search

intel / intel-xai-tools

Explainable AI Tooling (XAI). XAI is used to discover and explain a model's prediction in a way that is interpretable to the user. Relevant information in the dataset, feature-set, and model's algorithms are exposed.
Apache License 2.0
37 stars 6 forks source link

pytorch multipoint regression model cards #3

Open mkurczew opened 1 month ago

mkurczew commented 1 month ago

Hi! I went over the notebooks and some of the code for the Model Card Generator, but it seems that all three examples are classifiers.

What to do in case where I would like to create a model card for a multi-point regressor? For example, let's assume that we have a torch model that outputs 21 2-D facial landmarks. Let's say that I would like to show that NME(normalized mean error) is the same e.g., across all geos or races.

The make_eval_dataframe() method in model_card_gen/intel_ai_safety/model_card_gen/analyze/torch_analyzer.py assumes that there's a single numerical column representing labels and predictions so that probably wouldn't work for 21x2 numbers.

Any hints on how to tackle this kind of model with your toolkit?

tybrs commented 1 month ago

@mkurczew Thanks for the comment! We had a similar issue with multi-label support with pytorch. The workaround we received can be found here: https://github.com/tensorflow/model-analysis/issues/162

In short i think this could be solvable by creating a custom beam.Pipeline to generate evaluation results.