yujun-shi / blip Goto Github PK

Hi,

I am trying this very interesting and novel approach on some experiments with sparse-reward environments. In my particular case I am making use of a QMLPBase policy. Training over a task goes fine, but I get an error at the post processing with agent.ng_post_processing(rollouts, task_idx) as follows:

AttributeError: 'Linear_Q' object has no attribute '_costate'

Apparently the error has its origin at the estimate_fisher() function, where supposedly a backward hook is registered for the Linear_Q layers on both actor and critic networks (as defined under QMLPBase). It seems that this backward hook does not get triggered for the critic network, therefore the attribute _costate never gets created. This subsequently raises an error at the update_fisher_exact() function, as this function expects the _costate attribute to exist for every Linear_Q layer when calculating H = (f_state*m._costate).sum().

Any suggestions on how to overcome this issue?

Many thanks

老师，您好：
请教您修改哪部分代码，才使得您的BLIP能够在我自己的数据集上运行？或者是能给出一个数据加载的思路？

谢谢您！