Caffe accuracy is different from the confusion matrix ?

[Shisho Sama]

I train a cifar10 model in caffe and got 92% accuracy. 

then I tried the confusion matrix on the test data and the total accuracy was like  87.4% ! 

What is happening here. 

Isn't it that in the test phase, the whole test set is presented to the network and the average of the accuracies is calculated ? so this should give the same result as in confusion matrix right?

'm confused, Any help is appreciated