Hello Colin,

I downloaded the Basketball dataset from AWS exchange as documented here: https://github.com/ezhan94/multiagent-programmatic-supervision. However, the dataset seems to only contain a train.npz and a test.npz file, which are not the files that you use in your experiments. Is there an extra preprocessing step involved that I missed?

Thank you in advance.

Hello Colin,

Thank you for releasing the source code of this excellent work. I have some trouble replicating your results for the InD traffic trajectory dataset. I am running the default script you provide to train and evaluate the model and using the results from eval_results_test_last_driver5burnin.txt. The error that I get from these files is ~0.23 at 40 steps (averaged across 5 seeds). However, in your figures, the reported number is ~0.025. Do the reported numbers come from a different experimental setting and if so, how can I access them?

Thank you in advance.

Hi Colin,

Thanks for sharing the code of the excellent work.

When I run the code on motion-118, I notice that the output val loss is negative value (-0.56...). This is caused by the reconstruction loss (nll_gaussian in dnri.py). The normalize_nll is defaultly set to be True so the negative constant is added with the reconstruction loss.

It is not reasonable for me if any loss if negative. I thought it might be caused by different versions of pytorch. Just confirm whether you got negative loss in your result.

Thanks.

Hi, Colin. This is Yandong. I am really impressed with your work. Is that possible to get the same Basketball Dataset from you? It's only for academic purposes. I found this the original dataset from STATS is not available.

Hi Colin,

Thanks for releasing the code of this wonderful work. I was going through dnri_dynamicvars.py and have trouble finding where the reverse LSTM in equation 14 of the paper is implemented. I noticed that DNRI_DynamicVars_Encoder.reverse_rnn is a recurrent layer, but it seems like it isn't used anywhere. Thanks in advance for your help!

Dear author,

Thanks for your wonderful work.

After executing the dNRI of 'run_motion_118.sh', I found the edges don't have zero element. How could I yield the sparsity (no interaction) of interaction networks?

Besides, what does '2 edge types' or '4 edge types' mean in CMU motion data?

Hi,

I have a few trivial questions:

• are the edges in the graph not undirectional, i.e. the edge type for node $i$ to node $j$ could be different than the edge type from node $j$ to node $i$? If that is the case how can we impose symmetry on the connectivity matrix? (for example the same edge type from node $i$ to node $j$ and vice versa)
• how are the edges ordered? to be precise, for example after this line the output is just a 2D vector, whose rows are the one hot encoding of an edge type, with $(N*N - N)$ number of rows, which is due to the fact there is no connection between the node onto itself (please correct me if I'm wrong in any of those statements). But how could one know if for example the $n$-th row is representing the connection between node 1 to node 0 or node 0 to node 1? In other words, if one were to create an adjacency matrix whose elements represent the edge type, e.g.

$$\begin{bmatrix} 0 & 0 & 1\ 2 & 0 & 0\ 3 & 1 & 0 \end{bmatrix},$$

where ${ 0,1,2,3 }$ indicate the edge types. How could that be done using the edges from self.single_step_forward.

Thank you very much in advance.

Hello,

I've noticed that the edges generated by the NRI paper are in a symmetric matrix format, whereas the edges expected by dNRI are in a vector format. Likewise, it seems that -- for n particles -- the vector is of length n * (n-1), which makes sense since each particle's relation to itself is a given. That being the case, if I flatten the matrix (excluding the diagonal values) and pass that in as the vector, the edge predictions are always about a 50-50 guess. Any help would be greatly appreciated.

Thanks.