features not matching after fc7 layer

[nithinraok]

As this repo was inspired from https://github.com/peteanderson80/bottom-up-attention. It perfectly suits for me to get pre trained features. But as in bottom-up-attention pool5_flat features are these flat features are same as fc7 in this repo? because I am getting different features for same image. One possibly reason would be, I am guessing, as the original features were trained on VisualGenome, but in this I guess only coco dataset was used. Please clear this doubt. Thanks