Compulsory Pre-Processing required on images while fine-tuning AlexNet trained on imagenet data???

[Param Rajpura]

I am trying to fine-tune alexnet for object classification problem.
I need to clarify the following doubts....

1. I use IMAGE_DATA type data layer to feed images for training. Is any pre-processing compulsory on those images...like normalization,channel swap etc.
    Does caffe do any default-preprocessing??

2. After training(fine-tuning) the model, if we use the python for prediction and testing images, the example suggests using channel swap to BGR, raw scale=255,    etc. Why is it required???

3. Is this specific to python or the model?? Do imagenet images trained had any standard pre-processing done on them????

While experimenting, i tried testing my finetuned model using

1. Command line ie caffe test -model ..... -weights ...... -iterations
2. Using python net.predict.(I read about the oversampling as ON by default...i could understand that..but still not clear about other steps)

While using command line , the results are as expected but with python as we change the pre-processing the prediction is improper.....

Happy Brewing!!!!