Evaluation of analytics zoo vs scikit learn 4 layer CNN example

[Mahmood1968]

We took a small data sets of clothing with different colors which makes the problem multiple classes and multiple labels ( 6 labels ) using 4 CNN layers without transfer learning . In Scikitlearn code the AUC around 98% to 99%, while when the same dataset is used with transfer learning we got 87% AUC . What is the issues with analytics zoo ?

Please look at this example with dataset and can be used for testing analytics zoo
https://www.pyimagesearch.com/2018/05/07/multi-label-classification-with-keras/