Hi,
I am trying to figure out the accuracy gaps between the pytorch model accuracies compared to resnet50 paper .The published accuracies are slightly better that pytorch ones. I was wondering if that was due to changes in ImageNet datasets used for training. Did you use ImageNet 2012 training and validation datasets for your runs?
Most of the hyperparameters you posted are similar to that of resnet paper. I am curious to know if your final results are closer to either pytorch implementation or the resnet paper and if you were able to close the accuracy gap? Thanks!