Python end-to-end tutorial
The goal of this tutorial is to showcase different features of the SystemDS framework that can be accessed with the Python API. For this, we want to use the Adult dataset and predict whether the income of a person exceeds $50K/yr based on census data. The Adult dataset contains attributes like age, workclass, education, marital-status, occupation, race, […] and the labels >50K or <=50K. Most of these features are categorical string values, but the dataset also includes continuous features. For this, we define three different levels with an increasing level of detail with regard to features provided by SystemDS. In the first level, we simply get an already preprocessed dataset from a DatasetManager. The second level, shows the built-in preprocessing capabilities of SystemDS. With the third level, we want to show how we can integrate custom-built networks or algorithms into our Python program.
Prerequisite:
Level 1
This example shows how one can work with NumPy data within the SystemDS framework. More precisely, we will make use of the built-in DataManager, Multinomial Logistic Regression function, and the Confusion Matrix function. The dataset used in this tutorial is a preprocessed version of the “UCI Adult Data Set”. If you are interested in data preprocessing, take a look at level 2. If one wants to skip the explanation then the full script is available at the end of this level.
We will train a Multinomial Logistic Regression model on the training dataset and subsequently we will use the test dataset to assess how well our model can predict if the income is above or below $50K/yr based on the features.
Step 1: Load and prepare data
First, we get our training and testing data from the built-in DataManager. Since the multiLogReg function requires the labels (Y) to be > 0, we add 1 to all labels. This ensures that the smallest label is >= 1. Additionally we will only take a fraction of the training and test set into account to speed up the execution.
from systemds.context import SystemDSContext
from systemds.examples.tutorials.adult import DataManager
sds = SystemDSContext()
d = DataManager()
# limit the sample size
train_count = 15000
test_count = 5000
train_data, train_labels, test_data, test_labels = d.get_preprocessed_dataset(interpolate=True, standardize=True, dimred=0.1)
# Train data
X = sds.from_numpy(train_data[:train_count])
Y = sds.from_numpy(train_labels[:train_count])
Y = Y + 1.0
# Test data
Xt = sds.from_numpy(test_data[:test_count])
Yt = sds.from_numpy(test_labels[:test_count])
Yt = Yt + 1.0
Here the DataManager contains the code for downloading and setting up NumPy arrays containing the data. It is noteworthy that the function get_preprocessed_dataset has options for basic standardization, interpolation, and combining categorical features inside one column whose occurrences are below a certain threshold.
Step 2: Training
Now that we prepared the data, we can use the multiLogReg function. First, we will train the model on our training data. Afterward, we can make predictions on the test data and assess the performance of the model.
from systemds.operator.algorithm import multiLogReg
betas = multiLogReg(X, Y)
Note that nothing has been calculated yet. In SystemDS the calculation is executed once compute() is called. E.g. betas_res = betas.compute().
We can now use the trained model to make predictions on the test data.
from systemds.operator.algorithm import multiLogRegPredict
[_, y_pred, acc] = multiLogRegPredict(Xt, betas, Yt)
- The multiLogRegPredict function has three return values:
m, a matrix with the mean probability of correctly classifying each label. We do not use it further in this example.
y_pred, is the predictions made using the model
acc, is the accuracy achieved by the model.
Step 3: Confusion Matrix
A confusion matrix is a useful tool to analyze the performance of the model and to obtain a better understanding which classes the model has difficulties separating. The confusionMatrix function takes the predicted labels and the true labels. It then returns the confusion matrix for the predictions and the confusion matrix averages of each true class.
from systemds.operator.algorithm import confusionMatrix
confusion_matrix_abs, _ = confusionMatrix(y_pred, Yt).compute()
print(confusion_matrix_abs)
Full Script
In the full script, some steps are combined to reduce the overall script.
