教程如何從tensorflow轉入pytorch

選自作者：IllarionPyTorch時，沒幾天就放棄了TensorFlowPyTorch的核心要領但是隨后不PyTorch論壇，你可以在其中詢問相關的問題，并很快得到回復：/API，所以有時候你可以把它用作是NumPy的替代品PyTorch的開發(fā)者們這么做的原因是希望這種框架可以完全獲得importtorchimportnumpyasnpnumpy_tensor=np.random.randn(10,#convertnumpyarraytopytorcharraypytorch_tensor=torch.Tensor(numpy_tensor)#oranotherwaypytorch_tensor=#converttorchtensortonumpyrepresentation#ifwewanttousetensoronGPUprovideanothertypedtype=torch.cuda.FloatTensorgpu_tensor=torch.randn(10,20).type(dtype)#orjustcall`cuda()`methodgpu_tensor=pytorch_tensor.cuda()#callbacktotheCPUcpu_tensor=#definepytorchtensorsx=torch.randn(10,20)y=torch.ones(20,5)#`@`meanmatrixmultiplicationfrompython3.5,PEP-0465res=x@y#gettheres.shape#torch.Size([10,的importfromtorch.autogradimport#defineanx_tensor=torch.randn(10,20)y_tensor=torch.randn(10,5)x=Variable(x_tensor,requires_grad=False)y=Variable(y_tensor,requires_grad=False)#definesomeweightsw=Variable(torch.randn(20,5),#getvariabletensorprint(type(w.data))#torch.FloatTensor#getvariablegradientprint(w.grad)#loss=torch.mean((y-x@w)**#calculatethegradientsprint(w.grad)#somegradients#manuallyapplygradientsw.data-=0.01*w.grad.data#manuallyzerogradientsafterupdateimportfromtorch.autogradimportVariableimporttorch.nn.functionalasFx=Variable(torch.randn(10,20),requires_grad=False)y=Variable(torch.randn(10,3),requires_grad=False)#definesomeweightsw1=Variable(torch.randn(20,5),requires_grad=True)w2=Variable(torch.randn(5,3),requires_grad=True)learning_rate=loss_fn=optimizer=torch.optim.SGD([w1,w2],lr=learning_rate)forstepinrange(5):pred=F.sigmoid(x@w1)pred=F.sigmoid(pred@w2)loss=loss_fn(pred,y)#manuallyzeroallpreviousgradients#calculatenewgradients#applynewgradientsPyTorchTensorFlow的另一個主要區(qū)別在于其不同的計算圖表現形式TensorFlow使用靜態(tài)圖，這意味著我們是先定動態(tài)圖會在你希望調試代碼，或定義一些條件語句時顯現出自己的優(yōu)勢debugger一樣！importtensorflowasfirst_counter=tf.constant(0)second_counter=tf.constant(10)some_value=tf.Variable(15)#conditionshouldhandlealldefcond(first_counter,second_counter,*args):returnfirst_counter<second_counterdefbody(first_counter,second_counter,some_value):first_counter=tf.add(first_counter,2)second_counter=tf.add(second_counter,1)returnfirst_counter,second_counter,some_valuec1,c2,val=cond,body,[first_counter,second_counter,withtf.Session()assess:counter_1_res,counter_2_res=sess.run([c1,c2])importfirst_counter=torch.Tensor([0])second_counter=torch.Tensor([10])some_value=torch.Tensor(15)while(first_counter<second_counter)[0]:first_counter+=2second_counter+=Keras的即開即用的層構造函數：以保存內部狀態(tài)，例如包含可學習參數的變量nn包還定義了一組在訓練神經網絡時常用的損失函數fromcollectionsimportOrderedDictimporttorch.nnasnn#ExampleofusingSequentialmodel=nn.Sequential(nn.Conv2d(1,20,5),nn.Conv2d(20,64,5),)#ExampleofusingSequentialwithOrderedDictmodel=nn.Sequential(OrderedDict([('conv1',nn.Conv2d(1,20,('relu1',('conv2',nn.Conv2d(20,64,('relu2',output=fromtorchimportnnclassModel(nn.Module): init(self):super().init()self.feature_extractor=nn.Conv2d(3,12,kernel_size=3,padding=1,stride=1),nn.Conv2d(12,24,kernel_size=3,padding=1,stride=1),)self.second_extractor=24,36,kernel_size=3,padding=1,stride=1)defforward(self,x):x=self.feature_extractor(x)x=self.second_extractor(x)#notethatwemaycallsamelayertwiceormodex=self.second_extractor(x)returnimportclassMyFunction(torch.autograd.Function):defforward(ctx,input):output=torch.sign(input)returnoutputdefbackward(ctx,#savedtensors-tupleoftensors,soweneedgetfirstinput,=ctx.saved_variablesgrad_output[input.ge(1)]=grad_output[input.le(-1)]=0returngrad_output#x=torch.randn(10,20)y=MyFunction.apply(x)#ormy_func=MyFunction.applyy=my_func(x)#andifwewanttouseinsidenn.ModuleclassMyFunctionModule(torch.nn.Module):defforward(self,returnward(input)的形式，并以input,_=self.saved_tensors的方式接入import###tensorx_cpu=torch.randn(10,20)w_cpu=torch.randn(20,10)#directtransfertotheGPUx_gpu=x_cpu.cuda()w_gpu=w_cpu.cuda()result_gpu=x_gpu@w_gpu#getbackfromGPUtoCPUresult_cpu=###modelexamplemodel=model.cuda()#trainstepinputs=Variable(inputs.cuda())outputs=model(inputs)#getbackfromGPUtoCPUoutputs=outputs.cpu()class init(self,model,use_cuda=False,self.use_cuda=use_cudaself.gpu_idx=gpu_idxself.model=self.to_gpu(model)defto_gpu(self,tensor):ifself.use_cuda:returntensor.cuda(self.gpu_idx)returndeffrom_gpu(self,tensor):ifself.use_cuda:returntensor.cpu()returndeftrain(self,inputs=self.to_gpu(inputs)outputs=self.model(inputs)outputs=self.from_gpu(outputs)改變張量的權重權重可以用調用torch.nn.init包中的多種方法初始化為直接訪問張量的屬性這個決定或許并不直接了importfromtorch.autogradimport#newwaywith`init`modulew=torch.Tensor(3,5)#workforVariablesalsow2=Variable(w)#oldstyleddirectaccesstotensorsdataattribute#exampleforsomemoduledefweights_init(m):classname=m.class.ifclassname.find('Conv')!=-1:m.weight.data.normal_(0.0,0.02)elifclassname.find('BatchNorm')!=-1:m.weight.data.normal_(1.0,0.02)#forloopapproachwithdirectaccessclassMyModel(nn.Module): initforminifisinstance(m,n=m.kernel_size[0]*m.kernel_size[1]*m.out_channelsm.weight.data.normal_(0,math.sqrt(2./n))elifisinstance(m,nn.BatchNorm2d):elifisinstance(m,nn.Linear):PyTorchflag：requires_gradvolatile第一個可以禁用當前層的梯度，但子節(jié)點仍然可以計算第二個可以importfromtorch.autogradimport#requires#Ifthere’sasingleinputtoanoperationthatrequiresgradient,#itsoutputwillalsorequiregradient.x=Variable(torch.randn(5,y=Variable(torch.randn(5,z=Variable(torch.randn(5,5),requires_grad=True)a=x+ya.requires_grad#Falseb=a+zb.requires_grad#True#Volatilediffersfromrequires_gradinhowtheflagpropagates.#Ifthere’sevenasinglevolatileinputtoanoperation,#itsoutputisalsogoingtobex=Variable(torch.randn(5,5),requires_grad=True)y=Variable(torch.randn(5,5),volatile=True)a=x+ya.requires_grad#False允許/禁止批規(guī)范層和dropout如果你想要做的話，讓CPU和GPU的隨機算子不同也是可以的#schedulerfromtorch.optimimportoptimizer=torch.optim.SGD(model.parameters(),scheduler=lr_scheduler.StepLR(optimizer,step_size=30,gamma=0.1)forepochinrange(100):#Trainflagcanbeupdatedwith#todisabledropoutandbatchnormlearning#executetrainstep#runinference#CPUseed#GPUseedfromcollectionsimportOrderedDictimporttorch.nnasnnmodelnn.Sequential(OrderedDict([('conv1',nn.Conv2d(1,20,5)),('relu1',('conv2',nn.Conv2d(20,64,('relu2',#Sequential (conv1):Conv2d(1,20,kernel_size=(5,5),stride=(1,1)) (relu1):ReLU() (conv2):Conv2d(20,64,kernel_size=(5,5),stride=(1,1)) (relu2):ReLU()##save/loadonlythemodelparameters(preferedsolution)torch.save(model.state_dict(),save_path)#savewholemodeltorch.save(model,save_path)model=torch.load(save_path)文本記錄Python了，你也可以試試一些第三方庫：logger：/oval-group/loggerCrayon：/torrvision/crayontensorboard_logger：pytorch：/lanpa/tensorboard-pytorchVisdom：importimporttorchvisionasclassImagesDataset(torch.utils.data.Dataset): init(self,df,transform=None,self.df=dfself.transform=transformself.loader=loader getitem(self,index):row=self.df.iloc[index]target=path=img=ifself.transformisnotNone:img=self.transform(img)returnimg, lenn,_=self.df.shapereturnn#whattransformationsshouldbedonewithourimagesdata_transforms=tv.transforms.Compose([tv.transforms.RandomCrop((64,64),padding=4),train_df=pd.read_csv('path/to/some.csv')#initializeourdatasetatf