Caffe linking to caffe CPU only build (without GPU support)

[bisher3]

Hello,
I wanted to use javacpp presets for caffe on machines without a GPU. I had already successfully built and ran caffe in CPU only mode and now I'm trying to access it from java but the javacpp-presets downloads and links to a pre-built caffe with GPU support.
First I kept getting this error:

Caused by: java.lang.UnsatisfiedLinkError: /private/var/folders/g3/nyry1nsx6lg0vjb5nkbw7jt40000gn/T/javacpp143953538120369/libjnicaffe.dylib: dlopen(/private/var/folders/g3/nyry1nsx6lg0vjb5nkbw7jt40000gn/T/javacpp143953538120369/libjnicaffe.dylib, 1): Library not loaded: @rpath/libcudart.7.5.dylib

I then installed CUDA on the device and this error disappeared.
Now I get

F0513 19:49:44.501734 2203648 internal_thread.cpp:26] Check failed: error == cudaSuccess (35 vs. 0)  CUDA driver version is insufficient for CUDA runtime version.

This is obviously because the machine doesn't contain a GPU and the nVidia Cuda driver won't work on it.
Is there a way to link to the CPU only caffe built libraries ??
Thanks,
Bisher

[saudet]

Sure, you can make a build without CUDA support. It's probably easier to simply not use CUDA at runtime though.

[bisher3]

I set -gpu=-1 in the command line and solver_mode: CPU in the solver.prototxt file but still got this error.
Is there any other way to disable CUDA usage at runtime?

[saudet]

Something like that should do it yes. IIRC, works fine here.

[saudet]

Anyway, if it doesn't for you, please report it upstream! Thanks

[bisher3]

Sorry to reopen issue, but the compilation isn't working correctly.

Please notice that if you build caffe with GPU support and then disable it at runtime only, it will still search for CUDA and CUDA driver but not use them. What I want to achieve is get a caffe build that doesn't search for the CUDA driver and can work on a machine without any GPU at all. (such as old macbook)

These are the steps I tried to recompile the javacpp-presets for caffe WITHOUT GPU support (CPU_ONLY=1):

1. Pull javacpp and javacpp-presets from github
2. Build javacpp correctly using maven. The javacpp-1.2-SNAPSHOT.jar is correctly deployed to Maven repository
3. Modify cppbuild.sh in javacpp-presets/caffe
   on line 30:
   macosx-*) export CPU_ONLY=1
   and also line 145
   make -j $MAKEJ BLAS=$BLAS BLAS_INCLUDE=/usr/local/opt/openblas/include BLAS_LIB=/usr/local/opt/openblas/lib OPENCV_VERSION=3 DISTRIBUTE_DIR=.. # lib
   This is to use the openBlas libraries in my mac instead of default ones. OpenBlas is faster
4. Run : mvn clean install --projects .,opencv,caffe
5. After a while all jars are deployed correctly in Maven repository: caffe-rc3-1.2-SNAPSHOT.jar , opencv-3.1.0-1.2-SNAPSHOT.jar and caffe-rc3-1.2-SNAPSHOT-sources.jar that supposedly contains /org/bytedeco/javacpp/macosx-x86_64/libcaffe.so.1.0.0-rc3

I add
<dependency> <groupId>org.bytedeco.javacpp-presets</groupId> <artifactId>caffe</artifactId> <version>rc3-1.2-SNAPSHOT</version> </dependency>

