How to inference next item in sequential model CORE?

[Jeriousman]

I see CORE has predict, full_sort_predict functions. But they all need the label data (the very last item) which is test_item_emb = self.item_embedding.weight from interaction[self.ITEM_ID]which is again, the last item (label data).
But to make real inference function, we should not have supervising last item in our function but solely have input sequential data. So what should inference function look like? Can you please provide it?

I tried

def inference(self, interaction):
  item_seq = interaction[self.ITEM_SEQ]
  seq_output = self.forward(item_seq)
  values, indices = torch.topk(scores, k=3)  ##just to get 3 top most probably next item candidates
  return indices

But I am not sure just using self.forward is a right way to do inference. Please guide me with code. Thank you.

[Jeriousman]

I think I have wrongly understood the functions. Now new question is, I think full_sort_predict is checking input sequence with all available items so that we can inference item among them. So my question is,

1. I guess full_sort_predict is the inference function I was looking for. Is that right? Just the model is not inductive as embeddings of all items must be learned again? So model is transductive?
2. But then why do we have predict function that does with last items of each sequence? What is the purpose of it?

And, from

https://recbole.io/evaluation.html

I see Acceleration Strategy for top-k Evaluation.

3. Is there any way to inference last item with topk of the functions? Any code?

Thank you.