ys)#End____________________拿KO去卷積____________________范例：Al來學(xué)習(xí)專家直覺kO=np.array([1,kl=npfarrayt[0,yO=conv(x,kOf3)piint("\nYD=",np,round(yO,2))#............ContInued--------------------x=np.array([255,0,0,0,0.255,0,255,。,255,0,0,0,255,01)0,0,255,print("\nYZ=",yz)kz=np.array([1,1])yz=conv(z,kz,I)z=np.zeros((yO.size),dtype='int32')far1inrange(yO.Size):z[i]=y0[i]+yl[i]printC'VnZ=”，z)y1=Conv(x,k1,3)print("\nYlnp,round(yL2))拿Kl去卷積-卷積核的W，直接從專家心中來。]]loQoD-1o1oo_______________范例：Al來學(xué)習(xí)專家直覺kO=np.array([1,kl=np.array([0,#............ContInued--------------------..............................x=np.array([255,0,0,0,0.255,0,255,。,255,0,0,0,255,01)____________________________________yO=conv(x,k。，____________prmt(F,\nYlJ=",np,round(yU,27J_________________________________________Iyl=Conv(x,kl,3)0,0,255,primei=*,np,round(yl,ZTJprint("\nYZ=\yz)kz=np.array([1,1])yz=conv(z,kz,I)z=npnzeros((yO,size),dtype='int32')fciriinrange(yD.SIze):z[i]=y0[i]+yl[i]printC'VnZ=”，z)___yO特征表[FeatureMap)____________________________汰_丿______________yl特征表(FeatureMap)____________丿-卷積核的w，直接從專家心中來。^1/]]1OQOD-1O1OOkO=np.array([1,kl=npfarrayt[0,yl=Conv(x,k1,3)print("\nYl=",np,round(yl,2))yO=conv(x,kOf3)piint("\nYD=",np,round(yO,2))范例：Al來學(xué)習(xí)專家直覺?卷積核的w，直接從專家心中來。Continuedx=np.array([255,0,0,0,0,255,0,255,丄0,0,255,255,0,0,0,255,01)z=npnzeros((yO,size),dtype='int32,)fciriinrange(yD.SIze):z[i]=y0[i]+yl[i]printfAnZ=”，z)將兩個(gè)特征表相加______________丿kz=np.array([1,1])yz=conv(z,kz,1)print("\nYZ=",yz)]]loQoD-1o1oo_______________范例：Al來學(xué)習(xí)專家直覺kO=np.array([1,kl=npfarrayt[0,kz=np.array([1,1])yz=conv(z,

