Can't reproduce the results with grid features on single GPU

[woodfrog]

Can't reproduce the results with grid features on single GPU

[woodfrog]

Hi, I didn't try to use torch.utils.checkpoint to run this model before, but if all other setups were kept the same, this is likely the problem. Did you change any other settings like image resolution, batch size, or the pre-trained BUTD ResNet for the grid-feature experiments?

BTW, did you train the models with region features to reproduce the results? If models with region features work well, then I believe the problem should be the use of torch.utils.checkpoint.

[liuyyy111]

yep, i also train the model with region features, and it works well.
I just remove 'CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 ' in the trian_grid.sh, without any other changes

[woodfrog]

I see. I found this post in Pytorch Forum discussing similar issues, not sure if it can help.

[liuyyy111]

Thank you!!