Which one of mean.binaryproto files to use when finetuning?

[Bumby]

When finetuning a pre-trained model on a new dataset, should I use the mean file that comes with the pre-trained model or should I generate my own mean file from the new dataset? What would be the pros and cons for each choice?