import numpy as np
from systemds.context import SystemDSContext
from systemds.examples.tutorials.adult import DataManager
from systemds.operator.algorithm import multiLogReg, multiLogRegPredict, confusionMatrix
sds = SystemDSContext()
d = DataManager()
# limit the sample size
train_count = 15000
test_count = 5000
train_data, train_labels, test_data, test_labels = d.get_preprocessed_dataset(interpolate=True, standardize=True, dimred=0.1)
# Train data
X = sds.from_numpy(train_data[:train_count])
Y = sds.from_numpy(train_labels[:train_count])
Y = Y + 1.0
# Test data
Xt = sds.from_numpy(test_data[:test_count])
Yt = sds.from_numpy(test_labels[:test_count])
Yt = Yt + 1.0
betas = multiLogReg(X, Y)
[_, y_pred, acc] = multiLogRegPredict(Xt, betas, Yt)
confusion_matrix_abs, _ = confusionMatrix(y_pred, Yt).compute()
print(confusion_matrix_abs)
Level 2
This part of the tutorial shows an overview of the preprocessing capabilities that SystemDS has to offer. We will take an unprocessed dataset using the csv format and read it with SystemDS. Then do the heavy lifting for the preprocessing with SystemDS. As mentioned before, we want to use the Adult dataset for this task. If one wants to skip the explanation, then the full script is available at the end of this level.
Step 1: Metadata and reading
First of all, we need to download the dataset and create a mtd-file for specifying different metadata about the dataset. We download the train and test dataset from: https://archive.ics.uci.edu/ml/datasets/adult
The downloaded dataset will be slightly modified for convenience. These modifications entail removing unnecessary newlines at the end of the files and adding column names at the top of the files such that the first line looks like:
age,workclass,fnlwgt,education,education-num,marital-status,occupation,relationship,race,sex,capital-gain,capital-loss,hours-per-week,native-country,income
We also delete the line holding the string value Cross validator inside the test dataset.
After these modifications, we have to define a mtd file for each file we want to read. This mtd file has to be in the same directory as the dataset. In this particular example, the dataset is split into two files “train_data.csv” and “test_data.csv”. We want to read both, which means that we will define a mtd file for each of them. Those files will be called “train_data.csv.mtd” and “test_data.csv.mtd”. In these files, we can define certain properties that the file has and also specify which values are supposed to get treated like missing values.
The content of the train_data.csv.mtd file is:
{
"data_type": "frame",
"format": "csv",
"header": true,
"naStrings": [ "?", "" ],
"rows": 32561,
"cols": 15
}
The “format” of the file is csv, and “header” is set to true because we added the feature names as headers to the csv files. The value “data_type” is set to frame, as the preprocessing functions that we use require this datatype. The value of “naStrings” is a list of all the string values that should be treated as unknown values during the preprocessing. Also, “rows” in our example is set to 32561, as we have this many entries and “cols” is set to 15 as we have 14 features, and one label column inside the files. We will later show how we can split them.
After these requirements are completed, we have to define a SystemDSContext for reading our dataset. We can do this in the following way:
sds = SystemDSContext()
train_count = 32561
test_count = 16281
With this context we can now define a read operation using the path of the dataset and a schema. The schema simply defines the data types for each column.
As already mentioned, SystemDS supports lazy execution by default, which means that the read operation is only executed after calling the compute() function.
