self.saved_searching_ternimal_state = self.root_state_tensor.clone()[self.lego_indices.view(-1), :].view(self.num_envs, 108, 13) RuntimeError: shape '[1, 108, 13]' is invalid for input of size 1716

[xiaobai0906]

Hi,
When I try to train each sub-task by 'python train_rlgames.py --task=BlockAssemblyOrient --num_envs=1024'.
An error has occurred：
"self.saved_searching_ternimal_state = self.root_state_tensor.clone()[self.lego_indices.view(-1), :].view(self.num_envs, 108, 13)
RuntimeError: shape '[1, 108, 13]' is invalid for input of size 1716."
How can I solve it?
Thanks!

[cypypccpy]

Hi @xiaobai0906 ,
This error is due to the environment dimension of the search terminal state you saved is different from the orient task. It seems that you are using one environment's searching_ternimal_state, and you are running 1024 environments' orient, so you also need to use the corresponding 1024-dimensional searching_ternimal_state.
Hope this can help you.

[RalphFH]

@cypypccpy @xiaobai0906
hi~ I am replicating this paper as a baseline method and have also encountered this problem. Here 1716 is equal to 132 x 13, and 132 means the block nums loaded in the search task. It seems to have nothing to do with the env_num.