izrI)______________________________________________yO=conv(x,

kOf3)print("\nY0=H,np,round(yO,2))______________________yl=Conv(x,k1,3)print("\nYl=",np,round(yl,2))z=npnzeros((yO,size),dtype='int32')fciriinrange(yD.SIze):z[i]=y0[i]+yl[i]printfAnZ=”，z)_________________________________________print("\nYZ=\yz)______再一次卷積___________丿-卷積核的W，直接從專家心中來。#............ContInued.............................x=np.array([255,0,0,0,0,255,0,255,丄0,0,255.255,0,0,0,255,01)]]loQoD-1o1ooYO=[51002550255]Y1=[025505100]Z=[510255255510255]YZ=[765510765765]?>兩個(gè)特征表Y0=[51002550255]Y1=[025505100]Z=[510255255510255]YZ=[765510765765]?>丿YO=[51002550255]Y1=[025505100]Z=[510255255510255]YZ=[765510765765]?>特征表相加丿-發(fā)現(xiàn)了2次跳機(jī)YO=[51002550255]Z=Y1=[025505100]匝)255255(510)255]YZ=[765510765765]?>YO=[51002550255]Y1=[025505100]Z=[510255255510255]-YZ□都小于1020，沒有〈連續(xù)跳機(jī)〉的現(xiàn)象Y0=[51002550255]Y1=[025505100]Z=[510255255510255]YZ=「7655107657651?>最后的卷積表專家只提供直覺判斷范例：Al來學(xué)習(xí)專家直覺-設(shè)計(jì)一個(gè)分類器，來吸納專家智慧ABCDEFGHIJKLMN0126RGBRGBTZ3紅2550002550綠0（沒問題）4S6綠0255000255藍(lán)0（沒問題）5藍(lán)0025502550綠0（沒問題）6Epoch500綠0255025500紅0（沒問題）7紅2550000255藍(lán)_1（跳機(jī)）8藍(lán)0025525500紅1（跳機(jī)）910111213141516Initial學(xué)習(xí)EpochInitial學(xué)習(xí)正規(guī)化機(jī)臺(tái)的各種狀態(tài)變化（沒問題）（沒問題）（沒問題）（沒問題）ABCDEFGHIJKLM.N0126RGBRGBTTZ3紅2550002550綠0（沒問題）4S6綠0255000255藍(lán)0（沒問題）5藍(lán)0025502550綠0（沒問題）6Epoch500綠0255025500紅0（沒問題）7紅2550000255藍(lán)_1（跳機(jī)）8藍(lán)0025525500紅1（跳機(jī)）91011Initial12范例：Al來學(xué)習(xí)專家直覺-正規(guī)化ABCDEFGHIJKLMN126RGBRGBT3紅100010綠0（沒問題）4s6綠010001藍(lán)0（沒問題）5藍(lán)001010綠0（沒問題）6Epoch500綠010100紅0（沒問題）7紅100001藍(lán)_1（跳機(jī)）8藍(lán)001100紅1（跳機(jī)）91011Initial120Z17范例：Al來學(xué)習(xí)專家直覺-展幵訓(xùn)練ABCDEHJKLMNO1N6TZ2（沒問題）00.03（沒問題）S600.034（沒問題）00.035（沒問題）Epoch50000.036（跳機(jī)）10.977（跳機(jī)）10.97891011Initial1213W-5.211.741.6-5.341.614正規(guī)化B1516x0xlx2藍(lán)紅藍(lán)紅遷移到卷積核1.740RGBRGB紅100010綠010001藍(lán)001010綠010100紅100001藍(lán)001100第1天第2天第3天第4天第5天第6天00002550255000255255000255W1.74-5.211.741.65341.6第1天第2天第3天第4天第5天第6天255000255第1天第2天第3天第4天第5天第6天255000255ATX[]00002550255001.6QRs第6天EFG第2天12KLM第4天NOP第5天BCD第1天HIJ第3天z255J56J_8Ar'255000--tT"二_——w1.74「5211.741.6P-5.34B0請看看Python程序，來實(shí)踐上述的情境。#ex_All_04.pyimportiiiiiiipy；技】項(xiàng)importkurasfLik?iil：,iiiudE!1siipui'

記qu日ritialfro/iikeras.layersimportDensefromkeras.optionizersimpo11SGDfromkeras.modelsijiiportMod&ldefsigmoid(y):return1/(1+np.6xp(-y))1,dtype=np.f1Bl132)丿準(zhǔn)備分類器的訓(xùn)練數(shù)據(jù)t=np.array([[0],[0],[0],[0],[1],[1]],dtypNnp,f1oat32)...........Continued......................................EditFormatRunOptionsWindowHelp05005]525025>50o505?22?o?>55o02020050055505505?5?72o2?o55ff.f200020np請看看Python程序，來實(shí)踐上述的情境。Continued建立分類器模型d.$at_w日ights(wb)i=

Dense(0,activation=1

si^moid'?

name="dM

,

input_di葉”model=Sequentia1()modelpadd(d)modelnCDinpilet1dss=keras.losses.MSE,optimizer=SGD(1r=0.15),metrics=['accuracy'])wb=[tip,array([[0.5],[-0.5],[0.5],[04],[-04],[04]],dtype=np.fLoat32),np.ariay([0.0],dtyp&=np,float32)]kkw=Nuiiekkb=Mone#...........continued...............................6￡-!______-=--訓(xùn)練分類器wo=d.g日t_weights()[0]kkw=】項(xiàng)，￡qiMPZ：頃WQ)print(n\n-----training-----*')printf=‘，np^roundtwo,2))bo=d.^et_weights()[1]kkb=】項(xiàng)，阪旺莢頃bo)print(ll\nB=11,np.round(bo,2))y=npdot(x,kkw)+kkbz

