bender python客户端
bender-client的Python项目详细描述
用于python的bender客户端
:warning: The full DOCUMENTATION on bender-python-client can be found HERE.
设置
- 在bender.dreem.com
- 使用
pip install bender-client在python环境中安装bender
用法示例
Let's use the famous MNIST example where we try to recognize handwritten digits in images.
使用PyTorch的算法代码如下：
To use this example, do not forget to
pip install numpy torch torchvision.
from__future__importprint_functionimporttorchimporttorch.nnasnnimporttorch.nn.functionalasFimporttorch.optimasoptimfromtorchvisionimportdatasets,transformsclassNet(nn.Module):def__init__(self,dropout=True,activation="relu",kernel_size=5,conv_depth=10,linear_depth=50):super(Net,self).__init__()self.conv1=nn.Conv2d(1,conv_depth,kernel_size=kernel_size)self.conv2=nn.Conv2d(conv_depth,20,kernel_size=kernel_size)self.conv2_drop=nn.Dropout2d()ifdropoutisTrueelselambdax:xself.fc1=nn.Linear(320,linear_depth)self.fc2=nn.Linear(linear_depth,10)self.activation=getattr(F,activation)defforward(self,x):x=self.activation(F.max_pool2d(self.conv1(x),2))x=self.activation(F.max_pool2d(self.conv2_drop(self.conv2(x)),2))x=x.view(-1,320)x=self.activation(self.fc1(x))x=F.dropout(x,training=self.training)x=self.fc2(x)returnF.log_softmax(x,dim=1)deftrain(model,device,train_loader,optimizer,epoch):model.train()forbatch_idx,(data,target)inenumerate(train_loader):data,target=data.to(device),target.to(device)optimizer.zero_grad()output=model(data)loss=F.nll_loss(output,target)loss.backward()optimizer.step()ifbatch_idx%10==0:print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(epoch,batch_idx*len(data),len(train_loader.dataset),100.*batch_idx/len(train_loader),loss.item()))deftest(model,device,test_loader):model.eval()test_loss=0correct=0withtorch.no_grad():fordata,targetintest_loader:data,target=data.to(device),target.to(device)output=model(data)test_loss+=F.nll_loss(output,target,reduction='sum').item()pred=output.max(1,keepdim=True)[1]correct+=pred.eq(target.view_as(pred)).sum().item()test_loss/=len(test_loader.dataset)print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(test_loss,correct,len(test_loader.dataset),100.*correct/len(test_loader.dataset)))return(correct/len(test_loader.dataset))defrun(epochs=3,lr=0.01,momentum=0.5,dropout=True,activation="relu",kernel_size=5,conv_depth=10,linear_depth=50):torch.manual_seed(1)device=torch.device("cpu")train_loader=torch.utils.data.DataLoader(datasets.MNIST('../data',train=True,download=True,transform=transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.1307,),(0.3081,))])),batch_size=32,shuffle=True,)test_loader=torch.utils.data.DataLoader(datasets.MNIST('../data',train=False,transform=transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.1307,),(0.3081,))])),batch_size=1000,shuffle=True,)model=Net(dropout,activation).to(device)optimizer=optim.SGD(model.parameters(),lr=lr,momentum=momentum)accuracy=0forepochinrange(1,int(epochs)+1):train(model,device,train_loader,optimizer,epoch)accuracy=test(model,device,test_loader)returnaccuracyif__name__=='__main__':# HYPERPARAMETERS (That's what bender is interested in)# Here we select them on our own in an arbitrary wayhyperparameters={"kernel_size":5,"epochs":3,"lr":0.05,"momentum":0.2,"dropout":True,"activation":"relu","conv_depth":10,"linear_depth":50,}run(epochs=hyperparameters["epochs"],lr=hyperparameters["lr"],momentum=hyperparameters["momentum"],dropout=hyperparameters["dropout"],activation=hyperparameters["activation"],kernel_size=hyperparameters["kernel_size"],conv_depth=hyperparameters["conv_depth"],linear_depth=hyperparameters["linear_depth"],)
