search

liuzechun / MetaPruning

MetaPruning: Meta Learning for Automatic Neural Network Channel Pruning. In ICCV 2019.
MIT License
351 stars 74 forks source link

Why the accuracy is low in the start of training about mobilenetV2 part? #5

Closed YANGWAGN closed 5 years ago

YANGWAGN commented 5 years ago

During the training, I only changed the batch size from the default 512 to 256, but the Acc@1 is only about 5% and the Acc@5 is only about 15% before the 5th epoch. Do you encounter this situation during the training? Thank you !

liuzechun commented 5 years ago

Low training accuracy is absolutely normal in training the PruningNet. We also experienced that. Thinking that the PruningNet need to predict weights for various different Pruned Networks while the weight in a normal network only need to suit themselves to a single network, it is understandable for PruningNet achieving not as high accuracy as we expect for a normal network.