=sigMoid(y)Print("\nZ=",np.round(zt2))|#...........coniinued.............................*...........coniinued......................defone_ruiiiid(x,t):globalmodeldxx=xriptiEwaxis,：：dtt=tjnp,newaxis,二model.fit(dtt,1,1,D,stuffle=False)deftiainir^()：globalmoiel,

kkw,kkbX=DX/255forepir；r；in^&(2000):for1inrange(S)ioneround(x[i],t[i])im~i=1.gwt^urwi(J[。]kkw=】項(xiàng)，￡qiMPZ：頃WQ)H*("，n■=、■'*儼*ig，ge----_11

Jprintf,np.rQundtwQ,2))bo=d.^et_weights()[1]kkb=】項(xiàng)，阪旺莢頃bo)print(ll\nB=11,np.round(bo,2))y=npdot(x,kkw)+kkbz

=sigMoid(y)Print("\nZ=",np.round(zt2))|#...........coniinued.............................#...........coniinued......................defane_ruiiiid(x,t):globalmodeldxx=xrip臨船乂is,：：dtt=tjnp,newaxis,二jiiode1hfit(dxx,d11,1,1,0,shuffle=aIse)deftiainir^()：globalmoiel,

kkw,kkbX=DX/255forepir；r；in^&(2000):for1inrange(S)ione_round(x[i],t[i])遷移到卷積核___________)_____#...........Continued.............................ddx=np.array([1,0,0,0,0,1],dtype=np.float32)defgetY(dxf

kw,kb):y=np.sum(ilx*kw)+kbreturnydefcdhv(x,kw,kb,stride)：

/xz=x.sizekz=kw.$izesteps=int((xz-kz)/stride)+1y=np.zeros((Steps),dtype=np.float32)foriinrange(steps):start=i*stridedx=x[start:start+l:z]y[i]=?etY(dx7kw,kb)returny#............continued.............................準(zhǔn)備卷積運(yùn)算函數(shù)continued...............................#...........Conv-1.............”)_____________________yl=Conv(tx,kkw,kkb,3)zlSIgmoldC^Ff/.....print(h\nprint(MZ1=",np.round(zlt

2))diefconvolution():TX=nP.array([255.0,0,0,0,255.0,255,。，0,0,255,255,0』，0?255f0],dtype=np.float32)tx=TX/255_k神2=np.array([1,1])z2=Conv(zl,kw2,0,1)...................print("\n-----conv-2--------print("Z2=",npround(z2,2))_________________________使用遷移來的卷積核Convolutioii()#End................................#_一二……training()#continuedConv-1IJT專家直接提供的卷積核丿yl=Conv(tx,kkw,kkb,3)zl=signu)id(yl)print(h\nprint(MZ1np.round(zlt

