目 第一章 張量 創(chuàng)建操作Creation 索引,切片,連接,換位Indexing,Slicing,Joining,Mutating 隨機抽樣Random 序列化 并行化 數(shù)學(xué)操作Math Reduction 比較操作Comparison 其它操作Other BLASandLAPACK 第二章 第三章 第四章 第五章 第六章 第七章 第八章Automaticdifferentiationpackage- 第九章 第十章 第十一章 第十二章 第十三章 第十四章 torch包含了多維張量的數(shù)據(jù)結(jié)構(gòu)以及基于其上的多種數(shù)學(xué)操作。另外，它也提供了多它有CUDANVIDIAGPU上進行張量運算(計算能力>=2.0)objpytorchTrueobj(Object)–判斷對象objpytorchstorageinputObject)–torch. 張量中的元素個數(shù):input(Tensor)>>>a=>>>torch.numel(a)>>>a=>>>torch.numel(a) edgeitems=None,linewidth=None,profile=None)Numpyprecision–(8threshold–閾值，觸發(fā)匯總顯示而不是完全顯示(repr)（edgeitems–匯總顯示中，每維（軸）兩端顯示的項數(shù)（linewidth–用于插入行間隔的每行字符數(shù)（80）。Thresholdedmatricieswillthisprofilepretty(shortCreationtorch.eye(n,m=None,nint)–mintoptional)–列數(shù).如果為None,out(Tensor,optinal)-Output返回值:102維張量返回值類型Tensor>>>100100[torch.FloatTensorofsizetorch.from_numpy(ndarray)→Numpynumpy.ndarraypytorchTensorndarray共享同一內(nèi)存空間。修改一個會導(dǎo)致另外一個也被修改。返回的張量不能改變大>>>a=numpy.array([1,2,>>>t=>>>torch.LongTensor([1,2,>>>t[0]=->>>array([-1,2,torch.linspace(start,end,steps=100,out=None)→1startend上均勻間隔的step1維張量start(float)–end(float)–stepsint)–startendoutTensoroptional)–>>>torch.linspace(3,10,[torch.FloatTensorofsize>>>torch.linspace(-10,10,--[torch.FloatTensorofsize>>>torch.linspace(start=-10,end=10,--[torch.FloatTensorofsizetorch.logspace(start,end,steps=100,out=None)→1量的長度為stepsstartfloat–endfloat–stepsint–startendoutTensoroptional)–結(jié)果張量>>>torch.logspace(start=-10,end=10,[torch.FloatTensorofsize>>>torch.logspace(start=0.1,end=1.0,[torch.FloatTensorofsizetorch.ones(*sizes,out=None)→1sizessizesint...)–outTensoroptional)–>>>torch.ones(2,1111[torch.FloatTensorofsize>>>[torch.FloatTensorofsizetorch.rand(*sizes,out=None)→返回一個張量，包含了從區(qū)間[0,1)的均勻分布中抽取的一組隨機數(shù)，形狀由可變參數(shù)sizessizesint...)–outTensoroptinal)->>>[torch.FloatTensorofsize>>>torch.rand(2, [torch.FloatTensorofsizetorch.randn(*sizes,out=None)→返回一個張量，包含了從標準正態(tài)分布(01，即高斯白噪聲)中抽取一組隨sizessizesint...)–outTensoroptinal)->>>--[torch.FloatTensorofsize>>>torch.randn(2,1.43390.3351-1.5458-0.9643-[torch.FloatTensorofsizetorch.randperm(n,out=None)→給定參數(shù)n0n-1nint)–上邊界(不包含>>>[torch.LongTensorofsizetorch.arange(start,end,step=1,out=None)→1floor((end?start)/step)startendstep為步長start(float)–end(float)–step(float)–outTensoroptional)–>>>torch.arange(1,[torch.FloatTensorofsize>>>torch.arange(1,2.5,[torch.FloatTensorofsizetorch.range(start,end,step=1,out=None)→1floor((end?start)/step)+1個元素。包含在半開區(qū)間[start,end）startstart(float)–end(float)–stepint)––>>>torch.range(1,[torch.FloatTensorofsize>>>torch.range(1,4,[torch.FloatTensorofsizetorch.zeros(*sizes,out=None)→0sizessizesint...)–outTensoroptional)–>>>torch.zeros(2,0000[torch.FloatTensorofsize>>>[torch.FloatTensorofsize索引,切片,連接,換位IndexingSlicingJoiningMutatingtorch.cat(inputs,dimension=0)→torch.cat()可以看做torch.split()torch.chunk()cat()函數(shù)inputssequenceofTensors)–可以是任意相同Tensorpythondimensionintoptional)–>>>x=torch.randn(2,>>>0.5983-0.03411.5981-0.5265-[torch.FloatTensorofsize>>>torch.cat((x,x,x),0.5983-0.03411.5981-0.5265-0.5983-0.03411.5981-0.5265-0.5983-0.03411.5981-0.5265-[torch.FloatTensorofsize>>>torch.cat((x,x,x),0.5983-0.03412.49180.5983-0.03412.49180.5983-1.5981-0.5265-0.87351.5981-0.5265-0.87351.5981--[torch.FloatTensorofsizetorch.chunk(tensor,chunks,tensorTensor)–chunksint)–dimint)–torch.gather(inputdimindexout=None)Tensordimindex指定位置的值進行聚合。