Accuracy of VGG16 and Squeezenet Models very low

I downloaded the VGG16 and Squeezenet Models. I used the preprocessing steps mentioned in the page and when I ran them using caffe2 backend I got pretty low accuracy as mentioned in the site.
For VGG16, I got accuracy 54.02% where as the mentioned accuracy if 72%, For Squeezenet, I got accuracy around 34% which is also way off than whatever has been mentioned in the site.
Has anyone else faced this issue? Something I am doing wrong?
Preprocessing steps:
1- Resize (bilinear interpolation)
2 - Crop
3 - mean and stddev subtraction - (pixel[i]/255 - channel_mean)/channel_stddev
4 - Save in BGR(To be run in caffe2 backend)