tsinghua-fib-lab/LCSim

mismatched shape of the agent

Len-Li opened this issue · 4 comments

Len-Li commented

Hi,

Thanks for open-sourcing the awesome project!

I encounter an issue when training the diffusion model. The shape is mismatched in here.

assert (
            self.num_historical_steps + self.num_future_steps
            == data["agent"]["xyz"].shape[1]
        )

Here is the dimension of the agent:
    
print(data["agent"]["xyz"].shape, self.num_historical_steps + self.num_future_steps)

torch.Size([50, 198, 3]) 91                                                                                                                          
torch.Size([70, 198, 3]) 91                                                                                                                          
torch.Size([129, 199, 3]) 91                                                                                                                         
torch.Size([26, 199, 3]) 91                                                                                                                          
torch.Size([44, 198, 3]) 91                                                                                                                          
torch.Size([41, 198, 3]) 91

It seems like we should add a sliding window to crop the time dimension of input data in the dataloader?

Thanks!

lionel-me commented

It seems that you used the 20s scenarios of WOMD. As mentioned in the readme, our diffusion model is trained on the 9s scenarios of WOMD, which were selected from the 20s scenarios by Waymo company by adding a sliding window.

Len-Li commented

It seems that you used the 20s scenarios of WOMD. As mentioned in the readme, our diffusion model is trained on the 9s scenarios of WOMD, which were selected from the 20s scenarios by Waymo company by adding a sliding window.

Thanks for your reply. I will add a sliding window to it.

Len-Li commented

by the way, how to make a visualization like this on your website? The animation is very cool.
image

lionel-me commented

LCSim provides tools for rendering the scenario (see example here. By specifying returen_his = True in the sampling function, it will return the whole intermediate results of the denoising process. Then, you can draw these intermediate results in different figures and make a video by mediapy.

  • 👍1