Pytoch uses torchnet Classerrormeter in meter

Use torchnet.meter Medium ClassErrorMeter seek Top-x Score

 

from torchnet.meter import ClassErrorMeter
# Other omitted 

metric = [ClassErrorMeter([1,2], True)]  # Calculation top-1  top-2  ACC

take metric Transfer to training class

Call in the training function or verification function metric

 if self.metric is not None:  # every last epoch Reset , Calculate each epoch The accumulation of acc
       self.metric[0].reset()   


if self.metric is not None:
      prob     = F.softmax(outputs, dim=1).data.cpu()
      self.metric[0].add(prob, labels.data.cpu())     # Add to metric[0]

  Print out

 if i == len(self.train_data_loader) - 1 and self.metric is not None:
                    top1_acc_score = self.metric[0].value()[0]
                    top2_acc_score = self.metric[0].value()[1]

Accumulating top-1 acc You can also use the following methods

# stay for Define before the loop 
 presum = 0
 for i, (inputs, labels) in enumerate(self.train_data_loader):  
     #....

    prob = F.softmax(outputs, dim=1).data.cpu()
    pre=torch.argmax(prob, 1)
    a=(pre==labels.data.cpu()).int()
    presum+=a.sum().numpy()
    acc=100*presum/len(train_datasets)