Is timm/utils/metrics.py reliable for completely imbalanced datasets?

[bryanpiguave]

I am currently working on long-tailed recognition. Everything seems to be ok when I trained the model. The model obtained a 91% accuracy in top1 in the validation set. However, when I created the confusion matrix, the weighted average accuracy was 72%. It seems to me that the accuracy does not consider a weighted accuracy, it is calculated in terms of the batch and it is gradually updated. Should I consider the implemented function as reliable for my study case?

[smv620]

does this repo have an example of how to build and display the confusion matrix after train.py is run?