Inconsistency

[pjyazdian]

Hi, thank you for the great job. I trained the model and the result is not similar to what you reported in your paper. Also, I see you redistribute HumanML3D which in fact needs licence. Also, I see that it is not exactly same as HumanMl3D and have several extra samples with "w_<HumanML3D_ID>".
Could you please provide more info?
Thanks!

[mingyuan-zhang]

Thanks for your interest in our work.

1. May I know what's your model's performance now? One possible reason may be smaller batchsize. In our experiments, we use 8 gpus to achieve a large batchsize, which is essential to achieve higher accuracy.

2. Thanks for your pointing out. I have deleted the related files. In next week I will provide scripts about preparing this dataset.

3. In HumanML3D dataset, each motion data may be assigned with different text annotation and different time intervals. We follow the official preparation to regard them as different items and store them explicitly. You may find the official implementation from this link