Interface Java with Numpy. Try not to be reasonable.
What are you trying to accomplish? Get Java access to the highly tuned Numpy array library.
Why bother? At least one reason is that there is no reliable pure-Java way to get vectorized performance for dense matrix operations. This is a big deal(TM). Our preliminary investigation shows that (even with the JNI overhead) we can do array-array multiplication several hundred times faster by using the numpy and JNI over a naive pure-java solution. Another is just to explore working with these two technologies at once.
How do you get the JVM to talk with Python libraries? Its a combination of nio, JNI and an embedded python interpreter. By using nio's direct byte buffer's, we can (usually) avoid copying data as it moves in and out of the JVM. Using memoryview and buffer objects on the Python avoids copying as the values move back in.
What works: If you make an NPArray (that's the Java class that represents a Numpy array), you can execute max, min, log and array-array multiply. For log and multiply, the resulting NPArray is actually backed by memory that numpy allocated for its results array. The DType, array-order (C/Fortran), and byte order are properly communicated from the python to the Java side (but non-default NPType values don't move from Java to Python consistently).
A number of default properties have been set in the default.properties file. However, defaults can be overriden through a system.properties file in the same directory. The system.properties file has the same entries as default.properties and the same syntax.
- java.path= location of the JNI headers for your system. Can be found by locating jni.h on your system.
- java.os= JNI header subdirectory for OS-specific files. Commonly includes jni_md.h. Just the OS-directory name is required.
- python.path= location of the python root. Needs to have both the 'lib' and 'include' directories in it.
- python.version=python version name. Something like "python2.7". Running "python --version" will probably give a good idea of what to put here.
With system.properties in place, in the Numpy4J directory, run "ant".
Unit tests are run with "ant test". To do a quick perofrmance comparison, run "ant perf".
Performance is one of the goals, so we are monitoring it already. We have two major comarisons groups. The first compares Numpy4J to using a ByteBuffer wrapped inside of a Java NPArray. The second compares Numpy4J to using Java arrays. Our tests are Max, Log and Multiply. For the sake of curiosity, we ran the "max" test wit the ByteBuffer veresion returning Number objects and returning ints. In general, if an array of values is returned, Numpy4J is eventually faster by a significant amount.
Results were acquired via the "ant perf" task.
Running with 100 iterations (let the JVM know we're serrious?) and array sizes betweeen 10 and 5.2 million:
- Max Numpy4J vs. primtive array: Pritimite arrays register essentially zero time...Numpy4j always looses
- Max Numpy4J vs. NPArray returning ints: Numpy4j is 80% slower than pure java
- Max Numpy4J vs. NPArray returning Numbers: Numpy4j is five times faster
- Log Numpy4J vs. NPArray returning ints: Numpy4j is a little more than twice as fast
- Log Numpy4J vs. primtive array: Numpy4J is a little less than twice as fast
- Multiply Numpy4J vs. NPArray returning ints: Numpy4J is a little more than three times faster
- Multiply Numpy4J vs. primitive array: Numpy4J is a little less than twice as fast