Pruning increases inference time for detectron2 model

I have pruned a detectron2 model for 20%, 40%, 60% and 80% of the weights.
However, while checking the measuring the inference time for such a model, it’s performing inconsistently and surprisingly, the model with 80% weights pruned is taking more time than an unpruned model.

This is surprising because I had pruned a Resnet101 model previously and the inference time decreased consistently.

Note: The pruning parameters have been removed using prune.remove()