3維張量，輸出可以定義為：out[i][j][k]=tensor[index[i][j][k]][j][k]#out[i][j][k]=tensor[i][index[i][j][k]][k]#out[i][j][k]=tensor[i][j][index[i][j][k]]#>>>t=>>>torch.gather(t,1, [torch.FloatTensorofsizeinputTensor)–dimint)–indexLongTensor)–outTensoroptional)–torch.index_select(input,dim,index,out=None)→_Tensor_有相同的維度(在指定軸上)。注意：返回的張量不與原始張量共享內(nèi)存空間。input(Tensor)dimint)–indexLongTensor)–outTensoroptional)–>>>x=torch.randn(3,>>>1.20452.40840.40010.55961.56770.6219-1.3635-1.2313-0.5414-[torch.FloatTensorofsize>>>indices=torch.LongTensor([0,>>>torch.index_select(x,0,1.20452.40840.40011.3635-1.2313-0.5414-[torch.FloatTensorofsize>>>torch.index_select(x,1,1.20450.55961.3635-[torch.FloatTensorofsizetorch.masked_select(input,mask,out=None)→mask中的二元值，取輸入張量中的指定項mask為一個ByteTensor)，將取1Dinput(Tensor)maskByteTensor)–outTensoroptional)–>>>x=torch.randn(3,>>>1.20452.40840.40010.55961.56770.6219-1.3635-1.2313-0.5414-[torch.FloatTensorofsize>>>indices=torch.LongTensor([0,>>>torch.index_select(x,0,1.20452.40840.40011.3635-1.2313-0.5414-[torch.FloatTensorofsize>>>torch.index_select(x,1,1.20450.55961.3635-[torch.FloatTensorofsizetorch.nonzero(input,out=None)→中非零元素索引的張量。輸出張量中的每行包含輸入中非零元素有n維，則output的形狀zxn,z inputTensor)–outLongTensoroptional)–>>>>>>torch.nonzero(torch.Tensor([1,1,1,0,[torch.LongTensorofsize>>>torch.nonzero(torch.Tensor([[0.6,0.0,0.0, [0.0,0.4,0.0, [0.0,0.0,1.2, [0.0,0.0,0.0,-0123[torch.LongTensorofsizetorch.split(tensor,split_size,tensorTensor)–split_size(int)–dimint)–torch.squeeze(input,dim=None,當給定dim為:(A×1×B),squeeze(input,0)將squeeze(input1)，形狀會變成(A×B)input(Tensor)dimintoptional)–只會在outTensoroptional)–>>>x=>>>(2L,1L,2L,1L,>>>y=>>>y.size()(2L,2L,>>>y=torch.squeeze(x,>>>(2L,1L,2L,1L,>>>y=torch.squeeze(x,>>>y.size()(2L,2L,1L,2L)torch.stack(sequence,sqequenceSequence)–dimint)–0torch.t(input,out=None)→input(Tensor)outTensoroptional)–>>>x=torch.randn(2,>>>0.48340.6907-0.13000.5295[torch.FloatTensorofsize>>>0.4834-0.69071.3417[torch.FloatTensorofsizetorch.transpose(input,dim0,dim1,out=None)→的轉(zhuǎn)置dim0dim1input(Tensor)dim0int)–dim1int)–>>>x=torch.randn(2,>>>0.5983-0.03411.5981-0.5265-[torch.FloatTensorofsize>>>torch.transpose(x,0,0.5983-0.0341-2.4918-[torch.FloatTensorofsizetorch.unbind(tensor,tensor(Tensor)dimint)–torch.unsqueeze(input,dim,dimdim+input.dim()+1tensor(Tensor)dimint)––>>>x=torch.Tensor([1,2,3,>>>torch.unsqueeze(x,123[torch.FloatTensorofsize>>>torch.unsqueeze(x,[torch.FloatTensorofsizeRandomseedintorlong)–種子.返回生成隨機數(shù)的原始種子值（pythonlong）。new_statetorch.ByteTensor)–torch.default_generator=<torch._C.Generatorobject>torch.bernoulli(input,out=None)→從伯努利分布中抽取二元隨機數(shù)(01)［0,1］ii1。