Data Preprocessing for Inferencing

Data Preprocessing for Inferencing

Inference is decomposed into three distinct steps, and this is the first: data preprocessing for inferencing. The inf_data_preprocessing pipeline includes nodes for:

  1. Input Image Preprocessing
  2. Out-of-Distribution (OOD) Detection

You can extend this pipeline by adding nodes to perform additional operations on input images before they are passed to the model.

OOD Detection

For Out-of-Distribution (OOD) Detection, we use the pytorch-ood library. A separate pipeline, titled ood_detection, is provided to train and integrate custom OOD detectors.

Features of OOD Detection Pipeline

  • Preconfigured Templates
    We provide templates for the following OOD detectors:
    • MSP Detector: Maximum Softmax Probability Detector
    • RMD Detector: Relative Mahalanobis Distance Detector
    • Multi-Mahalanobis Detector: An enhanced version of the Mahalanobis-based detector.
  • Dataset Example
    The pipeline is configured to use CIFAR-10 as the “out-of-distribution” dataset by default. However, you can adapt the pipeline to use any dataset of your choice.

Getting Started with OOD Detection

  1. Train your custom OOD detector by modifying the templates provided in the ood_detection pipeline.
  2. For further customization or advanced features, refer to the PyTorch OOD Documentation.

Customization Tips

  • Add Preprocessing Steps: Extend the inf_data_preprocessing pipeline to include additional transformations or preprocessing steps suited to your use case.
  • Experiment with OOD Detectors: Customize the OOD pipeline to use different datasets or fine-tune the provided detectors to improve performance.