现在让我们将bender插入其中！
- 导入折弯机
frombenderclientimportBenderbender=Bender()
This will ask for your email and password. The client will use these to login and retrieve a TOKEN. This TOKEN is personal, it should not be shared, it will be stored in your home folder as "
.bender_token", and you will not be asked for your login/password again until it expires. :warning: Again, your TOKEN is personal. You should not give it or add it to any public repository :warning:
- 创建一个实验
An experiment is related to the problem you are trying to solve, here : MNIST classification
bender.new_experiment(name='MNIST Classification',description='Simple image classification on handwritten digits',metrics=[{"metric_name":"algorithm_accuracy",# It's just a name and there can be multiple watched metrics."type":"reward",# The type can either be "reward" or "loss" depending on if you want to maximize or minimize it.}],dataset='MNIST')
- 创建算法
An algo is simply corresponding to ONE solution to an Experiment problem : here it's as we saw a Neural Net with PyTorch
bender.new_algo(name='PyTorch_NN',# The parameters below are actually the Hyper-Parameters of your algo described in a listparameters=[{"name":'kernel_size',"category":"categorical","search_space":{"values":[3,5,7],},},{"name":'conv_depth',"category":"uniform","search_space":{"low":1,"high":100,"step":1,},},{"name":'linear_depth',"category":"uniform","search_space":{"low":1,"high":100,"step":1,},},{"name":'epochs',"category":"uniform","search_space":{"low":1,"high":4,"step":1,},},{"name":'lr',"category":"loguniform","search_space":{"low":1e-5,"high":1e-1,"step":1e-6,},},{"name":'momentum',"category":"uniform","search_space":{"low":0,"high":1,"step":0.05,},},{"name":'dropout',"category":"categorical","search_space":{"values":[True,False],},},{"name":'activation',"category":"categorical","search_space":{"values":["relu","softmax","sigmoid","tanh"],},},])
- 从bender获取超参数集建议
The whole goal of what we did up there is to use Bender to get a new set of Hyperparameters to try according to the settings of your Experiment and Algo.
suggestion=bender.suggest()# suggestion would for example contain something like :{"kernel_size":5,"epochs":3,"lr":0.05,"momentum":0.2,"dropout":True,"activation":"tanh","conv_depth":10,"linear_depth":50,}
- feed a trial to bender
A Trial is simply an attempt of you trying a Hyperparameters Set with your algorithm associated with the result metrics obtained. If you want bender to improve over time, feed him every trial you make.
bender.new_trial(parameters={"kernel_size":5,"epochs":3,"lr":0.05,"momentum":0.2,"dropout":True,"activation":"tanh","conv_depth":10,"linear_depth":50,},results={"algorithm_accuracy":0.7,# We put an arbitrary value here just for the example.})
- 全部代码放在一起
Psssssst... The magic starts at line 443... ;)
To use this example, do not forget to
pip install numpy torch torchvision bender-client.
