Question about training configuration setting.

[BB88Lee]

Hi, thanks so much for sharing such great work!

I have some questions about the parameter settings of the experiment. In the paper, the training epoch is 12 with an initial LR 1e-4, but the epoch is 24 in code with an initial LR 2e-4.

Do all the experiments (Table 1 & Table 2) in the paper under the setup mentioned in the paper (epoch 12)? (I understand that CBGS will extend the training time) Could you report an experimental config from the paper? One might expect to do a comparison experiment with you in the same experimental configuration.

The released code with epoch 24 and initial LR 2e-4, are there any results under such a setup? How much can it improve over when set epoch to 12 with an initial LR 1e-4?

[lxtGH]

@WangYueFt Hi! I have the same question about that.

[zhby99]

same question here

[a1600012888]

Hi, Thanks for point it out. Sorry for the late reply

For the experiments without CBGS, the 1e-4 LR and 12 epochs are typos in the paper. In most of the experiment we used 2e-4 LR and 24 epochs.

I didn't remember the detailed performance difference between these two setup, but training the detection models longer can boost the performance a lot, especially using no pretrain models. If training DETR3D with only ImageNet pretrained model, you can continue training the models to 48, 72 epochs to still observe some performance improvments.

Thanks!

[exiawsh]

Hi, Thanks for point it out. Sorry for the late reply

For the experiments without CBGS, the 1e-4 LR and 12 epochs are typos in the paper. In most of the experiment we used 2e-4 LR and 24 epochs.

I didn't remember the detailed performance difference between these two setup, but training the detection models longer can boost the performance a lot, especially using no pretrain models. If training DETR3D with only ImageNet pretrained model, you can continue training the models to 48, 72 epochs to still observe some performance improvments.

Thanks!

Hello, thanks for your great work! Did you use cbgs in your experiment? 24 epochs are trained by default, and no cbgs is the default setting?