Jerry🐭

I met the same error as you did. Hope for solutions!

> > @MobiusLqm oh sorry, that variable is only used to keep the gradient w.r.t. uv, the content itself is still zero. The values are stored in this variable https://github.com/graphdeco-inria/diff-gaussian-rasterization/blob/main/cuda_rasterizer/rasterizer_impl.cu#L265...

Hi! I have the same question with you. Have you figured it out?

我在微调后用cli_demo.py运行模型也出现了和楼主差不多的问题： 而且在无法进行正常的对话，会出现```nan in logits, set nan_default_token to proceed in BaseStrategy.forward.```的错误提示，请问有什么解决办法么？

> You can use the preprocessing code from NeuMan (see https://github.com/apple/ml-neuman/tree/main/preprocess) for most of the preprocessing but HUGS also needs some preprocessing from 4D-Humans (see issue 6: #6) Thanks for...

> The output structure of the PHALP pkl file is described [here](https://github.com/brjathu/PHALP/blob/677074a9bd7acac58c0b98a31b04ae54b93dcd2f/README.md?plain=1#L158). > > When you run the `track.py` demo, you use the PHALP tracking algorithm. PHALP has its own...

> Do you want to use HMR2.0 with a different tracking algorithm, or do you want to use PHALP with a different 3D human pose estimation system? Sorry for missing...

The issue turned out to be that the learning rate was too small, which made the parameters barely trainable. Never trust the official fine-tuning documentation that suggests a learning rate...