How to prepare prediction dataset for manifold

[imneonizer]

Summary

I am not very much familiar with react, but i somehow managed to run manifold demo example inside the container, though i have to modify package file to run it on 0.0.0.0 to be accessible outside the container.

I visualized the provided Regression demo dataset, and its working fine. Now I wanted to use manifold to debug my object detection model, I am confused about preparing the csv files for input I.e., what should I possibly fill inside feature columns since these models takes in pixel values as features.

Context and Environment

• Docker version 19.03.8
• Base Image: ubuntu 18.04.3 LTS Bionic Beaver
• node v13.7.0
• npm 6.13.6
• manifold 1.1.4

Suggestion / Request

please provide an example on usecase of manifold with image classification and object detection model, how to prepare the data for demo application, and which parameters are effective in debugging through manifold.

[gnavvy]

@imneonizer, the open-source version of Manifold only supports categorical, numeric, and geospatial features. We have limited bandwidth to push for image/video/textual features at the moment, unfortunately.

My hunch is that you can try to use the output from the softmax layer as the numeric proxy of the input images, and see which class ranked the top (largest discrepancy) given the two selected subsets with Manifold. Or you can try other tools (e.g. shap) that give you pixel-level attention maps.

Hope it can give you some insights. 🤞