This directory has Jupyter notebooks that demonstrate explainability and model card generation with the Intel® Explainable AI Tools.
| Notebook | Domain: Use Case | Framework | Description |
|---|---|---|---|
| Explaining Image Classification and Object Detection Using the CAM Explainer | CV: Image Classification & Object Detection | PyTorch*, TensorFlow* and Intel® Explainable AI API | Two separate notebooks that demonstrate how to use the CAM explainer API to explain ImageNet classification and detection examples using a ResNet50 CNN from the TorchVision & Torch model hub and TF's keras.applications model hub. |
| Explaining Custom CNN MNIST Classification Using the Attributions Explainer | CV: Image Classification | PyTorch and Intel® Explainable AI API | Demonstrates how to use the attributions explainer API to explain an MNIST classification example using a Custom PyTorch CNN. |
| Explaining Custom NN NewsGroups Classification Using the Attributions Explainer | NLP: Text Classification | PyTorch and Intel® Explainable AI API | Demonstrates how to use the attributions explainer API to explain a NewsGroups dataset text classification example using a Custom TensorFlow NN. |
| Explaining Custom CNN CIFAR-10 Classification Using the Attributions Explainer | CV: Image Classification | PyTorch and Intel® Explainable AI API | Demonstrates how to use the attributions explainer API to explain the CIFAR-10 dataset image classification example using a Custom PyTorch CNN. |
| Multimodal Breast Cancer Detection Explainability using the Intel® Explainable AI API | CV: Image Classification & NLP: Text Classification | PyTorch, HuggingFace, Intel® Explainable AI API & Intel® Transfer Learning Tool API | Demonstrates how to use the attributions and metrics explainer API's to explain the classification of a text and image breast cancer dataset using a PyTorch ResNet50 CNN and a HuggingFace ClinicalBert Transformer. |
| Explaining Fine Tuned Text Classifier with PyTorch using the Intel® Explainable AI API | NLP: Text Classification | PyTorch, HuggingFace, Intel® Explainable AI API & Intel® Transfer Learning Tool API | Demonstrates how to use the attributions explainer API's to explain the classification of a text using HuggingFace Transformer. |
| Explaining a Custom Neural Network Heart Disease Classification Using the Attributions Explainer | Numerical/Categorical: Tabular Classification | TensorFlow & Intel® Explainable AI API | Demonstrates how to use the attributions explainer API's to explain the classification of a Tabular data using a TensorFlow custom NN. |
| Notebook | Domain: Use Case | Framework | Description |
|---|---|---|---|
| Generating a Model Card with PyTorch | Numerical/Categorical: Tabular Classification | PyTorch | Demonstrates training a multilayer network using the "Adult" dataset from the UCI repository to predict whether a person has a salary greater or less than $50,000, then uses the Model Card Generator to create a model card with interactive graphics to analyze the model. |
| Detecting Issues in Fairness by Generate Model Card from TensorFlow Estimators | Numerical/Categorical: Tabular Classification | TensorFlow | Uses a TFX pipeline to train and evaluate a model using the COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) dataset to generate a risk score indended to determine a defendant's likelihood of reoffending. The Model Card Generator is then used to create interative graphics visualizing racial bias in the model's predictions. |
| Creating Model Card for Toxic Comments Classification in TensorFlow | Numerical/Categorical: Tabular Classification | TensorFlow | Adapts a TensorFlow Fairness Exercise notebook to use the Model Card Generator. The notebook trains a model to detect toxicity in online coversations and graphically analyzes accuracy metrics by gender. |
| Creating Model Card for Hate Speech Detection using Hugging Face model | Numerical/Categorical: Tabular Classification | PyTorch | Utilizes a model hosted on Hugging Face Hub for detecting hatespeech in English language using the HateXplain dataset. The Model Card Generator is then used to create a model card with interactive graphics to analyze the model performance metrics at threshold and Bias AUC metric for target groups. |
| Multiclass classification of Hate Speech using Hugging Face model | Numerical/Categorical: Tabular Classification | PyTorch | Uses a model hosted on Hugging Face Hub for classifying hate speech into Hate, Offensive, or Normal categories using the HateXplain dataset. The Model Card Generator is then used to create a model card with individual interactive graphics for each class to analyze the model performance metrics at threshold and the Bias AUC metric for target groups. |
*Other names and brands may be claimed as the property of others. Trademarks