QuickStart

Let’s take a look at some code examples.

Matrix Operations

Making use of SystemDS, let us multiply an Matrix with an scalar:

# Import SystemDSContext
from systemds.context import SystemDSContext
# Create a context and if necessary (no SystemDS py4j instance running)
# it starts a subprocess which does the execution in SystemDS
with SystemDSContext() as sds:
    # Full generates a matrix completely filled with one number.
    # Generate a 5x10 matrix filled with 4.2
    m = sds.full((5, 10), 4.20)
    # multiply with scalar. Nothing is executed yet!
    m_res = m * 3.1
    # Do the calculation in SystemDS by calling compute().
    # The returned value is an numpy array that can be directly printed.
    print(m_res.compute())
    # context will automatically be closed and process stopped

As output we get

[[13.02 13.02 13.02 13.02 13.02 13.02 13.02 13.02 13.02 13.02]
 [13.02 13.02 13.02 13.02 13.02 13.02 13.02 13.02 13.02 13.02]
 [13.02 13.02 13.02 13.02 13.02 13.02 13.02 13.02 13.02 13.02]
 [13.02 13.02 13.02 13.02 13.02 13.02 13.02 13.02 13.02 13.02]
 [13.02 13.02 13.02 13.02 13.02 13.02 13.02 13.02 13.02 13.02]]

The Python SystemDS package is compatible with numpy arrays. Let us do a quick element-wise matrix multiplication of numpy arrays with SystemDS. Remember to first start up a new terminal:

import numpy as np
from systemds.context import SystemDSContext

# create a random array
m1 = np.array(np.random.randint(100, size=5 * 5) + 1.01, dtype=np.double)
m1.shape = (5, 5)
# create another random array
m2 = np.array(np.random.randint(5, size=5 * 5) + 1, dtype=np.double)
m2.shape = (5, 5)

# Create a context
with SystemDSContext() as sds:
    # element-wise matrix multiplication, note that nothing is executed yet!
    m_res = sds.from_numpy(m1) * sds.from_numpy(m2)
    # lets do the actual computation in SystemDS! The result is an numpy array
    m_res_np = m_res.compute()
    print(m_res_np)

More complex operations

SystemDS provides algorithm level functions as built-in functions to simplify development. One example of this is l2SVM, a high level functions for Data-Scientists. Let’s take a look at l2svm:

# Import numpy and SystemDS
import numpy as np
from systemds.context import SystemDSContext
from systemds.operator.algorithm import l2svm

# Set a seed
np.random.seed(0)
# Generate random features and labels in numpy
# This can easily be exchanged with a data set.
features = np.array(np.random.randint(
    100, size=10 * 10) + 1.01, dtype=np.double)
features.shape = (10, 10)
labels = np.zeros((10, 1))

# l2svm labels can only be 0 or 1
for i in range(10):
    if np.random.random() > 0.5:
        labels[i][0] = 1

# compute our model
with SystemDSContext() as sds:
    model = l2svm(sds.from_numpy(features),
                  sds.from_numpy(labels)).compute()
    print(model)

The output should be similar to

[[ 0.02033445]
 [-0.00324092]
 [ 0.0014692 ]
 [ 0.02649209]
 [-0.00616902]
 [-0.0095087 ]
 [ 0.01039221]
 [-0.0011352 ]
 [-0.01686351]
 [-0.03839821]]

To get the full performance of SystemDS one can modify the script to only use internal functionality, instead of using numpy arrays that have to be transfered into systemDS. The above script transformed goes like this:

# Import SystemDS
from systemds.context import SystemDSContext
from systemds.operator.algorithm import l2svm

with SystemDSContext() as sds:
    # Generate 10 by 10 matrix with values in range 0 to 100.
    features = sds.rand(10, 10, 0, 100)
    # Add value to all cells in features
    features += 1.1
    # Generate labels of all ones and zeros
    labels = sds.rand(10, 1, 1, 1, sparsity = 0.5)

    model = l2svm(features, labels).compute()
    print(model)

When reading in datasets for processing it is highly recommended that you read from inside systemds using sds.read(“file”), since this avoid the transferring of numpy arrays.