Estimate of work for integrating data layer, solver, etc

[sh1r0]

Estimate of work for integrating data layer, solver, etc

[sh1r0]

I just updated the repo, now solver is compiled together with Caffe-Mobile. If you need image_data_layer only, there's no need to make data_layer.cpp included in Makefile, this android lib always supports it. (I used image_data to deploy on mobile before, and now memory_data is used.)
If you really want all those ignored stuff to get back, you probably need to cross compile not only hdf5 but also leveldb and lmdb.

BTW, I wonder why you need solvers, do you use them for training on Android?? Just out of curiosity :p

Thanks.

[jbeorse]

Thanks for the quick response and updates.

Hmm, I'm actually integrating code that was written by a partner and I don't fully know all of its inner workings. I'm not trying to train on Android, but its certainly possible I've missed some things when trying to remove that code. I'll work with that partner and make sure we've really seperated things out correctly. We still have references to some of the db dependencies too, which it seems like are only used for training.

On a side note, I'm a little confused with how you are using glog and gflags. You don't seem to be building them for Android, and in fact I see an area where you use glog_alternative.hpp, but then there are files like caffe-android-lib/caffe-mobile/jni/caffe/tools/compute_image_mean.cpp that still include <glog/logging.h>. How does this work?

[sh1r0]

In short, I avoid using glog and gflags in porting.
Tools are not used in my case, thus they are excluded. (In Android.mk, only things under src/caffe/ are included for compilation.) glog_alternative.hpp is a simple wrapper with CHECK* macros taking place of glog ones.

Thanks.

[jbeorse]

Ok, that makes sense. I'm working on creating a header that translates glog calls into native Android logger calls. If I get that working I'll share that with you.