There have been multiple issues regarding transformations not being applied properly in memory data layer and Ignored in HDF5 Data layer. I have also noticed that there is a separate transform function for

• Datum (used by lmdb backend)
• OpenCV Mat ( used by memory data layer - > addMatVector )
• Caffe Blobs ( which, i think, can be used along with hdf5 data layer )

Is there a possibility in combining everything by adding Transformation as a layer?
This will enable us to apply transformations to in-memory 4D array using OpenCV, Magick++, etc. with just a reference to the memory blob ( OpenCV also supports GPU ops (using CUDA and OpenCL which can help batch processing ).

Thoughts?