how to calculate the speedup in the paper

when I run the compressed VGG16 model, the runtime is even longer than VGG16 before compressed. how to run the compressed models getting the same results (Table 5) in the paper https://arxiv.org/abs/1903.09973?