scale issue on using pre-trained model

[a18700]

Hello, I'm writing an issue for letting others know about input scaling issue.

I was struggling with nan and memoryerror on training VOC and coco dataset both, and finally figured out that the problem was the difference on input preprocessing to given caffe model and custom models.

On my case, my custom model is trained with image data which is divided by 255, but given caffe-model is not.

I want anyone this article helps others suffering with 'Memoryerror', 'NaN'.