01:––輸出張量(可選>>>a=torch.Tensor(3,3).uniform_(0,1)#generateauniformrandommatrixwithrange[0,1]>>> [torch.FloatTensorofsize>>> [torch.FloatTensorofsize>>>a=torch.ones(3,3)#probabilityofdrawing"1"is>>> [torch.FloatTensorofsize>>>a=torch.zeros(3,3)#probabilityofdrawing"1"is>>> [torch.FloatTensorofsizetorch.multinomial(input,num_samples,replacement=False,out=None)→相應(yīng)行中定義的多項分布中抽取的num_samples個樣[注意]:每行1(這里我們用來做權(quán)重)，但是必須非負且總和不0。是一out也是一個相同長度num_samples的向moutm×n的矩陣。replacement為True,則樣本抽取可以重復(fù)。否則，一個樣本在每行不能被重復(fù)抽num_samples必須(如果)–num_samplesint)–replacementbooloptional)––>>>weights=torch.Tensor([0,10,3,0])#createaTensorof>>>torch.multinomial(weights,[torch.LongTensorofsize>>>torch.multinomial(weights,4,[torch.LongTensorofsize4]torch.normal(means,std,means,stdmeans是一meansTensor)–stdTensor)–outTensor)–[torch.FloatTensorofsize10]torch.normal(mean=0.0,std,out=None)meansTensor,optional)–stdTensor)–outTensor)–可選的輸出張量>>>torch.normal(mean=0.5,std=torch.arange(1,[torch.FloatTensorofsizetorch.normal(means,std=1.0,meansTensor)–std(floatoptional)–outTensor)–可選的輸出張量[torch.FloatTensorofsizeRecommendedapproachforsavingaobj–f(返回文件描述符)pickle_module–picklingpickle_protocol–pickleprotocalmap_location動態(tài)地進行內(nèi)存重映射，使其能從不動設(shè)備中讀取文件。一般調(diào)用時，需兩個參數(shù):storage和locationtag.storageNone(此時地址可以通過默認方法進行解析).認情況下，locationtags中"cpu"hosttensors，‘cuda:device_ide.g.‘cuda:2’)cudaf–(返回文件描述符)map_location–remappickle_module–unpickling(pickle_module>>>#Loadalltensorsontothe>>>torch.load('tensors.pt',map_location=lambdastorage,loc:storage)#MaptensorsfromGPU1toGPU0數(shù)學(xué)操作Mathtorch.abs(input,out=None)→outTensoroptional)–>>>a=>>>>>> torch.add(input,value,FloatTensororDoubleTensorvalue必須為實數(shù)，否則須為整數(shù)?！咀g注：似乎并非如此，無關(guān)輸入類型，value取整數(shù)、實數(shù)皆可?！縱alue(Number)––>>>a=>>>>>>torch.add(a,[torch.FloatTensorofsize4]torch.add(input,value=1,other,out=None)otherFloatTensororDoubleTensorvalue必須為實數(shù)，否則須為整數(shù)?！咀g注：似乎并非如此，無關(guān)輸入類型，value取整數(shù)、實數(shù)皆可。】–value(Number)–otherTensor)––>>>import>>>a=>>>>>>b=torch.randn(2,>>>1.0663-0.1513>>>torch.add(a,10,用tensor2對tensor1逐元素相除，然后乘以標量值value并加到。 tensorTensor)–tensor1tensorvalue(Numberoptional)–tensor1tensor2tensor1Tensor)–張量，作為被除數(shù)(分子(–>>>t=torch.randn(2,>>>t1=torch.randn(1,>>>t2=torch.randn(6,>>>torch.addcdiv(t,0.1,t1,0.0122-0.0188-0.7396-1.5721。張FloatTensororDoubleTensorvalue必tensorTensor)–tensor1tensorvalue(Numberoptional)–tensor1tensor2tensor1Tensor)––>>>t=torch.randn(2,>>>t1=torch.randn(1,>>>t2=torch.randn(6,>>>torch.addcmul(t,0.1,t1,0.0122-0.0188-0.7396-1.5721torch.asin(input,out=None)→–>>>a=>>>>>> torch.atan(input,out=None)→–>>>a=>>>>>>torch.atan2(input1,input2,out=None)→input1Tensor)–input2Tensor)––>>>a=>>>>>>torch.atan2(a,torch.ceil(input,out=None)→張量每個元素向上取整,即取不小于每個元素的最小整數(shù)，并返回結(jié)果到輸–>>>a=>>>>>>torch.clamp(input,min,max,out=None)→ |min,ifx_i<y_i=|x_i,ifmin<=x_i<= |max,ifx_i>FloatTensororDoubleTensorminmax必須為實數(shù)，否則須為整數(shù)?！咀g注：似乎并非如此，無關(guān)輸入類型，minmax取整數(shù)、實數(shù)皆可?！