some improvement on pair fast loss in pytorch
I reproduce pair fast loss in HDC with pytorch and make some improvement.
I use average pooling to get online confusion matrix and resample the train list.
I use resnet18 and single pair fast loss and reach 99.43% in 50sku.
Resampled list from online confusion matrix: 99.60%
Resampled list from offline confusion matrix: 99.64%
Use nn.CrossEntropy() as loss function.