Question about mirror parameter in caffe

[Yunzhi]

Hi guys, I'm just new to Caffe. And I'm wondering what the function of the mirror parameter is. It seems that Caffe just randomly flips the input images. So the volume of the training set is not expanded, am I right?  Is there anybody who can give me a definite answer? Thanks in advance.  

[Przemek D]

mirror: true means that Caffe will randomly flip the images horizontally - see the relevant lines in data_transformer.cpp here and here. Depending on what you mean by the training set being expanded: no, the actual volume of the set stays the same - the transformation is done on the fly, after the data has been loaded into memory - but yes the network will see more examples than it normally would have. It is a disk-space-efficient method of data augmentation, as you only store each example once but produce multiple versions of it during the training.