A Appendix

We provide the hyperparameters of our baselines and those of FQA in this section. All code was written in Python 3.6 with neural network architectures defined and trained using PyTorch v1.0.0. We adapted the authors' official repository from Network based on GRU-style recurrence followed by a Graph Net decoder. This neighborhood size is also the same as the distance cutoff used in section 4.3. We used 8 attention heads to match the number of FQA's decisions.