2))print("\nprint("Z2=",npround(z2,2))diefconvolution():TX=nP.array([255.0,0,0,0,255.0,255,。，0,0,255,255,0』，0?255f0],dtype=np.float32)tx=TX/255np.ar髯頁-ft,z2=Conv(zl1kw2,.0,1)Convolutioii()#Endtraining()-發(fā)現(xiàn)了2次跳機(jī)Z=[0.050.030.030,030.970.97]Z1=(0.970.03□.03(0,97)0.03]-----Conv-222=[0.990.050.990,99]t[88515rL7Io__yo-■41114-42242_=--woB]J1L-9-Z2[]都小于1.0，沒有〈連續(xù)跳機(jī)〉的現(xiàn)象Z=[0.050.030.050,030.970,97]Z2=[0.990.050.990,99]-----conv-1-----------*-Z1=[0.970.030.030,970.03]-d2owB]J1L-9t[88515rL7Io__yo-■41114-42242把專家直覺納入Al模型里|說明?專家直覺(ExpertIntuition)就是您可以看出來眼前的情況與過去發(fā)生情況的某些相似點(diǎn)(即相似特征)。?您的專門知識(shí)愈深，就愈能看出許多相似情況，而在菜鳥眼中，每個(gè)情況都是新且獨(dú)立的情況。?專家直覺帶給人們瞬間洞察力，也就是鑒往知來的能力。把專家直覺納入Al模型里|說明-例如，下圖是醫(yī)院里的護(hù)理師排班表:7ABCBCDBCEBCFBCGOFHOFIOFJBCKBCLBCMBCNBC0OFPBCQBCRBCsOFTOFuOFVOFwBCXBCYBCzBCAAOFABBCACBCADBCAEBC8BCBCOFBCBCBCBCOFBCBCBCOFBCBCBCBCOFOFOFJBBCBCOFOFBCBCBCBCOF9JBJBJBJBOFOFBCBCBCOFOFOFBCBCBCBCOFOFOFOFBCBCOFOFBCBCOFOFBC0R=OFOFOFBCBCOFBCBCOFJBJBJBJBOFOFOFOFBCBCBCBCBCBCBCOFBCBCBCOFOF11BE=BCOFBCBCOFBCBCBCBCOFBCBCBCOFOFOFJBBCBCBCBCOFOFBCBCOFBCBCBCOF2BCOFOFBCBCBCBCOFBCBCBCBCOFOFOFRABCBCBCBCOFOFOFBCBCOFOFBCBC3白底=BC=白班JBJBOFBCBCBCOFBCBCBCBCBCOFOFOFOFBCRABCBCOFOFOFBCBCOFOFOFBC14黃底夜BCBCBCBCBCOFOFOFBCBCBCBCOFOFOFOFBCBCJBBCOFOFBCBCBCOFOFBCBC5藍(lán)底攻人=大夜BCBCOFOFBCBCBCBCOFBCBCBCOFOFOFOFBCJBBCBCOFOFBCBCBCBCBCOFOF.6BCBCBCOFOFJBJBJBJBJBOFBCBCOFOFOFOFBCRABCOFOFOFBCBCBCOFBCBC70=100BCBCOFOFBCBCBCBCBCOFBCBCOFOFOFOFBCBCOFBCRAJBBCOFOFOFBCBCBC8E&I0IBCBCBCOFOFBCBCBCBCOFOFBCBCJBRABCOFOFOFOFBCBCBCOFOFBCBCBCBC.9|jB=110OFOFBCBCBCBCOFBCBCBCBCOFBCJBJBBCOFOFOFOFBCBCBCBCOFOFOFBCBC!0RA=111OFOFBCBCBCOFOFBCBCBCOFOFBCRARABCBCOFBCBCBCBCOFBCBCBCBCOFOF?其中：OF代表休假；BC代表白天班；JB代表小夜班；RA代表大夜班。把專家直覺納入Al模型里I說明-有經(jīng)驗(yàn)的護(hù)理師，一眼就能看出這不是一張好的排班內(nèi)容，其憑借的就是專家直覺。-如果我們能夠探知這位資深護(hù)理師所觀察到的特征，并且將其表現(xiàn)于AI模型里，就能大大提升AI系統(tǒng)的質(zhì)量。-例如，護(hù)理師們有一個(gè)概念稱為：花式排班。AI人員就去探知〈花式排班〉的特征，表達(dá)于AI模型上。，把專家直覺納入Al模型里|設(shè)計(jì)分類器?藉由分類器來吸納專家的智慧。-一旦分類器訓(xùn)練好了，就將分類器的W遷移過來，成卷積核(Kernel)。?有了卷積核就能進(jìn)行卷積運(yùn)算(Convo山tion)來自動(dòng)提取特征了。?這樣就引入專家的經(jīng)驗(yàn)、智慧(又稱為〈專家直覺＞)，納入到AI模型里。例如，專業(yè)術(shù)語〈花式排班〉就蘊(yùn)含了專家智慧。把專家直覺納入Al模型里設(shè)計(jì)FX（FeatureExtractor）分類器BN0PQR102護(hù)理排班Al模型03代表：OF（休息）04代表：BC（日班）05代表：JB（小夜班）06代表：RA（大夜班）07080代表GckxI09101H1121131014151161017181920211代表Bad訓(xùn)練1500回合LZAT丫環(huán)-1丫環(huán)一2丫環(huán)-3CDEFGIIIJKX[]100010001000010010000010100000010100100000101000000■1100001000100010000100100000100100100000■101000010001000100000001001000000022FX分類器神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)，把專家直覺納入Al模型里|展幵訓(xùn)練?于是專家直覺就成為FX分類器的內(nèi)涵了。-當(dāng)您按下＜訓(xùn)練〉按鈕，F(xiàn)X分類器就幵始學(xué)習(xí)了。?學(xué)習(xí)之后，這位資深護(hù)理的專業(yè)直覺，就成為這分類器的智慧了。并以權(quán)重來表達(dá)這項(xiàng)智慧，如下圖：，把專家直覺納入Al模型里|展幵訓(xùn)練護(hù)理排班AI模型0代表Ggl0.06學(xué)習(xí)完成!0.080.06-3.952.6850.0222.57969訓(xùn)練1500回合代表：QF（休息）代表：BC（日班）代表：JB（小夜班）代表：RA（大夜班）-6.798457.25021-6.842732.7792-0.2960.030.020.020.03丫環(huán)-1丫環(huán)-2丫環(huán)-30.450.2720.4591代表Bad權(quán)2.7585-3..121.18-0.162.765-3.980.581.1-0.2960.0225.46-3.91-0.3-3.885.472.77921.2010.62-3.952.685-0.21.12-3.2、-017585-3.121.18-0.161765-3.980.581.1把專家直覺納入Al模型里123YYYrI-6.79845-7.25021_-6.84273u2'

