/temperature-scaling

🌡 Temperature scaling of convolutional neural networks using TensorFlow 2.X API

Primary LanguageJupyter Notebook

🌡 Temperature scaling using TensorFlow 2

An example of performing 🌡 temperature scaling using the TensorFlow 2.X API based on the paper On Calibration of Modern Neural Networks

Overview

  • A simple 🚀 MiniVGGNet model is built and trained on the 👕Fashion MNIST 👗 dataset using the TensorFlow 2.X API
  • The expected calibration error (ECE) is calculated and reliability diagrams plotted for the trained model
  • Reusable functions provided for calculating ECE, as well as plotting reliability diagrams
  • Temperature scaling is performed using gradient descent and the Adam optimiser
  • Based on PyTorch implementation by paper authors.

Requirements

  • Python 3.X
  • TensorFlow 2.X