Why choose the last column of Q0, Q1 and g

[zyxnlp]

Hi Reid, thanks for your implementation. I have two questions about the code.

I was wondering why the last column of Q0, Q1, and g was selected in the code of lines 165-167.

Q0 = sm(Q_logits_T0)[:, 1]
Q1 = sm(Q_logits_T1)[:, 1]
g = sm(g)[:, 1]

Would that be equal to calculate the P(Y=1 |T=0, C, text) and P(Y=1 |T=1, C, text) and P(T=1 | C, text)?

However, according to the definition in the paper, what we should calculate are P(Y|T=0, C, text), P(Y |T=1, C, text), and P(T | C, text) ? Please correct me if I'm wrong.

Another question is why the Y labels of T=0 and T=1 should be set to -100 when calculating the cross-entropy of Q1 and Q0?

T0_indices = (T == 0).nonzero().squeeze()
Y_T1_labels = Y.clone().scatter(0, T0_indices, -100)
T1_indices = (T == 1).nonzero().squeeze()
Y_T0_labels = Y.clone().scatter(0, T1_indices, -100)
Q_loss_T1 = CrossEntropyLoss()(Q_logits_T1.view(-1, self.num_labels), Y_T1_labels)
Q_loss_T0 = CrossEntropyLoss()(Q_logits_T0.view(-1, self.num_labels), Y_T0_labels)

I'm a freshman in causal inference. Really appreciated it if you can help me out!