Junbin Xiao comments

Results 55 comments of


                                            Junbin Xiao

How to get the word representations

Thx for pointing it out. The code follows the official repo, and yes that the attention-pooling is changed to sum-pooling here. I will fix it soon.

Hi, I wanna know how to set specific frames per second?

Hi, please refer to frame rate in skvideo or ffmpeg..

Hi, can you pleaese release the pre-extracted TGIF feature?

Hi, you can obtain them from [VGT](https://github.com/sail-sg/VGT). The features are similar and can help you reproduce a close results.

Features of MSRVTT-QA

Hi, please find the feature ![here](https://drive.google.com/drive/folders/1Kfo_TE1UYE3uX6MjkiNPZSO3s2NxAeYQ?usp=sharing)

About the box region feature for NExT-QA

Hi, I have given it in the readme..

About the BERT feature

Thanks for your interest. Please find the edited code for bert finetuning [here](https://drive.google.com/file/d/1_wwJrB7r974Eq3VkXUnUlPhBwQRDLMbZ/view).

how much GPU memory does the next QA need?

Thanks for the question. It needs about 24G for training of the model with batch size 64 with 8 clips for each video, whereas 8G is enough for inference. If...

how much GPU memory does the next QA need?

Hi, please find the [code and mode](https://drive.google.com/file/d/1Z0RMnIJrqQcFQEhhuhiRvDvyRmaKdPbg/view?usp=sharing) for BERT finetuning/feature extraction. You should launch a new issue with a proper title, otherwise I may slow in finding your questions..

Visualization

Thanks for your interest, we use this [tool](https://github.com/visjs/vis-network) to visualize the learned graphs.

Visualization

No such guideline :), and you need to format the learned adj matrices to adapt to the tool. Basically, you need a JS file like this: ``` var nodes =...