2579692.77921.2010.62-3.952.685-0.21.12-3.2-0.2960.0225.46-391-0.3-3.885.47-0把專家直覺納入Al模型里17585-3.121.18-0.1617653980.581.10.450.2720.4592.57969123.丫丫丫2把專家直覺納入Al模型里1319-6.798452.77921.2010.62-3.952.685-0.21.12-3.20.45207.250^1/丫環(huán)-2-0.2960.0225.46-3.91-0.3-3.885.47-0027221-6.^3-丫環(huán)-32.7585-3.121.18-0.161765-3.980.581.10.459222.57969*把專家直覺納入Al模型里從分類器遷移出來,成為卷積核丫環(huán)-12.7792丫環(huán)-2-0.2960.0220.272丫環(huán)一32.7585-0.162.765-3.980.459丫環(huán)的卷積核1819-6.79845207.2502121-5.84273222.57969r-3.952.685-0.21.12-3.2.從FX分類器遷移出來，成為卷積核丫環(huán)-2的卷積核2.579692223-6.798457.25021-5.84273181920210.450.2720.459從FX分類器遷移出來，成為卷積核丫環(huán)-12.77921.2010.62-3.952.685-0.2231.1：丫環(huán)-3的卷積核0.450.2720.459-6.798457.25021-6.842732.579591819202122準(zhǔn)備排班表（卷積的對象）匯入卷積核2.78Conv2.76庭4通道匯合匯入丫環(huán)智慧0.450.270.460ABCDEFGHIJKLMN0PQRSTUVWXYZAA1OFRARARAOF210000001000100010001000110034OFOFOFRARAOFOF510001000100000010001100010067亨JBOFOFOFJBJBOF00101000100010000010001010091011RARARAOFRARARA0001000100011000000100010003個(gè)丫環(huán)的卷積核把專家直覺納入Al模型里訓(xùn)練完成,得到卷積核-為了簡單起見，上圖里只列出4筆排班的原始資料（4位護(hù)士的本本月份排班表）。?其中：OF代表休假；BC代表白天班；JB代表小夜班；RA代表大夜班。?請按下〈卷積〉，就展開對原始數(shù)據(jù)進(jìn)行卷積運(yùn)算，來提取特征（具有花式排班的表征）。0.27-6.840.950.810.9870.9970.810.810.810.990.4980.020.530.990.5270.9970.310.310.5170.00]0.4880.0250.990.960.9870.8020.960.3060.990.00]0.990.990.0250.0190.990.710.8020.9550.310.]]0.00]0.020.990.]80.9970.0250.9390.960.990.8020.020.990.]7匯入丫環(huán)智慧卷積的結(jié)果針對第1筆資料3個(gè)丫環(huán)的卷積結(jié)果2.咫].20.6-0.305.5r卷積4通道ConvH匯合0.990.950.50.020.530.32-45.5-0-4().6].]]]0.&。.&0.50000.5().300ABCDEFGH1JKLMN0PQj]]000]000]000]100012AAABACADAEA]]1213141a1516上17182237卷積的結(jié)果F0A00450.270.46B0C0E0卷積ConvD]J0通道匯合H]]_0L]針對第2筆資料3個(gè)丫環(huán)的卷積結(jié)果G0W0V0]00匯入丫環(huán)智慧Z000]00.030.S000.8000.800.03。盅00.03。盅00.03。盅00.03。盅0.310.020.9610.421-6.87.25-6.842.5S0.50.5]3.2].]0.520.490.99MN0P10002.咫].20.6-4-0.305.5392.763]1.2-0.2190.5200000.3]0.950.810.80.9幻0.9970.950.&0.S0..]0.&。.&。屈0.990.950.810.&]0.81。