MACHINE LEARNING BERKELEY A more practical solution Pretrain at lower resolution,finetune at higher resolution o Lower resolution training makes training cheaper/less computationally intensive o For more info,check out"Fixing the train-test resolution discrepancy' Benchmarked on ImageNet classification o Pretrained on massive dataset,then fine-tuned on actual ImageNet The large datasets include ImageNet-22k(a superset of ImageNet with 14M images and 22k classes)and JFT-300M(300M images and 18k classes) A more practical solution ● Pretrain at lower resolution, finetune at higher resolution ○ Lower resolution training makes training cheaper / less computationally intensive ○ For more info, check out “Fixing the train-test resolution discrepancy” ● Benchmarked on ImageNet classification ○ Pretrained on massive dataset, then fine-tuned on actual ImageNet ○ The large datasets include ImageNet-22k (a superset of ImageNet with 14M images and 22k classes) and JFT-300M (300M images and 18k classes)