MrClownC

“回到凑零钱问题，为什么说它符合最优子结构呢？假设你有面值为 1, 2, 5 的硬币，你想求 amount = 11 时的最少硬币数（原问题），如果你知道凑出 amount = 10, 9, 6 的最少硬币数（子问题），你只需要把子问题的答案加一（再选一枚面值为 1 的硬币），求个最小值，就是原问题的答案。因为硬币的数量是没有限制的，所以子问题之间没有相互制，是互相独立的。” 括号里的“再选一枚面值为1的硬币”，感觉描述有些问题，应该是在子问题的基础之上选一枚能满足原问题的面值硬币，比如10,9,6最少硬币数是子问题6的硬币数目，那么这里应该选面值为5的硬币。虽然我能理解作者能表达的意思~但第一次读到这还是困惑了很久。

> actions depend on arm_action_mode and I found out the actions from arm mode JointVelocity and JointPosition can be found in observations. For other arm modes, you should find out...

I have the same problem. Do you have a solution now?

> There is another problem. As you can see the name `checkpoint_gripper_post_hist_1_aug_10_4_traj_cons_ws_12_mpt_dolly_3b_2.pth`, this ckpt is trained on `mpt_dolly_3b`, but you load `mpt_dolly_1b`, can you try `mpt_dolly_3b`? I have a question...

> no, MPT-3b-IFT is MPT-3b with instruction fine-tuning I have some confusion. I see in the readme document of the project that the longest model of "AvgLen" is model 3B,...

> Sorry, you are right, my mistake, `checkpoint_gripper_post_hist_1_aug_10_4_traj_cons_ws_12_mpt_dolly_3b_2.pth` corresponds to the language model `mpt_dolly_1b` Ok, I understand. Thank you for your quick reply

Thank you for reminding me, I will try. Another thing that makes me confused is about the setting of the parameter `--cross_attn_every_n_layers`. As you can see in the picture, the...

> Thanks for reminding me. You are right. The arg is replaced by the one defined in the dict, so `--cross_attn_every_n_layers 4` does not make any difference. I apologize for...

> thanks! I tried with --cross_attn_every_n_layers 4 but it's not working, I didn't use pt_eval_ckpts.bash since I didn't have a multi-gpu environment, but now I have a server with several...