縨inNumber)–maxNumber)––>>>a=>>>>>>torch.clamp(a,min=-0.5,[torch.FloatTensorofsize4]torch.clamp(input,*,min,out=None)→Tensorvalue(Number)––>>>a=>>>>>>torch.clamp(a,torch.clamp(input,*,max,out=None)→FloatTensororDoubleTensormax必須為實數(shù)，否則須為整數(shù)。【譯注：似乎并非如此，無關(guān)輸入類型，max取整數(shù)、實數(shù)皆可?！縱alue(Number)––>>>a=>>>>>>torch.clamp(a,torch.cos(input,out=None)→–>>>a=>>>>>>–>>>a=>>>>>>torch.div(input,value,out=tensor/valueFloatTensororDoubleTensorvalue必須為實數(shù)，否則須為整數(shù)。【譯注：似乎并非如此，無關(guān)輸入類型，value取整數(shù)、實數(shù)皆可。】value(Number)––>>>a=>>>>>>torch.div(a,[torch.FloatTensorofsize5]torch.div(input,other,out=None)和other逐元–張量(分子otherTensor)–張量(分母–>>>a=>>>-0.18100.40170.2863-0.61832.06960.9012-0.56790.4743-0.0117--0.12130.96290.2682>>>b=torch.randn(8,>>>0.87740.8866-0.64901.4259-1.4633-0.4643-0.34921.6103->>>torch.div(a,-0.20620.52510.3229--0.95281.85250.63200.3881-3.8625-0.0253-0.34730.63060.1666-[torch.FloatTensorofsize–>>>torch.exp(torch.Tensor([0,math.log(2)]))torch.FloatTensor([1,2])torch.floor(input,out=None)→:–>>>a=>>>>>>-–divisorTensororfloat)–除數(shù)，一個數(shù)或與被除數(shù)相同類型的張量-–輸出張量>>>torch.fmod(torch.Tensor([-3,-2,-1,1,2,3]),torch.FloatTensor([-1,-0,-1,1,0,>>>torch.fmod(torch.Tensor([1,2,3,4,5]),參考torch.remainder(),python%torch.lerp(start,end,weight,start，endstartTensor)–endTensor)–weightfloat)––>>>start=torch.arange(1,>>>end=>>>>>>[torch.FloatTensorofsize>>>torch.lerp(start,end,torch.log(input,out=None)→–>>>a=>>>>>> –>>>a=>>>>>>torch.mul(input,value,FloatTensororDoubleTensorvalue必須為實數(shù)，否則須為整數(shù)?！咀g注：似乎并非如此，無關(guān)輸入類型，value取整數(shù)、實數(shù)皆可?！縱alue(Number)––>>>a=>>>>>>torch.mul(a,[torch.FloatTensorofsize3]torch.mul(input,other,out=None)注意：當形狀不匹配時，的形狀作為輸入張–otherTensor)––>>>a=>>>-0.72800.0598-1.4327--0.1427-0.06900.0821--0.92410.51100.4070--0.83080.7426-0.6240->>>b=torch.randn(2,>>>0.0430-1.07750.60151.1647-0.65490.0308-0.1670-1.25930.0292-0.08490.45301.2404-0.4659-0.1840>>>torch.mul(a,-0.0313-0.0645-0.8618-0.0934-0.0021-0.0137-1.16380.0149-0.0346--1.0304-0.34600.1148-–>>>a=>>>>>>的按exponent次冪exponent可以為單float相同元素數(shù)的張量。exponentfloatorTensor)––>>>a=>>>>>>torch.pow(a,>>>exp=torch.arange(1,>>>a=torch.arange(1,>>>>>>>>>torch.pow(a, [torch.FloatTensorofsize4]torch.pow(base,input,out=None)base為標量浮點值,為張out與輸basefloat)–inputTensor)––>>>exp=torch.arange(1,>>>base=>>>torch.pow(base,–>>>a=>>>>>>張量–divisorTensororfloat)––>>>torch.remainder(torch.Tensor([-3,-2,-1,1,2,3]),torch.FloatTensor([1,0,1,1,0,>>>torch.remainder(torch.Tensor([1,2,3,4,5]),參考:torch.fmod()C–>>>a=>>>>>>–>>>a=>>>>>> –>>>a=>>>>>>[torch.FloatTensorofsize–>>>a=>>>>>>torch.sin(input,out=None)→–>>>a=>>>>>>–>>>a=>>>>>>torch.sqrt(input,out=None)→–>>>a=>>>>>> torch.tan(input,out=None)→–>>>a=>>>>>>–>>>a=>>>>>>–>>>a=>>>>>>Reductiontorch.cumprod(input,dim,out=None)→yi=x1?x2?x3?...?xidimint)––>>>a=>>>>>>a[5]=torch.cumsum(input,dim,out=None)→yi=x1+x2+x3+...+xidimint)––>>>a=>>>torch.dist(input,other,p=2,out=None)→(other)的potherTensor)–pfloat,optional)––>>>x=>>>>>>y=>>>>>>torch.dist(x,y,3.5)>>>torch.dist(x,y,3)>>>torch.dist(x,y,0)>>>torch.dist(x,y,1)torch.mean(input)→>>>a=torch.randn(1,>>>-0.2946-0.9143>>>torch.mean(a)torch.mean(input,dim,out=None)→Tensordim上每行的均值。