Ablation study on model fusion

Hi.
Given the attached architecture, the model is already trained and has the weights loaded. I would like to perform an ablation study by disabling or removing one component (example: text). after some research I have found that I can create an empty input containing zeros and fed it to the network but I am not sure about that.

Please can you share some doc or tell me how to do it.
I don’t want to create the model from scratch that handle the tree components (scene, face and audio)

Thank you

Hi @youb, As you are trying to perform ablation study to know how important a feature is for the model performance. As your model has 4 inputs and while inference and you need to test with 3 inputs and exclude one input and for that input passing zero values may add bias.

You can try by adding some noise to the zero values, or modify the input by adding/removing/replacing with someother values. Please refer to the feature ablation in this document. Thank You.