from__future__importprint_functionimportargparseimporttorchimporttorch.nnasnnimporttorch.nn.functionalasFimporttorch.optimasoptimfromtorchvisionimportdatasets,transformsfrombenderclientimportBenderclassNet(nn.Module):def__init__(self,dropout=True,activation="relu",kernel_size=5,conv_depth=10,linear_depth=50):super(Net,self).__init__()self.conv1=nn.Conv2d(1,conv_depth,kernel_size=kernel_size)self.conv2=nn.Conv2d(conv_depth,20,kernel_size=kernel_size)self.conv2_drop=nn.Dropout2d()ifdropoutisTrueelselambdax:xself.fc1=nn.Linear(320,linear_depth)self.fc2=nn.Linear(linear_depth,10)self.activation=getattr(F,activation)defforward(self,x):x=self.activation(F.max_pool2d(self.conv1(x),2))x=self.activation(F.max_pool2d(self.conv2_drop(self.conv2(x)),2))x=x.view(-1,320)x=self.activation(self.fc1(x))x=F.dropout(x,training=self.training)x=self.fc2(x)returnF.log_softmax(x,dim=1)deftrain(model,device,train_loader,optimizer,epoch):model.train()forbatch_idx,(data,target)inenumerate(train_loader):data,target=data.to(device),target.to(device)optimizer.zero_grad()output=model(data)loss=F.nll_loss(output,target)loss.backward()optimizer.step()ifbatch_idx%10==0:print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(epoch,batch_idx*len(data),len(train_loader.dataset),100.*batch_idx/len(train_loader),loss.item()))deftest(model,device,test_loader):model.eval()test_loss=0correct=0withtorch.no_grad():fordata,targetintest_loader:data,target=data.to(device),target.to(device)output=model(data)test_loss+=F.nll_loss(output,target,reduction='sum').item()pred=output.max(1,keepdim=True)[1]correct+=pred.eq(target.view_as(pred)).sum().item()test_loss/=len(test_loader.dataset)print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(test_loss,correct,len(test_loader.dataset),100.*correct/len(test_loader.dataset)))return(correct/len(test_loader.dataset))defrun(epochs=3,lr=0.01,momentum=0.5,dropout=True,activation="relu",kernel_size=5,conv_depth=10,linear_depth=50):torch.manual_seed(1)device=torch.device("cpu")train_loader=torch.utils.data.DataLoader(datasets.MNIST('../data',train=True,download=True,transform=transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.1307,),(0.3081,))])),batch_size=32,shuffle=True,)test_loader=torch.utils.data.DataLoader(datasets.MNIST('../data',train=False,transform=transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.1307,),(0.3081,))])),batch_size=1000,shuffle=True,)model=Net(dropout,activation).to(device)optimizer=optim.SGD(model.parameters(),lr=lr,momentum=momentum)accuracy=0forepochinrange(1,int(epochs)+1):train(model,device,train_loader,optimizer,epoch)accuracy=test(model,device,test_loader)returnaccuracydefinit_bender():bender=Bender()bender.create_experiment(name='MNIST Classification',description='Simple image classification on handwritten digits',metrics=[{"metric_name":"algorithm_accuracy","type":"reward"}],dataset='MNIST')bender.create_algo(name='PyTorch_NN',hyperparameters=[{"name":'kernel_size',"category":"categorical","search_space":{"values":[3,5,7],},},{"name":'conv_depth',"category":"uniform","search_space":{"low":1,"high":100,"step":1,},},{"name":'linear_depth',"category":"uniform","search_space":{"low":1,"high":100,"step":1,},},{"name":'epochs',"category":"uniform","search_space":{"low":1,"high":4,"step":1,},},{"name":'lr',"category":"loguniform","search_space":{"low":1e-5,"high":1e-1,"step":1e-6,},},{"name":'momentum',"category":"uniform","search_space":{"low":0,"high":1,"step":0.05,},},{"name":'dropout',"category":"categorical","search_space":{"values":[True,False],},},{"name":'activation',"category":"categorical","search_space":{"values":["relu","softmax","sigmoid","tanh"],},},])returnbenderif__name__=='__main__':# Create experiment and algo if they don't exist yet. Else, load them from the config file ./.benderconfbender=init_bender()whileTrue:# Get a set of Hyperparameters to testsuggestion=bender.suggest(metric="algorithm_accuracy")# Get algo result with themresult=run(epochs=suggestion["epochs"],lr=suggestion["lr"],momentum=suggestion["momentum"],dropout=suggestion["dropout"],activation=suggestion["activation"],kernel_size=suggestion["kernel_size"],conv_depth=suggestion["conv_depth"],linear_depth=suggestion["linear_depth"],)# Feed Bender a Trial, AKA => suggestion + resultbender.create_trial(hyperparameters=suggestion,results={"algorithm_accuracy":result})print('New trial sent -----------------------------------------------------\n\n')
欢迎加入QQ群-->： 979659372
