More about motion weights

[mxlin03]

Hi, I've noticed that the weights assigned to motions significantly impact the performance of ASE in downstream tasks. Therefore, I'm keen to understand more about how these weights are set.

You mentioned in a previous response:
'The weights adjust the sampling probability of data from different motion files to balance the variety of motions in the dataset. For instance, if there is an abundance of one type of motion data, it's necessary to adjust the weights of other types to ensure the model doesn't overemphasize certain skills. For example, if you have twice as much data for skill A (like walking left turn, which also includes aspects of skill B, such as walking) compared to skill B, you might assign a weight of 1 to motions for skill A and a weight of 2 for skill B.'

However, I'm puzzled about the scenario where skill A includes elements of skill B, like in the case of walking left turn versus just walking. How should the weights be set in such instances?

Could you provide a specific calculation process, for example, how the weight of 0.03105590 for the file 'RL_Avatar_Atk_Kick_Motion.npy' was determined?

[xbpeng]

It is possible that there are overlaps in the skills between different motion clips, but we don't need to analyze the clips in that much detail when setting the weights. The weights in the dataset file are determined by grouping the clips into similar skills and then assigning each group of skills weights so that the sampling process will roughly sample the same amount of data across different skills.