屈0.99]200.490.030.030.030.0300.420.020.0300.4980.420.02000.50000.03().50.020.030.030.030.03().502]{).00.R0.R0.R0.8]]0.32000.5270.9970.32000.5().30000.530.320000().530.31]]0.5200.3：]]0.52000.30.5170.00]004‘0.3]]]0.420.420.490.030.020.40.490.030.0300.4880.0250.030000.40.40.4*]]0.990.R0.96]0.990.&0E]0.9幻0.8020E0.&o.s]]]]r>、館290.9S]]0.990.9]0.5200.080.]0.00]0.3061]]]0.90.90.9]]0.990.90.90.90.9&0.990.90.90.930]0.420.42]]]0.490.030.86]0.0250.0190.420.40.4]]]]]0.4]]]1]]]]]310.99]]0.980.9]0.990.&0.710.90.8020.9551]]]0.90.90.9]]0.980.90.90.90.990.980.90.90.9323334000.3]0.52000.31]]]0.：]]0.00]0.020.9]]0.90.90.9]]0.990.90.980.5200.02o.is00.3350.030.030.020.490.0300.020.420.42]0.9970.0250]]]]]]]0.4]]]0.490.0300.570.030360.80.R0.960.990.8o.s0.96]]]0.9390.8020.020.]]]0.90.90.9]]0.980.90.990.990.80.020.170.8]38AAABACADAEA0.322]27323334353637卷積的結(jié)果針對第3筆資料3個(gè)丫環(huán)的卷積結(jié)果0450.270.46]00Z000]W000.50.51).3.]]、0.40.4]]]*3)00.03。盅00.03。盅0.810.030.8000.80.8ABCDEFGH1JKLMX0PQ000]000]000]10000121314匯入卷積通道2.咫].20.6-42.715丫環(huán)智慧Conv匯合-0.305.539-0162.763]1.2-0.22.81718zHi19().5200000.3]0.950.810.80.9幻0.9970.950.&0.S0.80.1200.490.030.030.030.0300.420.020.0300.4980.420.0200°1210.990.R0.R0.R0.8]]0.32000.5270.9970.32000J-6.87.25-6.842.580.990.950.810.&]0.R1。屈0.990.50.020.030.030.030.030.50.530.320000().530.3]10.5200000.020.420.490.030.030.0300.9610.990.&0.R0.8。盅2324]]0.5200.3：]]0.52000.30.5170.00]0000.3250.420.420.490.030.020.40.490.030.0300.4880.0250.0300026]]0.990.R0.96]0.990.&0E]0.9幻0.8020E0.&O.S]*0.98]]0.990.9]0.5200.080.]0.00]0.3061]]]0.90.90.9]]0.990.90.90.90.9&0.990.90.90.9]0.420.42]]]0.490.030.86]0.0250.0190.420.40.4]]]]]0.4]]]1]]]]]0.99]]0.980.9]0.990.&0.710.90.8020.9551]]]0.90.90.9]]0.980.90.90.90.990.980.90.90.9000.3]0.52000.31]]]0.：]]0.00]0.020.9]]0.90.90.9]]0.990.90.980.5200.02o.is00.30.030.030.020.490.0300.020.420.42]0.9970.0250]]]]]]]0.4]]]0.490.0300.570.0300.80.R0.960.990.8O.S0.96]]]0.9390.8020.020.]]]0.90.90.9]]0.980.90.990.990.80.020.170.8]38wAAABACADAEA-6.80.452.580.8000022]0.500.40.50]]0.8270.9().9]]]0.4Z90.90.9]]卷積的結(jié)果針對第4筆資料3個(gè)丫環(huán)的卷積結(jié)果7.25-6.840.270.46]00]]Z000]X00]ABCDEFGH1JKLMX0PQ000]000]000]10000121314匯入卷積通道2.