I found that using Inception scale and aspect ratio augmentation can boost the performance of ResNet by ~1% (http://torch.ch/blog/2016/02/04/resnets.html).
Is there any way to do the exact same scale and aspect ratio augmentation using ImageRecordIter?
The tutorial to generate RecordIO resizes the shorter side to 480 pixels, how can I sample a batch that is 8%-100% of the original image size? And the max_aspect_ratio parameter in ImageRecordIter only produces a random value in [1 - max_aspect_ratio, 1 + max_aspect_ratio] instead of [3/4, 4/3] proposed in the first inception paper.
Thanks in advance for your help.