Can I use pytorch-like training process instead of scikit-compatible way of use this lib?

[kaiwang0112006]

I want to do research on tabnet with federated learning, which means I need to get the model weight out and set it back during each epoch of training. It would be easier with a pytorch-like epoch training process of using this lib instead of scikit-compatible way of training.

[Optimox]

Hello,
I am not sure that I understand your request. But if you want to use the tabnet network simply as a pytorch module and insert it inside your own pipeline you can simply used the modules from here:

tabnet/pytorch_tabnet/tab_network.py

Line 508 in 2c0c4eb

class TabNet(torch.nn.Module):

[kaiwang0112006]

That;s great! What should I give to the parameter "group_attention_matrix" ?

[Optimox]

This is an advanced feature, you can leave to None, otherwise you'll need to dig a bit into the code to use it. It's just a matrix of weights on how the attention can work across different features.