Lower top1/top5 accuracy on ImageNet for reference model

Issue summary

I tried to reproduce the accuracy values (top1 and top5) for the reference ImageNet model (CaffeNet). However, I only get a top5 accuracy of ~67% which is lower than the reported 80% here.

Steps to reproduce

I used the model definition and weights from GitHub. I evaluated the model using this simple script. I used the pre-processing (transforming from RGB to BGR, image resizing, center cropping, and ImageNet mean subtraction) indicated in the ImageNet examples of Caffe.

I uploaded the predicted classes and the ground-truth labels as a python list here, in case someone wants to take a look at them,

Tried solutions

I redownloaded the network weights and verified the preprocessing.

System configuration

Operating system: see below
Compiler:
CUDA version (if applicable):
CUDNN version (if applicable):
BLAS:
Python version (if using pycaffe):
MATLAB version (if using matcaffe):

I used the latest gpu docker container.

Issue checklist

[x] read the guidelines and removed the first paragraph
[x] written a short summary and detailed steps to reproduce
[x] explained how solutions to related problems failed (tick if found none)
[x] filled system configuration
[x] attached relevant logs/config files (tick if not applicable)

BVLC / caffe