Gluon Model Zoo Accuracy Problem


#1

Hi.

I have a copy of ImageNet loading into .rec files by im2rec.py from MXNet ~v.12. I’m trying to clone the models from the gluon model zoo and test their accuracy based on the data in the image rec file using mx.io.ImageRecordIter based on the example here https://gluon-cv.mxnet.io/build/examples_classification/dive_deep_imagenet.html#validation

I can’t for the life of me get the pretrained models to give accurate results. The gluon models seem to want to push all of the examples into a set of ~20 categories. I’m wondering if it has to do with the data transformations the pretrained models used (i.e., here, https://github.com/apache/incubator-mxnet/blob/master/example/gluon/data.py).

Any ideas on making these models work with a .rec file?

Thanks,
Matthew


#2

I figured it out, the Gluon models were trained with images scaled by 255, doing the same thing in ImageRecordIter gets me the expected accuracy.