antoyang / TubeDETR

[CVPR 2022 Oral] TubeDETR: Spatio-Temporal Video Grounding with Transformers

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

About the training time

liuyang-ict opened this issue · comments

How much V100 GPU-time when TubeDETR was trained on HC-STVG?

20 epochs on HC-STVGv2 took about 7 hours on 16 V100 for the best configuration (stride 4 resolution 352).

Thanks a lot~~

If fewer GPUs are used for training, like 2~8 V100, will the accuracy decrease?
By the way, are your V100 cards all 32GB? Thanks in advance.

By the way again, I think the preproc_hstvgv2.py in preproc should be preproc_hcstvgv2.py? :)

I haven't tried training with a lower number of V100 but I don't think this would decrease the results. All the V100 used in experiments are 32GB, but you can use 16GB V100 with a lower resolution and/or higher temporal stride. The name of the preproc_hstvgv2.py file indeed had a typo, thanks for catching this!