fedorzh commented on issue #9185: Gluon provided ResNet does not get desirable accuracy on CIFAR10 URL: https://github.com/apache/incubator-mxnet/issues/9185#issuecomment-357552905 @zhreshold, I really appreciate you sharing the code, it helped debug a few things. It seems that the largest increase in accuracy indeed comes from using different data accessing methods. In comparison with my initial models, there are two increases in accuracy: 1. Random data modifications you have in `train_transform` give a significant boost over not having them in the `transform` function 2. Even with them, there is a difference with using `get_data_rec` instead. Unfortunately, in my use case I need a fixed training dataset rather than a dataset which changes at every epoch. Feel free to close this issue; however, I still don't understand though why there is a difference in accuracy in 2. (using `ImageRecordIter` vs `train_transform` in your code) - probably other augmentations do make the difference.
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
