RuntimeError because of groundtruth=0 in one of the video+not reproducible result

[jh-yi]

Hi thanks for the efforts and the great works！

I followed all the settings mentioned in previous issues and still get MAE: 0.582 and OBO: 0.283. (pytorch 1.7.0, mmcv1.4.0, mmcv_full1.7.1 cuda 10.1), and I tested with two different machines, also repeated downloading the dataset and preprocessing the MP4 to npz, but still got the same results.

And here is one of the error I met:
RuntimeError: round_vml_cpu not implemented for 'Long' because of the round function in "count = torch.sum(target, dim=1).round().to(device)", line 38 in test_lopping.py.
I found that one of the video has groundtruth of 0, and if I ignore the round function for that video, I got the result as mentioned above.

Could you please release the npz files for validation？

[ZYQbill]

You can just run the vid2npz.py to get the same npz files.