Hi,
I have a question regarding the predicate-to-entity message propagation mechanism mentioned in the paper. I couldn't locate its actual implementation in the code. Specifically, there's a comment in lines 282-283 of the file pysgg/modeling/roi_heads/relation_head/model_msg_passing.py that states: "currently no redo classification on embedding representation, we just use the first stage object prediction."
Could you clarify whether this mechanism is not implemented despite being discussed in the paper, or if I might be missing something? When I retrieve class predictions from both FasterRCNN and BGNN networks, I get identical results, which might be due to the lack of refinement in class predictions.
Thank you in advance for your assistance.
Hi,
I have a question regarding the predicate-to-entity message propagation mechanism mentioned in the paper. I couldn't locate its actual implementation in the code. Specifically, there's a comment in lines 282-283 of the file pysgg/modeling/roi_heads/relation_head/model_msg_passing.py that states: "currently no redo classification on embedding representation, we just use the first stage object prediction."
Could you clarify whether this mechanism is not implemented despite being discussed in the paper, or if I might be missing something? When I retrieve class predictions from both FasterRCNN and BGNN networks, I get identical results, which might be due to the lack of refinement in class predictions.
Thank you in advance for your assistance.