Function to Extract VST Feature Vectors

[zhiqing0205]

Amazing work on Video-XL! The model's ability to handle long videos efficiently and the use of Visual Summarization Tokens (VSTs) is truly impressive. It's exciting to see how well it balances accuracy and efficiency.

I’m currently exploring downstream tasks using Video-XL and was wondering if you could provide a function to directly extract the VST feature vectors. This would make it much easier to integrate the model into other applications like classification or event detection.

If such a feature is feasible, it would be incredibly helpful for further exploration and practical use of your model. Thanks again for your excellent work and looking forward to your response!

[shuyansy]

Thank you so much! I think the function is feasible. Just give me some time to realize it.

[shuyansy]

Hi Thanks for your patient waiting. I have implemented the function "get_memory" in videoxl/model/Language_model/modeling_beacon.py. Please check that. To use it in your inference phase, you can run

past_key_values = model.memory.get_memory()

If you have further issues, please let me know.

[zhiqing0205]

Thanks so much for implementing this! I really appreciate your quick response and effort. I'll check it out and let you know if I have any questions. Thanks again!