Sorry to admit, but your code is chaos.

[cosxsinxds]

Hi, I’m sorry to say this, but the current state of the codebase is quite disorganized. Here are a few issues I encountered:

1. There’s a clear inconsistency between the code and the comments in RGNet/feature_extraction/mad_clip_text_extractor.py. The comments mention processing MAD, but the actual code is operating on Ego4D.

2. The instructions in data/DATASET.md and feature_extraction/README.md also conflict with each other. Both describe how to preprocess MAD, but using different methods. It's unclear why both are necessary or which one is the correct pipeline.

3. The file utils/init.py is missing. This causes significant debugging overhead, as PyCharm and users can see files like utils/basic_utils.py, but Python itself cannot resolve the import, leading to errors such as ModuleNotFoundError: No module named 'utils.basic_utils'.

4. cone/train.py mentioned in rgnet/scripts/train_mad.sh is missing. Even the floder named cone is missing.

If the results in your paper were genuinely achieved without shortcuts, I’d strongly recommend going through your codebase at least once to ensure everything is coherent and reproducible.

[Condor-G]

@cosxsinxds Hello, sorry to bother you. I’ve seen your comments under several projects, so I’d like to ask for your advice. Have you ever successfully run the code for RGNet, ReVisonLLM, or VTimeLLM? I’ve been trying to reproduce ReVisonLLM recently. I’ve fixed many bugs, but I still can’t get it to work, and I’m not sure if it’s an environment issue on my side.

[cosxsinxds]

@Condor-G, I read your note recently. I gave up trying to reproduce RGNet, ReVisonLLM long time ago. For VTimeLLM, I forgot whether I figured it out.