Regarding the accuracy after pruning

[zirid]

I used your test code from the " High-level Pruners" section in the documentation:

Instead of using "resnet18" I used resnet50 on an easy classification task. the accuracy without pruning is 99% after 5 epochs but I get a very low accuracy after, unlike your data that suggests that at worst the accuracy drops only by 1%.

Note: (I didn't use the last part: # finetune the pruned model here), I used the model as it is after pruning.

My results are: pruning ratio = 0.1 gives 94% accuracy pruning ratio = 0.5 gives 65% accuracy

Do you think I missed something?

Thank you.

[lemingshen]

I also met this problem, after pruning even with a very small pruning ratio (0.001), the accuracy drops significantly (70% --> 50%). Even fine-tuning does not work.