咫].20.6-42.715丫環(huán)智慧Conv匯合-0.305.539-0162.763]1.2-0.22.8171819().5200000.3]0.950.810.80.9幻0.9970.950.&0.S0.8200.490.030.030.030.0300.420.020.0300.4980.420.02000210.990.R0.R0.R0.8]]0.32000.5270.9970.32000。屈0.990.950.810.&]0.81。屈0.99]0.03().50.020.030.030.030.03().5000.530.320000().530.3000.3]10.5200000.030.030.020.420.490.030.030.0300.R0.R0.9610.990.&0.R0.8。盅0.990.90.90.90.980.990.90.90.9]]]1]]]]]旃0.90.90.90.990.9&0.90.90.92324]]0.5200.3：]]0.52000.30.5170.00]0000.3i250.420.420.490.030.020.40.490.030.0300.4880.0250.0300o126]]0.990.R0.96]0.990.&0E]0.9幻0.8020E0.&O.S]J28290.98]]0.990.9]0.5200.080.]0.00]0.3061]]30]0.420.42]]]0.490.030.86]0.0250.0190.420.40.4],310.99]]0.980.9]0.990.&0.710.90.8020.9551]]]]32與，V鈿000.3]0.52000.31]]]0.：]]0.00]0.020.9]0.030.030.020.490.0300.020.420.42]0.9970.0250]]]0.80.R0.960.990.8o.s0.96]]]0.9390.8020.020.]]]0.90.90.9]]0.990.90.980.5200.02o.is00.3]]]]0.4]]]0.490.0300.570.0300.90.90.9]]0.980.90.990.990.80.020.170.8]38交由大丫環(huán)匯合特征表-得到大丫環(huán)的特征表丿這筆有兩處花式排班202]2200000023242526270000003800000000000000000000000000000000000000000002829303132000000000000000000000000000033343536370000000000000000000000000交由大丫環(huán)匯合特征表-得到大丫環(huán)的特征表這筆有許多花式排班丿I于是，得到大丫環(huán)的特征表-您可以看到了，青色底的部分特征值為1，表示發(fā)現(xiàn)到有〈花式排班〉的表征。I進(jìn)行池化(Pooling)-然后，進(jìn)行CNN的池化(Pooling)運(yùn)算，萃取一周內(nèi)是否出現(xiàn)〈花式排班〉現(xiàn)象。把池化的特征表,交給格格-最后，建立全連接層(FCL)分類器，并進(jìn)行訓(xùn)練。把專家直覺納入Al模型里|訓(xùn)練完成-這樣就完成了〈排班＞CNN深度學(xué)習(xí)的模型的設(shè)計(jì)與訓(xùn)練了。-目前已經(jīng)完成〈排班〉模型的訓(xùn)練階段了。I進(jìn)行測試-現(xiàn)在就拿兩位新護(hù)士的排班表，來給AI評(píng)估看看?？纯词欠裼胁涣嫉幕ㄊ脚虐喱F(xiàn)象。把專家直覺納入Al模型里|測試結(jié)果-預(yù)測的結(jié)果呈現(xiàn)于粉紅色底的部分：?第1位新護(hù)士的排班是正常的。-第2位新護(hù)士的排班則并不理想。把專家直覺納入Al模型里更上一層樓-以上的范例，只是展示〈特征提取器＞的基礎(chǔ)能力：支持基本的分類任務(wù)。例如，辨別＜好＞與〈不好〉的排班表。-基于這項(xiàng)基礎(chǔ)，未來可以進(jìn)一步組合AE模型，來進(jìn)行理想的排班補(bǔ)值（生成），由AI來幫助您做〈智慧排班〉的工作。?祝福您輕松愉快更上一層樓。