In this repo, "heteroskedastic_cifar.py" contains the code for adding noisy labels in cifar datasets but only has "step"-type imbalance. I tried to use "imbalance_cifar.py" from LDAM to produce experiments on long-tailed datasets. But the training seems abnormal: when I estimate the statistics through a pre-train step using "cifar_hetero_est.py" with IF-100 CIFAR10, the accuracy has achieved 81% during the pre-training. Could you explain why it happen?
In this repo, "heteroskedastic_cifar.py" contains the code for adding noisy labels in cifar datasets but only has "step"-type imbalance. I tried to use "imbalance_cifar.py" from LDAM to produce experiments on long-tailed datasets. But the training seems abnormal: when I estimate the statistics through a pre-train step using "cifar_hetero_est.py" with IF-100 CIFAR10, the accuracy has achieved 81% during the pre-training. Could you explain why it happen?