Using tensorflow-keras to reproduce AlexNet.
Load pre-trained weights.
Benchmark model on ILSVRC-2012 validation dataset.
Using 1 crop per image when testing (compare to 5 crops)
Top-1 error rate: 0.480 (compare to 0.407 in paper)
Top-5 error rate: 0.241 (compare to 0.182 in paper)