miranheo / genvis Goto Github PK

I would like to ask how long it will take for Genvis to train on Ovis, that is, to fully follow genvis_R50_bs8_online.yaml,

On the four A100 gpus, it shows that it will take about 28 hours, and I am not sure if this is normal

Dear Author, you said you would release the code in December, but now it is January.

I ran the inference and obtained the following results. How can I visualize them?

How to run the Genvis model with demo.py?

Thanks for your excellent work! When I read the implementation details in your paper, I am a little confused. Do you mean that GenVIS uses the finally trained model of VITA and then trains the proposed module? (It still consumes a lot of resources when training the VIS model.)

Thanks for your excellent work! I wonder when you will release the code.

Hi,
sukjunhwang，

Thanks for your excellent work!

I noticed that for GenVIS, SAMPLING_ FRAME_ SHUFFLE is false, which means that the frame id to be sampled is orderly. For clip training, there may be duplicate images in a clip, and the probability is high. Is there a problem with such sampling?

Dear author. I was very curious that GenVIS was trained only with youtebe2019 and not with COCO dataset. What will the test result be? What should the training configs be changed to?