不同纬度的action数据如何一起训练

[skyoux]

开放的两个预训练模型，貌似都是action_dim=16的，自采数据如果action_dim不是16的话，都需要初始化 action_adaptor、state_adaptor和输出层，请问一下，不同纬度的action数据如何一起训练？

[skyoux]

代码中似乎没有看到padding的操作

[ModiShi]

Yes, we didn't implement data padding. You can implement a custom padding function inheriting DataTransformFn and register it in the config.

[ModiShi]

If the action dimensions in your dataset are consistent, you can simply initialize the action adaptor and state adaptor, then finetune on your own data directly.

[skyoux]

Okay, thanks for your reply. @ModiShi