Requesting lmdb support in extract_features.cpp

[mgh1]

As lmdb is being used widely in caffe, including tools such as compute_image_mean.cpp, convert_imageset.cpp, it would be great to also extend lmdb support to extract_features.cpp.

This will improve tooling consistency as well as performance in scenarios which use the results of extract_features for various tasks which would benefit from the speed of lmdb.

Thank you for your consideration.

[kloudkl]

To easily switch between LevelDB and LMDB, the two backends should be wrapped in two classes that share the same interface or base abstract class. Otherwise, using them in more places will make the code as ugly as if not uglier than exposed by #1074 which tried to segregate each of them using guarding macros.

[kloudkl]

Maybe what you need is an output layer with the LevelDB or LMDB as the backing store similar to the HDF5OutputLayer.

[kloudkl]

#1238 is building a unified API for LevelDB and LMDB.

[shelhamer]

See #1238 but really the features can be extracted through pycaffe or matcaffe by just loading the model with its usual data layer defined.