Closed YeliTXK closed 4 years ago
@YeliTXK , to use cifar10, you can simple provide your own train_loader
(doc here). If you want to use a new search space based on MobileNetv2, you can use NNI NAS API to write the model.
For the question about experiment on cifar10, @ultmaster could provide more information on this. Also welcome to share your experience here.
I haven't seen many works running CIFAR10 on MobileNetV2 backbone. Haven't tried myself though.
An easy way to verify is to randomly sample a few models (e.g., 20) and train them sufficiently on CIFAR10. Then you will find out, whether it's the problem with search space, or training approach, or ProxylessNAS itself.
I haven't seen many works running CIFAR10 on MobileNetV2 backbone. Haven't tried myself though.
An easy way to verify is to randomly sample a few models (e.g., 20) and train them sufficiently on CIFAR10. Then you will find out, whether it's the problem with search space, or training approach, or ProxylessNAS itself.
Thanks a lot.
后续会添加ProxylessNASz中针对CIFAR10数据搜索的代码吗? 如果使用MobileNetv2的backbone来针对CIFAR数据集建立搜索空间,可能得到较好的效果吗? (我尝试了一下效果很差,是因为我的设置问题,还是MobileNetv2的backbone本身不适合CIFAR10数据集?) 希望得到您的回复。