edengabriel/taskweave

[CVPR 2024 Accepted] TaskWeave: Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection

Hi 👋

• 🎓 Education: I’m currently pursuing a Ph.D candidate.
• 😄 Research Interests: Multimodal learning, Video understanding, Embodied AI, Deep deinforcement learning, Cognitive Science
• ✨ Hobbies: Fitness, Outdoor sports, Tennis, Reading ...

taskweave's Issues

Cannot reproduce the product (Both training and inference)

I trained your model based on all your configurations on a single NVIDIA A40, and got the result that is about 1% lower than yours (As shown in eval.log.txt and my_inference.txt).
Moreover, I used the checkpoint you provided in ./results/best_model_mr/model_best.ckpt and found that the evaluation result on val dataset is also about 1% lower than reported (As shown in ckpt_inference.txt). Besides, that result do not match the evaluation result of any epoch in ./results/best_model_mr/eval.log.txt. That makes me confuse a lot...
My environment is Python 3.7.16 with Pytorch 1.11.0+cu113 and Torchvision 0.12.0+cu113. Basic settings are shown in my_inference.txt and ckpt_inference.txt. Could you please help me figure out the reason? Thanks.
eval.log.txt
my_inference.txt
ckpt_inference.txt

Asking for the link of extracted features

Thank you for your great work!
The original extracted feature from the Moment-DETR repository is now missing, and the author of Moment-DETR said he had lost that file as well.
Have you save this file? Could you share this file with me? Very much thanks!