Yongxin-Guo

/

trace-uni

like 3

Hi,

Thank you for your interest in our work. To configure the paths correctly in inference.py, please follow these steps:

• Modify lines 5-7 to reflect the correct paths on your system.
• Update line 26 with the path to the folder where your downloaded checkpoint is located. Please ensure that the trace exists within this path.
• Set num_frames on line 34 to num_frames=128.

These adjustments should help get everything working. If you encounter any issues while running the code, don’t hesitate to reach out, and we’ll be happy to assist you in resolving them.

thanks Guo. This is great response time. I really appreciate it. Model checkpoint will be downloaded from huggingface or there is some other workflow for doing it? Hugging face models do not have .pth format or so
I tried changing paths , and will do that

What is the link for checkpoint? Excuse my knowledge if you see limitations :)

Cheers

Hi,

We apologize for the misunderstanding. The entire huggingface folder is considered the "checkpoint." We provide the checkpoints in safetensors format here, and we did not include .pth files, as they can be sensitive to specific PyTorch versions.

You can directly run the following script to download the checkpoints and use the path of the downloaded folder to replace line 26 in inference.py:

git lfs install

git clone https://huggingface.co/Yongxin-Guo/trace-uni

thanks again. Really proactive approach, thumbs up.

I will try to run and see if I come across any other problems and tell you

the reference to clip-vit is incorrect or update. In checkpoint config file
"mm_vision_tower": "openai/clip-vit-large-patch14-336",
instead of
"mm_vision_tower": "model/clip-vit-large-patch14-336",

You are correct. The model paths in config.json should be updated. We are using our local "model/clip-vit-large-patch14-336" checkpoints here due to privacy concerns. You can either download the checkpoints or use the public version "openai/clip-vit-large-patch14-336."

Additionally, if you are running the training code, please ensure that the paths in the training scripts and system paths are also updated accordingly.

it goes into trace_uni (saying the tensor is on meta, could not be transported to cpu and gpu in a good way)

Could you kindly provide more details about the error you're encountering? Additionally, I’m not entirely sure what you mean by 'on meta.' Would you mind clarifying or providing context for this term? Thank you!

Hi Guo, thanks we were able to run it. Thanks for the great help. I feel it would be great if people provide a docker kind of tested inference. But it has great possibilities, and a very good algorithms that almost comes on par with the commercial algorithms and that too open source. We would like to fine tune it for our purposes.

Would you like to set up a chat.