SCHEMA = '"DOUBLE,STRING,DOUBLE,STRING,DOUBLE,STRING,STRING,STRING,STRING,STRING,DOUBLE,DOUBLE,DOUBLE,STRING,STRING"'
dataset_path_train = "adult/train_data.csv"
dataset_path_test = "adult/test_data.csv"
F1 = sds.read(
dataset_path_train,
schema=SCHEMA
)
F2 = sds.read(
dataset_path_test,
schema=SCHEMA
)
Step 2: Defining preprocess operations
Now that the read operation has been declared, we can define an additional file for the further preprocessing of the dataset. For this, we create a .json file that holds information about the operations that will be performed on individual columns. For the sake of this tutorial we will use the file “jspec.json” with the following content:
{
"impute":
[ { "name": "age", "method": "global_mean" }
,{ "name": "workclass" , "method": "global_mode" }
,{ "name": "fnlwgt", "method": "global_mean" }
,{ "name": "education", "method": "global_mode" }
,{ "name": "education-num", "method": "global_mean" }
,{ "name": "marital-status" , "method": "global_mode" }
,{ "name": "occupation" , "method": "global_mode" }
,{ "name": "relationship" , "method": "global_mode" }
,{ "name": "race" , "method": "global_mode" }
,{ "name": "sex" , "method": "global_mode" }
,{ "name": "capital-gain", "method": "global_mean" }
,{ "name": "capital-loss", "method": "global_mean" }
,{ "name": "hours-per-week", "method": "global_mean" }
,{ "name": "native-country" , "method": "global_mode" }
],
"bin": [ { "name": "age" , "method": "equi-width", "numbins": 3 }],
"dummycode": ["age", "workclass", "education", "marital-status", "occupation", "relationship", "race", "sex", "native-country"],
"recode": ["income"]
}
Our dataset has missing values. An easy way to deal with that circumstance is to use the “impute” option that SystemDS supports. We simply pass a list that holds all the relations between column names and the method of interpolation. A more specific example is the “education” column. In the dataset certain entries have missing values for this column. As this is a string feature, we can simply define the method as “global_mode” and replace every missing value with the global mode inside this column. It is important to note that we first had to define the values of the missing strings in our selected dataset using the .mtd files (naStrings”: [ “?”, “” ]).
With the “bin” keyword we can discretize continuous values into a small number of bins. Here the column with age values is discretized into three age intervals. The only method that is currently supported is equi-width binning.
The column-level data transformation “dummycode” allows us to one-hot-encode a column. In our example we first bin the “age” column into 3 different bins. This means that we now have one column where one entry can belong to one of 3 age groups. After using “dummycode”, we transform this one column into 3 different columns, one for each bin.
At last, we make use of the “recode” transformation for categorical columns, it maps all distinct categories in the column into consecutive numbers, starting from 1. In our example we recode the “income” column, which transforms it from “<=$50K” and “>$50K” to “1” and “2” respectively.
Another good resource for further ways of processing is: https://apache.github.io/systemds/site/dml-language-reference.html
There we provide different examples for defining jspec’s and what functionality is currently supported.
After defining the .jspec file we can read it by passing the filepath, data_type and value_type using the following command:
dataset_jspec = "adult/jspec.json"
jspec = sds.read(dataset_jspec, data_type="scalar", value_type="string")
Finally, we need to define a custom dml file to split the features from the labels and replace certain values, which we will need later. We will call this file “preprocess.dml”:
get_X = function(matrix[double] X, int start, int stop)
return (matrix[double] returnVal) {
returnVal = X[start:stop,1:ncol(X)-1]
}
get_Y = function(matrix[double] X, int start, int stop)
return (matrix[double] returnVal) {
returnVal = X[start:stop,ncol(X):ncol(X)]
}
replace_value = function(matrix[double] X, double pattern , double replacement)
return (matrix[double] returnVal) {
returnVal = replace(target=X, pattern=pattern, replacement=replacement)
}
The get_X function simply extracts every column except the last one and can also be used to pick certain slices from the dataset. The get_Y function only extracts the last column, which in our case holds the labels. Replace_value is used to replace a double value with another double. The preprocess.dml file can be read with the following command:
preprocess_src_path = "preprocess.dml"
PREPROCESS_package = sds.source(preprocess_src_path, "preprocess", print_imported_methods=True)
The print_imported_methods flag can be used to verify whether every method has been parsed correctly.
Step 3: Applying the preprocessing steps
Generally speaking, we would use the transform_encode function on the train dataset and with the returned encoding call the transform_apply function on the test dataset. In the case of the Adult dataset, we have inconsistent label names inside the test dataset and the train dataset, which is why we will show how we can deal with that using SystemDS. First of all, we combine the train and the test dataset by using the rbind() function. This function simply appends the Frame F2 at the end of Frame F1. This is necessary to ensure that the encoding is identical between train and test dataset.
