Jie Liu

1. This is the corresponding "battle won mean" in PyMARL. However, it reaches 80% win rate for 3s5z.  2. Why do you used training data (which is not greedy)...

I'm also able to resolve this by adding import numba; import torch; import habitat; in the first line in run.py You may need to rearrange the three " import numba;...

> Thanks @yifan123, but this wouldn't be sustainable -- in that I'll have to make this change in every habitat project file I encounter in the future. I'm also able...

Thank you very much for your passion on our work. Our hand-printed pictures come from Taobao comments. You can crawling them by web crawler or by hand manually. That doesn't...

The shallow and deep layers of the network require slightly different dilation. If you only want to use a single ICConv2d, I suggest simply manually setting the ratio of different...

I add the following code in the https://github.com/lm-sys/FastChat/blob/main/fastchat/llm_judge/gen_model_answer.py#L97, and it works. ``` if isinstance(model, PeftModelForCausalLM): model = model.merge_and_unload() ```

Hi, I also cross check the CQL scores reported in D4RL (arXiv-v4). 1. The mismatches has not been fixed @IcarusWizard @justinjfu @aviralkumar2907 2. In Table2 and Table3 of D4RL, there...

Cool! I have another questions: 1. I train ResNet18 and ResNet50 on ImageNet and cifar10 using EKFAC. However, the test accuracy is just as good as the sgd mothod with...

I'm using Batch Normalization. Changing eps will hurt the performance according to my experiment. Thanks for your reply😀， I will try it.