to my pom file and it imports the correct libraries into my project.
When I run caffe.java I get this error:
Exception in thread "main" java.lang.UnsatisfiedLinkError: no jnicaffe in java.library.path at java.lang.ClassLoader.loadLibrary(ClassLoader.java:1867) at java.lang.Runtime.loadLibrary0(Runtime.java:870) at java.lang.System.loadLibrary(System.java:1122) at org.bytedeco.javacpp.Loader.loadLibrary(Loader.java:654) at org.bytedeco.javacpp.Loader.load(Loader.java:492) at org.bytedeco.javacpp.Loader.load(Loader.java:409) at org.bytedeco.javacpp.caffe.<clinit>(caffe.java:16) at java.lang.Class.forName0(Native Method) at java.lang.Class.forName(Class.java:348) at org.bytedeco.javacpp.Loader.load(Loader.java:464) at org.bytedeco.javacpp.Loader.load(Loader.java:409) at org.bytedeco.javacpp.caffe$MessageLite.<clinit>(caffe.java:2967) at conpanna.caffe$2.command(caffe.java:182) at conpanna.caffe.main(caffe.java:396) Caused by: java.lang.UnsatisfiedLinkError: /private/var/folders/g3/nyry1nsx6lg0vjb5nkbw7jt40000gn/T/javacpp206074640320495/libjnicaffe.dylib: dlopen(/private/var/folders/g3/nyry1nsx6lg0vjb5nkbw7jt40000gn/T/javacpp206074640320495/libjnicaffe.dylib, 1): Library not loaded: libcaffe.so.1.0.0-rc3 Referenced from: /private/var/folders/g3/nyry1nsx6lg0vjb5nkbw7jt40000gn/T/javacpp206074640320495/libjnicaffe.dylib Reason: image not found at java.lang.ClassLoader$NativeLibrary.load(Native Method) at java.lang.ClassLoader.loadLibrary0(ClassLoader.java:1941) at java.lang.ClassLoader.loadLibrary(ClassLoader.java:1824) at java.lang.Runtime.load0(Runtime.java:809) at java.lang.System.load(System.java:1086) at org.bytedeco.javacpp.Loader.loadLibrary(Loader.java:637) ... 10 more

Sorry for long message, I just want to explain exactly what I did.
Thank you.

[saudet]

Please try to follow the instructions here to build on OS X:
https://github.com/bytedeco/javacpp-presets/wiki/Build-Environments#mac-os-x-x86_64

[saudet]

I just found out that is should be possible to get rid of that issue by linking with --cudart=static. Could you try this out and let me know what that gives? Let me know if you still have issues building.

[saudet]

I've just released an update that works well on machines without GPU as well. It has version "master-1.2":
https://github.com/bytedeco/javacpp-presets/tree/master/caffe
http://search.maven.org/#search%7Cga%7C1%7Cbytedeco%20caffe

We still need CUDA 7.5 installed, but I've tested it with Mac OS X (Yosemite) on a machine that has no GPU and it works fine. Just make sure that your code never calls either Caffe.SetDevice(gpu) or Caffe.set_mode(Caffe.GPU). Enjoy!

[bisher3]

Thank you Saudet! I can confirm this is working on a Mac without a GPU

[shiqin-zhang]

@saudet thanks for your https://github.com/bytedeco/javacpp-presets/tree/master/caffe , but I can not import the following dependence:

<dependency>
       <groupId>org.bytedeco.javacpp-presets</groupId>
       <artifactId>caffe-platform</artifactId>
       <version>1.0-1.3.3-SNAPSHOT</version>
</dependency>

I don't know weather it is out of date or some other reason, can you give a useful dependency?

[saudet]

@shiqin-zhang Make sure to enable the SNAPSHOT repository by adding something like the following in your pom.xml or settings.xml file:
https://github.com/bytedeco/javacpp-presets/blob/master/pom.xml#L118-L130

[saudet]

CUDA needs to be installed, even if you don't use or have a GPU.

[saudet]

FYI, separate CPU-only and GPU-enabled builds are now a reality! In addition to caffe-platform, adding something like the following will download and use the binaries for CUDA:

        <dependency>
            <groupId>org.bytedeco.javacpp-presets</groupId>
            <artifactId>caffe</artifactId>
            <version>1.0-1.3.4-SNAPSHOT</version>
            <classifier>linux-x86_64-gpu</classifier>
        </dependency>

If not, the CPU-only binaries are used.