Training HIGGS using IBM Snap ML

In this example we will train a Decision Tree model on the HIGGS dataset, using both scikit-learn and snap-ml-local.

Getting the Data

Download and decompress the data from the LIBSVM repository:

mkdir data
cd data
bunzip2 HIGGS.bz2
cd ../

Data Preprocessing

The data is in SvmLight format which is not very efficient since this dataset is dense. Therefore, we suggest to perform the following pre-processing, which converts it to dense format, performs normalization and then dumps it to numpy binary format for fast loading. Note that snapml is compatible with scikit-learn. Thus we can use the broad functionality of scikit-learn to perform preprocessing as needed.

import numpy as np

# import preprocessing functions from scikit-learn
from sklearn.datasets import load_svmlight_file
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import normalize

# load data from libsvm format
X,y = load_svmlight_file("data/HIGGS")

# Make the train-test split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25, random_state=42)

# Convert to numpy ararys
X_train = np.array(X_train.todense())
X_test  = np.array(X_test.todense())

# Normalize the training data
X_train = normalize(X_train, axis=1, norm='l1')
X_test  = normalize(X_test,  axis=1, norm='l1')

# Save the dense matrices"data/HIGGS.X_train", X_train)"data/HIGGS.X_test",  X_test)

# Save the labels"data/HIGGS.y_train", y_train)"data/HIGGS.y_test", y_test)

Training and Evaluating a Decision Tree

In the following we will show how to train a decision tree classifier using snapml on the HIGGS dataset. Therefore we first load the preprocessed data for numpy binary format

import time
import numpy as np
from scipy import sparse

# load the data
t0 = time.time()
X_train = np.load("data/HIGGS.X_train.npy")
X_test  = np.load("data/HIGGS.X_test.npy")
y_train = np.load("data/HIGGS.y_train.npy")
y_test  = np.load("data/HIGGS.y_test.npy")
print("Data load time (s):  {0:.2f}".format(time.time()-t0))

Then we specify the model parameters and initialize the decision tree classifier

# specify model parameters
max_depth = None

# import Snap ML DecisionTreeClassifier from pai4sk module directly
from pai4sk import DecisionTreeClassifier as SnapTree

# initialize classifier
dt = SnapTree(random_state=0, max_depth=max_depth)

In the next step we train our classifier on the training dataset. We will introduce a parameter num_ex_used for the user to specify how many examples should be used for training. This serves for reducing runtimes for testing.

# specify how many examples should be used for training
num_ex_used = X_train.shape[0]  # use the full training set

# Training
t0 = time.time()[0:num_ex_used], y_train[0:num_ex_used])
print("[snap] Training time (s):  {0:.2f}".format(time.time()-t0))

After the training has finished, we can validate the predictive performance of our model on the hold-out test set. Again, we have the option to reuse evaluation metrics implemented in scikit-learn to evaluate our model.

# Inference
pred_test = dt.predict(X_test)

# Evaluate accuracy_score on test set
from sklearn.metrics import accuracy_score
acc_snap = accuracy_score(y_test, pred_test)
print("[snap] Accuracy score:   {0:.4f}".format(acc_snap))

For the user interested in the performance comparison of snapml to the standard scikit-learn library, we will show how the same classifier can be trained using scikit-learn. This only requires minimal changes to the above code:

# load data and specify parameters as in the example above
# [...]

# Import DecisionTreeClassifier from sklearn
from sklearn.tree import DecisionTreeClassifier as skTree
dt = skTree(random_state=0, presort=False, max_depth=max_depth)

# Training time
t0 = time.time()[0:n_ex], y_train[0:n_ex])
print("[sklearn] Training time (s):  {0:.2f}".format(time.time()-t0))

# Inference
pred_test = dt.predict(X_test)

# Evaluate accuracy_score on test set
from sklearn.metrics import accuracy_score
acc_sklearn = accuracy_score(y_test, pred_test)
print("[sklearn] Accuracy score:   {0:.4f}".format(acc_sklearn))

© Copyright IBM Corporation 2018