dim(int)–thedimensionto–>>>a=torch.randn(4,>>>-1.2738-0.30580.1230-0.8771-0.5430-0.92331.41070.0317-0.6823-1.38540.4953-0.2160>>>torch.mean(a,torch.median(input,dim=-1,values=None,indices=None)->(Tensor,:dimint)–valuesTensoroptional)–indicesTensoroptional)–>>>-0.6891-0.26970.5254-0.5528->>>a=torch.randn(4,>>>0.4056-0.33721.0973-2.48842.13360.38410.1404-0.1821--0.24031.3975-2.00680.1298-1.5371-0.7257-0.4871-0.2359->>>torch.median(a,1)dimLongTensor，包含眾數(shù)職的索引。dim值默認:dimint)–valuesTensoroptional)–indicesTensoroptional)–>>>-0.6891-0.26970.5254-0.5528->>>a=torch.randn(4,>>>0.4056-0.33721.0973-2.48842.13360.38410.1404-0.1821--0.24031.3975-2.00680.1298-1.5371-0.7257-0.4871-0.2359->>>torch.mode(a,1)torch.norm(input,p=2)→ppfloat,optional)–>>>a=torch.randn(1,>>>-0.4376-0.5328>>>torch.norm(a,torch.norm(input,p,dim,out=None)→dimppfloat)–dimint)––>>>a=torch.randn(4,>>>-0.6891-0.26970.5254-0.5528->>>torch.norm(a,2,>>>torch.norm(a,0,d(input)→>>>a=torch.randn(1,>>>0.61700.3546>>>d(a)d(input,dim,out=None)→dimint)––>>>a=torch.randn(4,>>>0.1598--0.1831--0.9925--0.2416->>>d(a,torch.std(input)→->>>a=torch.randn(1,>>>-1.30631.4182->>>torch.std(a)torch.std(input,dim,out=None)→dimint)––>>>a=torch.randn(4,>>>0.1889-2.48560.0043-0.7701-0.4682-2.24100.1919-1.1856-1.03610.01731.06620.2143->>>torch.std(a,torch.sum(input)→>>>a=torch.randn(1,>>>0.61700.3546>>>torch.sum(a)torch.sum(input,dim,out=None)→dimint)––>>>a=torch.randn(4,>>>-0.46400.06090.1122-1.30631.64430.4714--1.3561-0.19591.0609-2.68330.5746-0.5709->>>torch.sum(a,torch.var(input)→>>>a=torch.randn(1,>>>-1.30631.4182->>>torch.var(a)torch.var(input,dim,out=None)→dim(int)–thedimensionto–>>>a=torch.randn(4,>>>-1.2738-0.30580.1230-0.8771-0.5430-0.92331.41070.0317-0.6823-1.38540.4953-0.2160>>>torch.var(a,Comparisontorch.eq(input,other,out=None)→–otherTensororfloat)––ByteTensoror10torch.equal(tensor1,tensor2)→torch.ge(input,other,out=None)→–otherTensororfloat)–float–ByteTensor或者與第torch.ByteTensor張量，包含了每個位置的比較結(jié)果(inputother回類型：Tensor10torch.gt(input,other,out=None)→，即–otherTensororfloat)–float–ByteTensor或者與第torch.ByteTensor張量，包含了每個位置的比較結(jié)果(inputother回類型：Tensor00(values,indices)indices中沿dim維的k個最小值–kint)–kdimint,optional)–outtuple,optional)–(TensorLongTensor)>>>x=torch.arange(1,>>>>>>torch.kthvalue(x,4)otherinput<=other第二個參數(shù)可以為一個數(shù)或與第一個參數(shù)–otherTensororfloat)–float–ByteTensor或者與第torch.ByteTensor張量，包含了每個位置的比較結(jié)果(inputother回類型：Tensor11torch.lt(input,other,out=None)→otherinput<other第二個參數(shù)可以為一個數(shù)或與第一個參數(shù)–otherTensororfloat)–float–ByteTensor或者與第torch.ByteTensor張量，包含了每個位置的比較結(jié)果(tensorother回類型：Tensor01>>>a=torch.randn(1,>>>0.4729-0.2266->>>dim1，其它與輸入形狀保持一致。dimint)––,–>>a=torch.randn(4,>>0.06920.31421.2513-0.92880.8552-0.20731.0695-0.0101-2.4507-0.7426-0.76660.4862->>>torch.max(a,1)otherTensor)––>>>a=>>>>>>b=>>>>>>torch.max(a,torch.min(input)→:input(Tensor)>>>a=torch.randn(1,>>>0.4729-0.2266->>>dim1，其它與輸入形狀保持一致。