X1 = F1.rbind(F2)
In order to use our jspec file we can apply the transform_encode() function. We simply have to pass the read .json file from before. In our particular case we obtain the Matrix X1 and the Frame M1 from the operation. X1 holds all the encoded values and M1 holds a mapping between the encoded values and all the initial values. Columns that have not been specified in the .json file were not altered.
X1, M1 = X1.transform_encode(spec=jspec)
We now can use the previously parsed dml file for splitting the dataset and unifying the inconsistent labels. It is noteworthy that the file is parsed such that we can directly call the function names from the Python API.
X = PREPROCESS_package.get_X(X1, 1, train_count)
Y = PREPROCESS_package.get_Y(X1, 1, train_count)
Xt = PREPROCESS_package.get_X(X1, train_count, train_count+test_count)
Yt = PREPROCESS_package.get_Y(X1, train_count, train_count+test_count)
Yt = PREPROCESS_package.replace_value(Yt, 3.0, 1.0)
Yt = PREPROCESS_package.replace_value(Yt, 4.0, 2.0)
Step 4: Training and confusion matrix
Now that we prepared the data we can use the multiLogReg function. These steps are identical to step 2 and 3 that have already been described in level 1 of this tutorial.
from systemds.operator.algorithm import multiLogReg
from systemds.operator.algorithm import confusionMatrix
from systemds.operator.algorithm import multiLogRegPredict
betas = multiLogReg(X, Y)
[_, y_pred, acc] = multiLogRegPredict(Xt, betas, Yt)
confusion_matrix_abs, _ = confusionMatrix(y_pred, Yt).compute()
print(confusion_matrix_abs)
Full Script
The complete script now can be seen here:
import numpy as np
from systemds.context import SystemDSContext
from systemds.operator.algorithm import multiLogReg, multiLogRegPredict, confusionMatrix
train_count = 32561
test_count = 16281
dataset_path_train = "adult/train_data.csv"
dataset_path_test = "adult/test_data.csv"
dataset_jspec = "adult/jspec.json"
preprocess_src_path = "preprocess.dml"
sds = SystemDSContext()
SCHEMA = '"DOUBLE,STRING,DOUBLE,STRING,DOUBLE,STRING,STRING,STRING,STRING,STRING,DOUBLE,DOUBLE,DOUBLE,STRING,STRING"'
F1 = sds.read(dataset_path_train, schema=SCHEMA)
F2 = sds.read(dataset_path_test, schema=SCHEMA)
jspec = sds.read(dataset_jspec, data_type="scalar", value_type="string")
PREPROCESS_package = sds.source(preprocess_src_path, "preprocess", print_imported_methods=True)
X1 = F1.rbind(F2)
X1, M1 = X1.transform_encode(spec=jspec)
X = PREPROCESS_package.get_X(X1, 1, train_count)
Y = PREPROCESS_package.get_Y(X1, 1, train_count)
Xt = PREPROCESS_package.get_X(X1, train_count, train_count+test_count)
Yt = PREPROCESS_package.get_Y(X1, train_count, train_count+test_count)
Yt = PREPROCESS_package.replace_value(Yt, 3.0, 1.0)
Yt = PREPROCESS_package.replace_value(Yt, 4.0, 2.0)
betas = multiLogReg(X, Y)
[_, y_pred, acc] = multiLogRegPredict(Xt, betas, Yt)
confusion_matrix_abs, _ = confusionMatrix(y_pred, Yt).compute()
print(confusion_matrix_abs)
Level 3
In this level we want to show how we can integrate a custom built algorithm using the Python API. For this we will introduce another dml file, which can be used to train a basic feed forward network.