范例實(shí)踐I延續(xù)上一小節(jié)的護(hù)理排班范例10RARARAOFRARARAOF110001000100011000000100010001101218范例實(shí)踐步驟1：建立FX(FeatureExtractor)模型-目標(biāo)一FX模型，能吸納專家的智慧：即分辨〈花式排班〉。-方法--設(shè)計(jì)一個(gè)分類器，讓專家貼上標(biāo)簽，進(jìn)行訓(xùn)練，學(xué)習(xí)專家直覺。范例實(shí)踐|步驟2:遷移出3個(gè)卷積核(Kernel)-目標(biāo)一建立特征提取器，例如卷積核(Kernel)。-方法一從(步驟1)已經(jīng)訓(xùn)練好的分類器里，遷移出Wh成為3個(gè)卷積核。范例實(shí)踐|步驟3:幵始進(jìn)行卷積-目標(biāo)一針對排班表進(jìn)行特征提?。淳矸e），如同專家審視排班表。-方法--讓3個(gè)卷積核（昵稱：丫環(huán)），對排班表進(jìn)行卷積。范例實(shí)踐|步驟4:將3個(gè)特征表（Featuremap）匯合-目標(biāo)一針對排班表的每一筆，得出一個(gè)特征表。-方法一使用分類器的Wo來3個(gè)（丫環(huán)）特征表，計(jì)算（匯合）出單一特征表。范例實(shí)踐|步驟5:設(shè)計(jì)高層分類器（昵稱：格格），并進(jìn)行訓(xùn)練。-目標(biāo)一設(shè)計(jì)&訓(xùn)練格格（分類器）。-方法一拿（步驟4）特征表，作為格格分類模型的輸入（訓(xùn)練）資料，并展幵訓(xùn)練。范例實(shí)踐|步驟6:匯出FX和FCL兩個(gè)模型-目標(biāo)一提供兩個(gè)*.pb檔案給OpenVINO。-方法一將訓(xùn)練好的FX和FCL模型導(dǎo)出到*.pb檔案。范例實(shí)踐|撰寫Python程序?撰寫一支Python程序。-訓(xùn)練好的FX和FCL模型導(dǎo)出到*.pb檔案。],dtype=np,floatB2)CIDI苴BFa/rly([[LO.0,0,1,0,0,13][1,0,0,0,OJ.O^CJ[l,0f0jc,0,03,0][1,0,0』，0,0,0,1][0,1,0.11,1,0.0,01[0,0,L0,l,0,0,0][0J.0J,1,0,0,01[0Jf0,0,03,0,0][DJ.OJO,0,0,1,01[04,0,0,CI,O』，1][0JJf0,0,1,0,01[0,0,0J,0,1,0,0][OfOJFOf0,0,1,03[OfOJ,O70,0,03][D,OfO,l,0,03,0][0,0,0,1,0,0,0,1]dtt=np.array([O,O,OfO,0,0,0,0,lfl,O],dtype=np,float32)#---------tinued---————-——-二^--------------------------defsigmoid(x):return1/(14-np.6xp(-x))準(zhǔn)備FX（特征提取）模型的訓(xùn)練數(shù)據(jù)#ex_Al1_06.pyimportnumpy豎npfromkerasmodelsimportSequentialfromkeras.layersimpo11ActIvationfDensefromkeras.optimizersimportSGDkerashack巳ndKimporttensorfIowastf訓(xùn)練1500回合N(LP護(hù)理排班Al梧代表：OF<代表：BC代表：IB（代表：RA0代表Good1代表Bad1000010010000010100000010100100000101000000I10000100卻1000100v00100100000■100100100000101000010001000100001000I00100001000■1AT0000FileEditFormatRunOptionsV/indowHelp撰寫Python程序whi=[np.lrriy(continued[[0[-0.[0,[0,ro.[o,[-0.[-0,np.aU.1],0.0,0J]0,1]0J]0.1]0J]0.1]0J]dtype^np,float32),0.0],dtype=np.f1oat32權(quán)重初期woi=[np.array([[-0.1],[0,5],>0,1]],dtype=np,float32),np.array([0,0],dtype=np.float32)]丿kernel=Nonebias=Nonewo=Nonebo=Nonefeature_map=Nonepo_fm=Nonemodel_l=NonAmodel2=NonpContinued準(zhǔn)備FX的丿o?5■555555