Number of kernels in alexnet_benmark

[deepali-c]

I notice that the number of kernels used in the convolutional layer of the alexnet_benchmark sample are different from the number specified in the Alexnet paper. So I am just curious to know the reason for doing so?

[ozabluda]

Because it's Alexnet from the "One weird trick" paper: https://arxiv.org/abs/1404.5997. Classic AlexNet 2-column topology was caused by him using 2 GPUs with 3GB RAM each (maximum at the time). As GPUs with more RAM appeared, it was replaced by simpler and better performing 1-column Alexnet. There is really no reason to benchmark 2-column classic Alexnet, except to preserve historic continuity, which often [accidentally] happens when people run Caffenet (since it's included with Caffe.