dimint)––(,–>>a=torch.randn(4,>>0.06920.31421.2513-0.92880.8552-0.20731.0695-0.0101-2.4507-0.7426-0.76660.4862->>torch.min(a,torch.LongTensorofsize4x1]torch.min(input,other,out=None)→Tensor注意：當形狀不匹配時，的形狀作為返回張量的形狀。otherTensor)––>>>a=>>>>>>b=>>>>>>torch.min(a,torch.ne(input,other,out=None)→otherinput!=other–otherTensororfloat)–float–ByteTensortorch.ByteTensor(tensorotherTrue1)。Tensor01torch.sort(input,dim=None,descending=False,out=None)->(Tensor,(sorted_tensorsorted_indices)sorted_indices–dimint,optional)–descendingbool,optional)–outtuple,optional)–ByteTensor相同類>>>x=torch.randn(3,>>>sorted,indices=>>>-1.67470.06100.1190-1.47820.71591.0341-0.3324-0.07820.3518>>>013210310>>>sorted,indices=torch.sort(x,>>>-1.6747-0.0782-1.4782-0.35180.06100.47631.03410.71591.4137>>>021202110torch.topk(input,k,dim=None,largest=True,sorted=True,out=None)->(Tensor,LongTensor)dim中kdim，則largestFalsek(values,indices)indices中測sorted為_True_k個值被排序。kint)“top-k”dimint,optional)–largestbool,optional)–sortedbool,optional)–outtuple,optional)–(TensorLongTensoroutput>>>x=torch.arange(1,>>>>>>torch.topk(x,3)>>>torch.topk(x,3,0,largest=False)Othertorch.cross(input,other,dim=-1,out=None)→dim和other的向量積（叉積）other必須有相dimsize3。otherTensor)–dimint,optional)–outTensor,optional)–>>>a=torch.randn(4,>>>0.22860.4446-0.04760.23210.61991.1924->>>b=torch.randn(4,>>>-0.1042-1.11560.99470.1149-1.01080.8319-0.9045-1.3754>>>torch.cross(a,b,-0.96190.20090.2696-0.6318--1.6805-2.01710.0163-1.5304->>>torch.cross(a,-0.96190.20090.2696-0.6318--1.6805-2.01710.0163-1.5304-torch.diag(input,diagonal=0,out=None)→diagonal0,diagonal0,diagonal0,diagonal(int,optional)––>>>a=>>>>>>1.04800.00000.0000-2.34050.00000.0000->>>torch.diag(a,0.00001.04800.00000.00000.0000-2.34050.00000.00000.0000-0.00000.00000.0000>>>a=torch.randn(3,>>>-1.5328-1.3210-0.85960.0471--0.66170.0146->>>torch.diag(a,>>>torch.diag(a,torch.histc(input,bins=100,min=0,max=0,out=None)→minmaxrangebins個直條，然后將排序好的數(shù)據(jù)劃分到各個直條(bins)minmax0,則利用數(shù)據(jù)中的最大最小值作為邊界。binsint)–bins(直條)的個數(shù)(100個minint)range的下邊界(包含maxint)range的上邊界(包含–>>>torch.histc(torch.FloatTensor([1,2,1]),bins=4,min=0,max=3)FloatTensor([0,2,1,0])torch.renorm(input,p,dim,maxnorm,out=None)→dimpmaxnorm。如果pmaxnorm，則當前子張量不需要修改。注意:torch7Hintonetal.2012,p.pfloat)–dimint)–maxnormfloat)––>>>x=torch.ones(3,>>>>>>>>>112233>>>torch.renorm(x,1,0,1.00001.00001.66671.66671.66671.6667torch.trace(input)→>>>x=torch.arange(1,10).view(3,>>>124578>>>torch.trace(x)torch.tril(input,k=0,out=None)→out，包含輸入矩陣(2D張量)的下三角部分，out0。這里所說的下三diagonal之上的元素。k控制對角線k0k0k0kint,optional)––>>>a=>>>1.32251.73041.24690.0064->>>1.32250.0000-0.3052-0.31111.24690.0064->>>torch.tril(a,1.32251.73041.24690.0064->>>torch.tril(a,k=-0.00000.0000-0.30520.00001.24690.0064torch.triu(input,k=0,out=None)→diagonal之上的元素。