search

ptuls / tabnet-modified

Modification of TabNet as suggested in the Medium article, "The Unreasonable Ineffectiveness of Deep Learning on Tabular Data"
60 stars 12 forks source link

Encoder-decoder architecture #2

Open astrogilda opened 4 years ago

astrogilda commented 4 years ago

Hey Paul,

Thanks for this repository! Could you help me understand how one would modify the code to function in an encoder decoder capacity, as referenced in the paper?

ptuls commented 4 years ago

Hey @astrogilda,

You'll need to create a new decoder function, similar to the encoder function

def decoder(self, data, reuse, is_training)

You'll have to implement a new loss function as per the paper for self-supervision image in Section 4.

Then, in here, add a decoder after the encoder to train I think. You might want to remove the classification part too.

At least that's the rough idea I have. Hopefully that helps.