Postprocessing on Predictions

The inf_pred_postprocessing pipeline represents the final stage of inference, where model predictions undergo postprocessing. This pipeline includes nodes for:
1. Conformal Predictions
2. Explainability using Integrated Gradients
3. Logging Predictions

Conformal Predictions

To improve the reliability of model predictions, we use Conformal Predictions. This approach ensures that the model provides a set of predictions, with a guarantee that the ground truth will fall within this set with a specified level of certainty.

Method: We use the RAPS (Regularized Adaptive Prediction Sets) method from the torchcp package.
References:
- TorchCP Documentation
- RAPS Research Paper

Conformal predictions help enhance trust in model outputs, especially in high-stakes applications.

Integrated Gradients

For explainability, we use Integrated Gradients, a technique that helps visualize the contribution of each input feature (e.g., pixels) to the model’s prediction.

Tool: The Captum library is used for this purpose.
Usage: Integrated Gradients provide insights into the importance of each pixel in the input image, enabling better interpretability of the model’s decision-making process.

Logging Predictions

We use Python’s logging module as a placeholder for recording predictions and other relevant information. However, you can extend this functionality to log additional metrics, such as:

Prediction confidence scores
Prediction sets from conformal methods
Gradients or feature importances

Logging can also integrate with external monitoring tools to facilitate real-time tracking of inference results.