Jumpy allows you to use nd4j from python via pointers (no network communication required like a lot of python libraries).

Jumpy requires an uber jar (a jar file that contains nd4j and all its dependencies) and the path to this file is expected to be found in the environment variable JUMPY_CLASS_PATH.

Easiest way to build the uber jar is by running mvn package on the pom.xml file:

git clone https://www.github.com/deeplearning4j/jumpy.git
cd jumpy
mvn package

This will create a jar file called dl4j-1.0.0-SNAPSHOT.jar in the target directory. Set JUMPY_CLASS_PATH environment variable to path of this file.

export JUMPY_CLASS_PATH='/...../jumpy/target/dl4j-1.0.0-SNAPSHOT.jar'

Finally, either install jumpy via pip:

pip install jumpy

Or from source:

python setup.py install

Just like numpy, you can initialize an array using .zeros() or .ones()

import jumpy as jp

x = jp.zeros((32, 16))
y = jp.ones((32, 16))

A numpy ndarray instance can be converted to a jumpy ndarray instance (and vice-versa) without copying the data

import jumpy as jp
import numpy as np

x_np = np.random.random((100, 50))
x_jp = jp.array(x_np)

Simply call the .numpy() method of jumpy.ndarray.ndarray

import jumpy as jp

x_jp = jp.zeros((100,50))
x_np = x_jp.numpy()

• Basic operators like + - * / += -= *= /= are overloaded and broadcasting is supported.
• Indexing, slicing and assignment behaviour has been made as close to numpy as possible.
• Check jumpy/ops/ to see available ops.

• Check for open issues, or open a new issue to start a discussion around a feature idea or a bug.
• We could use more ops! Have a look at available ops (jumpy/ops/), it's quite easy to add new ones.
• Send a pull request and bug us on Gitter until it gets merged and published. :)