search

madeleinegrunde / AGQA_baselines_code

MIT License
18 stars 4 forks source link

Training time consumption comparison among three baselines #4

Closed AlyssaYoung closed 2 years ago

AlyssaYoung commented 3 years ago

Thanks for providing the AGQA dataset and the exciting insights!

I have also trained these three baselines on my unpublished dataset which was collected by myself. I found that the official HME code needed too much time for one iteration and required much more time comsumption to achieve a good performance than the other two methods. I wonder whether you have also encountered this problem.

madeleinegrunde commented 3 years ago

Hi,

Thank you for reaching out. Yes, we encountered a similar problem that the HME code required much more time than the other two models. Our solution was to let it run for a longer amount of time, despite it being computationally intensive. We saved copies of the model and validated more frequently than the original code to compensate for the long processing time.

AlyssaYoung commented 3 years ago

Thanks for your reply. How many GPUs and how much time did you use to achieve the results in the paper?

madeleinegrunde commented 3 years ago

Hi Alyssa,

We used 1 GPU, and if I remember correctly approximately 1 week.

AlyssaYoung commented 3 years ago

I got it! Thank you very much for your reply!