Hi,
Thanks for sharing your implementation. I notice that the val1/val2 splits are used for training and testing on ImageNet follow the RCNN.
I would ask do you use the images from train set? Because rbgirshick mentioned that images from val1 and 1K images per class from train are both utilized.