Step 1: Obtain data
For the whole data setup please refer to level 2, Step 1 to 3, as these steps are identical.
import numpy as np
from systemds.context import SystemDSContext
train_count = 32561
test_count = 16281
dataset_path_train = "adult/train_data.csv"
dataset_path_test = "adult/test_data.csv"
dataset_jspec = "adult/jspec.json"
preprocess_src_path = "preprocess.dml"
sds = SystemDSContext()
SCHEMA = '"DOUBLE,STRING,DOUBLE,STRING,DOUBLE,STRING,STRING,STRING,STRING,STRING,DOUBLE,DOUBLE,DOUBLE,STRING,STRING"'
F1 = sds.read(dataset_path_train, schema=SCHEMA)
F2 = sds.read(dataset_path_test, schema=SCHEMA)
jspec = sds.read(dataset_jspec, data_type="scalar", value_type="string")
PREPROCESS_package = sds.source(preprocess_src_path, "preprocess", print_imported_methods=True)
X1 = F1.rbind(F2)
X1, M1 = X1.transform_encode(spec=jspec)
X = PREPROCESS_package.get_X(X1, 1, train_count)
Y = PREPROCESS_package.get_Y(X1, 1, train_count)
Xt = PREPROCESS_package.get_X(X1, train_count, train_count+test_count)
Yt = PREPROCESS_package.get_Y(X1, train_count, train_count+test_count)
Yt = PREPROCESS_package.replace_value(Yt, 3.0, 1.0)
Yt = PREPROCESS_package.replace_value(Yt, 4.0, 2.0)
Step 2: Load the algorithm
We use a neural network with 2 hidden layers, each consisting of 200 neurons. First, we need to source the dml file for neural networks. This file includes all the necessary functions for training, evaluating, and storing the model. The returned object of the source call is further used for calling the functions. The file can be found here:
:doc:tests/examples/tutorials/neural_net_source.dml
FFN_package = sds.source(neural_net_src_path, "fnn", print_imported_methods=True))
Step 3: Training the neural network
Training a neural network in SystemDS using the train function is straightforward. The first two arguments are the training features and the target values we want to fit our model on. Then we need to set the hyperparameters of the model. We choose to train for 1 epoch with a batch size of 16 and a learning rate of 0.01, which are common parameters for neural networks. The seed argument ensures that running the code again yields the same results.
epochs = 1
batch_size = 16
learning_rate = 0.01
seed = 42
network = FFN_package.train(X, Y, epochs, batch_size, learning_rate, seed)
Step 4: Saving the model
For later usage, we can save the trained model. We only need to specify the name of our model and the file path. This call stores the weights and biases of our model.
FFN_package.save_model(network, '"model/python_FFN/"').compute(verbose=True)
Full Script
The complete script now can be seen here:
import numpy as np
from systemds.context import SystemDSContext
train_count = 32561
test_count = 16281
dataset_path_train = "adult/train_data.csv"
dataset_path_test = "adult/test_data.csv"
dataset_jspec = "adult/jspec.json"
preprocess_src_path = "preprocess.dml"
neural_net_src_path = "neural_net_source.dml"
sds = SystemDSContext()
SCHEMA = '"DOUBLE,STRING,DOUBLE,STRING,DOUBLE,STRING,STRING,STRING,STRING,STRING,DOUBLE,DOUBLE,DOUBLE,STRING,STRING"'
F1 = sds.read(dataset_path_train, schema=SCHEMA)
F2 = sds.read(dataset_path_test, schema=SCHEMA)
jspec = sds.read(dataset_jspec, data_type="scalar", value_type="string")
PREPROCESS_package = sds.source(preprocess_src_path, "preprocess", print_imported_methods=True)
X1 = F1.rbind(F2)
X1, M1 = X1.transform_encode(spec=jspec)
X = PREPROCESS_package.get_X(X1, 1, train_count)
Y = PREPROCESS_package.get_Y(X1, 1, train_count)
Xt = PREPROCESS_package.get_X(X1, train_count, train_count+test_count)
Yt = PREPROCESS_package.get_Y(X1, train_count, train_count+test_count)
Yt = PREPROCESS_package.replace_value(Yt, 3.0, 1.0)
Yt = PREPROCESS_package.replace_value(Yt, 4.0, 2.0)
FFN_package = sds.source(neural_net_src_path, "fnn", print_imported_methods=True)
epochs = 1
batch_size = 16
learning_rate = 0.01
seed = 42
network = FFN_package.train(X, Y, epochs, batch_size, learning_rate, seed)
FFN_package.save_model(network, '"model/python_FFN/"').compute(verbose=True)