royelkabetz / citylearn-mad-cats Goto Github PK

Hello all,

First thanks a lot for such inspiring and well documented work.
I'd like to train your solution on another dataset, could you please give me more details on the training procedure so that I can use your implementation ?

Thanks a lot !

I came across your code : net_consumption_mlp/mlp_model.py that includes the following line:

 def forward(self, inp_data):
        # inp_data: [Batch, Input length, Channel]
        seq_last = inp_data[:, -1:].detach()
        output = inp_data - seq_last
        output = nn.Flatten()(inp_data)
        output = nn.ReLU()(self.linear_in(output))
        output = self.BN0(output)
        output = nn.ReLU()(self.linear_h0(output))
        output = self.BN1(output)
        output = nn.ReLU()(self.linear_h1(output))
        output = self.BN2(output)
        output = self.linear_out(output)
        output = output + seq_last[:, :, 0]
        return output

I understand that inp_data represents the input data with shape [Batch, Input length, Channel], and seq_last represents the last time step of the input data with shape [Batch, 1, Channel]. This operation subtracts each time step of the input data from the last time step. I am curious about the purpose of this operation. I would like to know more about the effect of this operation, and whether it has been shown to improve model performance in certain cases. Could you please provide some insight on it?

Thank you for your time and expertise.