k控制對角線k0k0k0kint,optional)––>>>a=>>>1.32251.73041.24690.0064->>>1.32251.73040.00000.0000->>>torch.triu(a,0.00001.73040.00000.0000-0.00000.0000>>>torch.triu(a,k=-1.32251.73040.00000.0064-BLASandLAPACKtorch.addbmm(beta=1,mat,alpha=1,batch1,batch2,out=None)→batch1batch2reducedadd步驟(所有矩陣乘結(jié)果沿著第一維相加)matbatch1batch23batch1b×n×m的張量，batch1b×m×poutmat的形狀n×pres=(beta?M)+(alpha?sum(batch1i@batch2i,i=0,b))FloatTensorDoubleTensor的輸入，alphaandbeta必須為實數(shù)，否則兩個參數(shù)須為beta(Numberoptional)–matmatTensor)–alpha(Numberoptional)–batch1@batch2batch1@batch2batch1Tensor)–batch2Tensor)––輸出張量>>>M=torch.randn(3,>>>batch1=torch.randn(10,3,>>>batch2=torch.randn(10,4,>>>torch.addbmm(M,batch1,-3.1162 0.1824- 0.4589-0.5641--9.3387-0.1794-1.2318-6.8841-torch.addmm(beta=1,mat,alpha=1,mat1,mat2,out=None)→mat1mat2matmat1n×m張量，mat2m×poutmatn×palphabeta分別是兩個矩mat1@mat2mat的比例因子，F(xiàn)loatTensorDoubleTensor的輸入，betaandalpha必須為實數(shù)，否則兩個參數(shù)須為beta(Numberoptional)–matmatTensor)–alpha(Numberoptional)–mat1@mat2mat1@mat2mat1Tensor)–mat2Tensor)––>>>M=torch.randn(2,>>>mat1=torch.randn(2,>>>mat2=torch.randn(3,>>>torch.addmm(M,mat1,-0.4095-1.97035.7674-4.9760torch.addmv(beta=1,tensor,alpha=1,mat,vec,out=None)→矩陣，vecmmoutmatn_alpha_beta分別mat?vec和matout=(beta?tensor)+(alpha?(mat@vec))beta(Numberoptional)–matmatTensor)–alpha(Numberoptional)–mat1@vecmatTensor)–vecTensor)––>>>M=>>>mat=torch.randn(2,>>>vec=>>>torch.addmv(M,mat,torch.addr(beta=1,mat,alpha=1,vec1,vec2,out=None)→vec2mmatn×m_beta_alpha分別是兩個matmatvec1@vec2的比例因子，即，beta(Numberoptional)–matmatTensor)–alpha(Numberoptional)–vec1，vec2vec1，vec2vec1Tensor)–vec2Tensor)––>>>vec1=torch.arange(1,>>>vec2=torch.arange(1,>>>M=torch.zeros(3,>>>torch.addr(M,vec1,123torch.baddbmm(beta=1,mat,alpha=1,batch1,batch2,out=None)→batch1batch2matbatch1batch23batch1b×n×m的張量，batch1b×m×poutmatn×p，beta(Numberoptional)–matmatTensor)–alpha(Numberoptional)–batch1@batch2batch1@batch2batch1Tensor)–batch2Tensor)––>>>M=torch.randn(10,3,>>>batch1=torch.randn(10,3,>>>batch2=torch.randn(10,4,>>>torch.baddbmm(M,batch1,batch2).size()torch.Size([10,3,5])torch.bmm(batch1,batch2,out=None)→batch1batch2內(nèi)的矩陣進行批矩陣乘操作。batch1batch2都為包含相同outmatn×p，F(xiàn)loatTensorDoubleTensor的輸入，alphaandbeta必須為實數(shù)，否則兩個參數(shù)須為batch1Tensor)–batch2Tensor)––>>>batch1=torch.randn(10,3,>>>batch2=torch.randn(10,4,>>>res=torch.bmm(batch1,>>>res.size()torch.Size([10,3,5])LUpivotsinfominibatch樣本進行分解。infoarefromdgetrfandanon-zerovalueindicatesanerroroccurred.CUDA的話，這CUBLASLAPACK。ATensor)–>>>A=torch.randn(2,3,>>>A_LU=torch.btrisolve(b,LU_data,LU_pivots)→Ax=bLUbTensorRHS張量LU_data(Tensor)–PivotedLUfactorizationofAfromLU_pivotsIntTensorLU>>>A=torch.randn(2,3,>>>b=torch.randn(2,>>>A_LU=>>>x=>>>torch.norm(A.bmm(x.unsqueeze(2))-b)torch.dot(tensor1,tensor2)→aTensor)–(–outtuple,optional)–輸出元組返回值：元組，包括：eTensor)avTensor):eigenvectorsTrue，則為包含特征向量的張量;否則為空張量返回值類型：(Tensor,Tensor)torch.gels(B,A,out=None)→m×nam>=n,gels對最小二乘minimize∥X∥FsubjectXXnn行包含解。余下的行包含以下殘差信息:n行開始計算的每列的歐(mBTensor)–ATensorm×nm×nouttuple,optional)–輸出元組返回值：元組，包括：XTensor):qrTensor)QR分解的細節(jié) [2,3, [3,5, [4,2, [5,4,>>>B=torch.Tensor([[-10,- [12, [14, [16, [18,>>>X,_=torch.gels(B,>>>2.00001.00001.0000QRQ,RLAPACKRatherthisdirectlycallstheunderlyingLAPACKfunction?geqrfwhichproducesasequenceof‘elementary–outtuple,optional)–(Tensortorch.ger(vec1,vec2,out=None)→vec1,vec2vec1n,vec2moutnxm的矩陣。