Accuracy difference between Data Layer and ImageData Layer

[Ark]

Hi,

I was just comparing the Data Layer and ImageData Layer performance using CIFAR-10 Quick model.

I could see there is not much difference in time, but there is a significant change in the accuracy. Data Layer accuracy I obtained is ~75% after 10K iterations while it is ~56% for ImageData Layer with  same number of iterations.

I am using the same model file except Data Layer replaced with ImageData Layer for training (shuffle is ON). Remaining everything is same.

Any idea why this difference in accuracy occurs?

Best Regards,
Ark.