flops_percentage unaccurate

[PatGlan]

Hello,
I just tried to reimplemented your Method, but it turns out, that the resulting total pruning-ratio of 91% ("total percentage: 91.200"), although I set the parameter "pruning_flops_percentage=0.5". Do you know, why the resulting-prune-ratio is that appart from the goal of 50%?
What do I have to change to end up with the goal-ratio?

Here are my trainings-settings:
I used all parameter as described in the readme for your pruning-procedure, except that I changed the dataloader.
For simplicity I replaced the dataloader of Imagenet by the standard Dataloader of CIFAR100 of pytorch, but rescaled the image to size 224. Since, I haven't got a pretrained model for CIFAR100, I did not used a pretrained model (flag: --fine_tune).

Can you help me with that problem?

best regards
Patrick