PocketFlow icon indicating copy to clipboard operation
PocketFlow copied to clipboard

Published 20 hours ago • verified publisher icon Tencent

ChannelPruner evaluate get unstable accuarcy

Open xiaomr opened this issue 6 years ago • 1 comments
trafficstars

I tried to evaluate a compressed model generated by channel prune learner, however the accuracy is unstable and lower than full precision evaluation. My resnet56_cifar10 checkpoint accuracy varies such as 0.9217, 0.9198, 0.9213, the same checkpoint get 0.93 accuracy in full precision eval mode.

xiaomr avatar Feb 20 '19 02:02 xiaomr

It seems like something in the evaluation mode of ChannelPrunedLearner works unexpectedly. We are looking into this issue. Thanks for your feedback.

jiaxiang-wu avatar Feb 25 '19 00:02 jiaxiang-wu