I find the prompt loss function (reg_loss) is not consistent with the original paper (Equ. 14). The original paper uses softmax over pairs of graph representation and class prototypical subgraph representation, whereas the code uses cross-entropy between predictions class and ground truth. This discrepancy does not seem to unify the pre-training loss and fine-tuning loss functions.
Additionally, which equation in the paper corresponds to the bp_loss loss?
The above code is located in prompt_fewshot.py.
I find the prompt loss function (reg_loss) is not consistent with the original paper (Equ. 14). The original paper uses softmax over pairs of graph representation and class prototypical subgraph representation, whereas the code uses cross-entropy between predictions class and ground truth. This discrepancy does not seem to unify the pre-training loss and fine-tuning loss functions.
Additionally, which equation in the paper corresponds to the bp_loss loss?