Even in Resnet era, data was the most relevant for accuracy. Google had this internal JFT data, training on which improves imagenet score much larger than diffe

Even in Resnet era, data was the most relevant for accuracy. Google had this internal JFT data, training on which improves imagenet score much larger than different architecture. Also most of the gains after that came from data augmentation techniques or bigger model rather than new architecture.