What data set did you use to train the InceptionV3 model? about dataset

[gkrasin]

@xinchunl the classifier was trained on machine image-level annotations available for download from the main page. The machine annotation is an output of another classifier internal to Google.

It's expected that there will be more annotations (including human rated) in both train and validation sets eventually.

The unbalanced story is trivial for the classifier released: we did nothing. That's why many of the rare labels don't fire much and the output needs to be calibrated to get at least some reasonable values on them. We believe that there are multiple possible approaches are viable to handle this, you're welcome to try and report the results. :)

[xinchunl]

Thanks Ivan for your explanation.