vec1Tensor1Dvec2Tensor1Douttuple,optional)–>>>v1=torch.arange(1,>>>v2=torch.arange(1,>>>torch.ger(v1, 8m×mXm×kBTensorm×km×kATensorm×mm×m>>>A=torch.Tensor([[6.80,-2.11,5.66,5.97, [-6.05,-3.30,5.36,-4.44, [-0.45,2.58,-2.70,0.27, [8.32,2.71,4.35,-7.17, [-9.67,-5.14,-7.26,6.08,->>>B=torch.Tensor([[4.02,6.19,-8.22,-7.57,- [-1.56,4.00,-8.67,1.75, [9.81,-4.09,-4.57,-8.61,>>>X,LU=torch.gesv(B,>>>torch.dist(B,torch.mm(A,X))Irrespectiveoftheoriginalstrides,thereturnedmatrixwillbetransposed,i.e.withstridesm)insteadof(m,–2–輸出張量>>>x=torch.rand(10,>>>0.78000.22670.78550.94790.59140.71190.44370.91310.12890.00450.04250.22290.46260.62100.02070.63380.70670.63810.83500.78100.85260.93640.75040.27370.06940.58990.85160.62800.60160.53570.29360.78270.27720.07440.26270.63260.78970.02260.31020.01980.94150.98960.35280.93970.20740.52350.61190.65220.33990.32050.55550.84540.37920.49270.10480.03280.57340.63180.98020.44580.09790.33200.37010.26160.34850.43700.56200.52910.82950.76930.18070.06500.16550.21920.69130.00930.01780.30640.67150.51010.25610.43700.46950.83330.11800.42660.41610.06990.42630.8865>>>x=torch.rand(10,>>>y=>>>z=torch.mm(x,>>>1.00000.00000.0000-0.00000.00000.00000.00000.0000-0.0000-0.00001.0000-0.00000.00000.00000.0000-0.0000-0.0000-0.0000-0.00000.00001.0000-0.0000-0.00000.00000.00000.0000-0.0000-0.00000.00000.00001.00000.00000.00000.0000-0.0000-0.00000.00000.0000-0.0000-0.00001.00000.00000.0000-0.0000-0.0000-0.00000.00000.0000-0.00000.00001.0000-0.0000-0.0000-0.0000-0.00000.00000.0000-0.00000.00000.00001.00000.0000-0.00000.00000.0000-0.0000-0.00000.00000.0000-0.00001.0000-0.0000-0.00000.0000-0.0000-0.00000.00000.0000-0.0000-0.00001.0000--0.00000.0000-0.0000-0.0000-0.00000.0000-0.0000-0.00000.0000>>>torch.max(torch.abs(z-torch.eye(10)))#Maxnonzerotorch.mm(mat1,mat2,out=None)→mat1mat2mat1n×m張量，mat2m×p張n×pout。mat1Tensor)–mat2Tensor)––輸出張量>>>mat1=torch.randn(2,>>>mat2=torch.randn(3,0.0519-0.33044.3910-5.1498torch.mv(mat,vec,out=None)→matvecmatn×m張量，vecm1維張量，將n元1維張量。matTensor)–vecTensor)––輸出張量>>>mat=torch.randn(2,>>>vec=>>>torch.mv(mat,QRq,rx=q?rqr是一個Irrespectiveoftheoriginalstrides,thereturnedmatrixqwillbetransposed,i.e.withstrides(1,m)insteadof(m,1).–2outtuple,optional)–tupleQ>>>a=torch.Tensor([[12,-51,4],[6,167,-68],[-4,24,->>>q,r=>>>-0.85710.3943-0.4286-0.9029-0.2857-0.1714>>>-14.0000- 0.0000- 0.0000->>>torch.mm(q,12- 6167- 24-1-0-0100U,S,V=torch.svd(A)。返回對形如n×mAA=USV'?Un×n，Sn×m，Vm×msomesome=Trueitcomputessomeandsome=Falsecomputesall.Irrespectiveoftheoriginalstrides,thereturnedmatrixUwillbetransposed,i.e.withstrides(1,n)insteadof(n,1).–2some(bool,optional)–outtuple,optional)–>>>a=torch.Tensor([[8.79,6.11,-9.15,9.57,-3.49, [9.93,6.91,-7.93,1.64,4.02, [9.83,5.04,4.86,8.83,9.80,- [5.45,-0.27,4.85,0.74,10.00,- [3.16,7.98,3.01,5.80,4.27,->>> 5.0400- -9.1500- 9.8000 0.1500-8.9900-6.0200->>>u,s,v=>>>-0.59110.26320.35540.3143-0.39760.2438-0.2224-0.7535--0.0335-0.6003-0.45080.2334--0.42970.2362-0.68590.3319-0.4697-0.35090.38740.1587-