search

TRAILab / CaDDN

Categorical Depth Distribution Network for Monocular 3D Object Detection (CVPR 2021 Oral)
Apache License 2.0
359 stars 62 forks source link

How many gpus were used ? #60

Closed ksh11023 closed 2 years ago

ksh11023 commented 2 years ago

hi, Thank you for sharing your work.

when training on Kitti benchmark, do you mean by batch size of 4 per GPU? which equals to batch size of 8 when using two GPUs? Am I getting this right..?

codyreading commented 2 years ago

Yup that's correct. I used the V100 GPU (32G), so I was able to fit those batch sizes in memory. You can always decrease the batch size if you need to fit with less GPU memory.

ksh11023 commented 2 years ago

Thank you for your reply!!

So, do I have to use batch size of 8 for Kitti benchmark, to get the results on the paper? I am having trouble getting the same(similar) results using batch size of 4 on Kitti benchmark.

codyreading commented 2 years ago

The results on the paper are all reported with a batch size of 4, so this should be fine for any of your experiments. What results are you getting for your models?

ksh11023 commented 2 years ago

When training on Kitti train set(3712) and validating on Val set(3769)

In the paper it says: Car APr_40 (IOU=0.7) easy: 23.57 mod: 16.31 hard: 13.84

And my results are : Car APr_40(IOU=0.7) easy: 21.92 mod: 15.04 hard:12.68

I used batch size 4, trained from the scratch with other settings same as the paper.

codyreading commented 2 years ago

That level of variance is expected between runs for CaDDN. Please see here for a more in depth discussion about the reproducibility. If you would like to achieve better results, I recommend performing multiple runs and select the one with the best results.

ksh11023 commented 2 years ago

Cody thank you for the swift reply!

I'll follow your advice. Thanks again for sharing your great work:)