Setting Train/Validation Split for an MoE Model

[kjrstory]

I have a question about setting up the train and validation splits for a "MoE Gating Network for External Aerodynamics". Is there a recommended guideline? Specifically, should I reuse the same train/validation sets that were used to train DoMINO, FigConvNet, and X-MeshGraphNet(XAeroNet)?

In my current pipelines, VTP files are generated only for the validation cases of those three models. For MoE training, I would need to generate VTPs for the training cases as well. Does this approach make sense, or is there a better practice you recommend?

[Dibyajyoti-Chakraborty]

It is not necessary to use the same train/val set that was used for training the individual experts. For example, in some corner cases, it may not be recommended if a model works really well for training but not on validation. Here, the MoE may learn to put a larger weight on that particular expert. In cases where the models consistently show similar skill in training vs validation, it does not matter.
So, feel free to use your particular mix of training and validation as long as they are independent and the training does not influence the validation in any way. I suggest checking the individual model performance in your train-val split. They should be consistent. For example, if expert 1 is relatively 10% better than expert 2 in the train set on average, it should be similar in the val set too.
Although it will be interesting to see what happens if this consistency is not present. Hope it answers your question.