chainer: Evaluate model performance when using chainer.links.BatchNormalization togather with chainer.config.train=False bug report
chainer version: 3.4.0 cupy version: 2.4.0
When I am using resnet-101 to classify images. The resnet-101 model have L.BatchNormalization layer.
After I trained over, I load model parameter and use chainer.using_config('train', False) to evaluate the model performance.
I am surprised that even I am using train dataset(not validation dataset), the accuracy is lower than When I was in training procedure observation ( it was 99% in the final iteration of training). It was only 80% in train split dataset.
I think the L.BatchNormalization has bug when you set chainer.using_config('train', False) to evaluate the pretrained model.
About this issue
- Original URL
- State: closed
- Created 6 years ago
- Comments: 18 (8 by maintainers)
By the way, bad experiment results are not other people’s faults. Please be polite to people who are helping you. It’s your own responsibility to do good research, understand and analyze the experiment results. At least you should read the paper I linked carefully and probably discuss or cite it in your paper.
Deadline is not the reason to be rude.