In matcher > common > logger.py 's class AverageMeter
It seems that there is a slight issue with the AverageMeter function. During initialization, a tensor of size [2, self.nclass] is initialized using torch.zeros, which causes the performance to be divided by larger number of the entire dataset( total number of data + self.nclass). As a result, when I modified your code and conducted inference, the recorded performance was approximately doubled.
In matcher > common > logger.py 's class AverageMeter
It seems that there is a slight issue with the AverageMeter function. During initialization, a tensor of size [2, self.nclass] is initialized using torch.zeros, which causes the performance to be divided by larger number of the entire dataset( total number of data + self.nclass). As a result, when I modified your code and conducted inference, the recorded performance was approximately doubled.