XAISuite General Demo

XAISuite General Demo

Welcome to the XAISuite General Demo. Here you’ll find countless examples about how to use XAISuite for easier machine learning model training, explanation, and explanation comparison.

First, start by installing the XAISuite library.

!pip install XAISuite==2.8.3

Then, import the library. You may need to restart runtime for this to work correctly.

from xaisuite import*

Run the following code to see a mini doc of XAISuite classes and functions.

help(DataLoader)
help(DataProcessor)
help(ModelTrainer)
help(InsightGenerator)

Let’s first start with Data Loading. The xaisuite.dataHandler.DataLoaderclass allows loading of data from different sources.

DataLoader(make_classification)

DataLoader(load_diabetes, return_X_y = True)

import numpy as np
data = np.zeros(20)
DataLoader(data)

import pandas as pd
data = pd.DataFrame([[1, 2, 3], [3, 4, 5]])
DataLoader(data)

Another option: DataLoader('path/to/local/file')

DataLoader also has options to specify the variable names, target name, categorical variables, etc. If not specified, these values are inferred.

Visualize variables loaded by DataLoader

load_data = DataLoader(make_classification, n_features = 10)
load_data.plot()

XAISuiteGeneralDemoOutput1

For optimal model training, additional processing must be done to the data. This is where xaisuite.dataHandler.DataProcessor comes into play. You can process data with default parameters or pass in your own transforms (like from the sklearn.preprocessing library).

load_data = DataLoader(make_classification, n_features = 10)
DataProcessor(load_data)

load_data = DataLoader(load_diabetes, return_X_y = True)
DataProcessor(load_data, test_size = 0.1)

You can also use transform components, a short example of which is given below with placeholders:

from sklearn.preprocessing import StandardScaler

load_data = DataLoader(load_diabetes, return_X_y = True)

DataProcessor(load_data, test_size = 0.1, target_transform = "component: TargetTransform()")

To train a model, simply do:

load_data = DataLoader(load_diabetes, return_X_y = True)
process_data = DataProcessor(load_data, test_size = 0.1)

ModelTrainer("SVR", process_data)

You can also pass in a model directly without using a String representation.

from sklearn.svm import SVR
load_data = DataLoader(load_diabetes, return_X_y = True)
process_data = DataProcessor(load_data, test_size = 0.1)

ModelTrainer(SVR, process_data, epsilon = 0.2)

For explaining, simply list the desired explanations.

from sklearn.svm import SVR
load_data = DataLoader(load_diabetes, return_X_y = True)
process_data = DataProcessor(load_data, test_size = 0.1)

ModelTrainer(SVR, process_data, explainers = ["lime", "shap"], epsilon = 0.2)

You can pass in arguments to the explainers:

from sklearn.svm import SVR
load_data = DataLoader(load_diabetes, return_X_y = True)
process_data = DataProcessor(load_data, test_size = 0.1)

ModelTrainer(SVR, process_data, explainers = {"lime": {"feature_selection": "none"}, "shap": {}}, epsilon = 0.2)

To access the explanations, use the getExplanationsFor, getAllExplanations, or getSummaryExplanations functions. Use plotExplanations for explanation visualization.

from sklearn.svm import SVR
load_data = DataLoader(load_diabetes, return_X_y = True)
process_data = DataProcessor(load_data, test_size = 0.1)
train_model = ModelTrainer(SVR, process_data, explainers = {"lime": {"feature_selection": "none"}, "shap": {}}, epsilon = 0.2)

explanations = train_model.getExplanationsFor([]) # Gets all explanations. You can also request explanations for a specific instance
train_model.plotExplanations("lime", 1) #Display the lime explainer results for the 2nd instance returned by getExplanationsFor()

Model score is 0.22886080630718109

Generating explanations.

0%|          | 0/45 [00:00<?, ?it/s]

XAISuiteGeneralDemoOutput2

Calculate similarity between explainers using the Shreyan Distance

from sklearn.svm import SVR
load_data = DataLoader(load_diabetes, return_X_y = True)
process_data = DataProcessor(load_data, test_size = 0.1)
train_model = ModelTrainer(SVR, process_data, explainers = {"lime": {"feature_selection": "none"}, "shap": {}}, epsilon = 0.2)
explanations = train_model.getExplanationsFor([])

insights = InsightGenerator(explanations)
print(insights.calculateExplainerSimilarity("lime", "shap"))

Model score is 0.14626289816154203

Generating explanations.

0%|          | 0/45 [00:00<?, ?it/s]

0.8081355932203389

NOTE: For examples using tensorflow or pytorch models, check out our other demos.