About the result of test.py

[abduallahmohamed]

About the result of test.py

[abduallahmohamed]

Hi,

I just rand the code:
here's my log;
can you make sure the you copied the repo correctly with the model weights?
the number of samples is 20 ?

You can notice a difference between my log and yours, it seems that the models checkpoints didn't load on your side correctly;

[zhanghuishun]

@abduallahmohamed Thanks you for reply!
I still have a question about the kernel size of st_gcn in model.py . I saw that you set graph convolving kernel size = 8 which is observe seq length. I dont understand that.
Another question is that I would like to know if you have tried attention mechanism. Because I read the ST-GCN paper you reference, they put a attention in their network. Thanks again!

[abduallahmohamed]

For obs len = 8, we needed a features from each pedestrians trajectory that is able to match the adjacency matrix A. take a look into this einsum:

Social-STGCNN/model.py

Line 74 in ebd57aa

x = torch.einsum('nkctv,kvw->nctw', (x, A))

Say for example we have 2 pedestrains observed over 8 time steps, the A will have dime of (8,2,2), the vertices have dim of (2,8,2) in order to have the einsum that represents the ST-GCNN to work we needed a graph kernel of 8.

I tested the attention mechanism and it wasn't promising on my side, but I could say it was a naiive test and out of the scope for this work.

[zhanghuishun]

Thanks a lot!