Questions about the training hyperparameters

[1349949]

Hi,
I have some questions about the training hyperparameters.

1. What’s the best model’s training batchsize(samples_per_gpu) exactly is in the Table.2 & 3 of your paper? All the training configs you provided in the repo set samples_per_gpu=1, is the same as the best model?

2. The ablation study in the paper using 20% data for training. What about other training hyperparameters? Such as training epochs, batchsize, training 3 classes together or separately.

I’m doing some reproducing experiments, so I need the training hyperparameters mentioned above.

[Abyssaledge]

@1349949 Thanks for using!

1. We use samples_per_gpu=1 for all experiments.
2. For ablation in Table 1, we train 3 classes together for 12 epoch. For other ablations, we train Vehicle and Pedestrian separately. Training 3 classes together will obtain better performances on Pedestrian and Cyclist.

[minwang-ai]

Hi,
I have two questions :

1. Can we only use 20% data for ablation study?
2. May I ask how long it takes to use the whole dataset for training with one frame and three frame?

[Abyssaledge]

@MinWang1997

1. Although I don't know what you're worried about, you can certainly do it using 20% data.
2. ~ 60h for single frame, maybe 80h for 3 frames with 8 2080Ti.

[ziqipang]

@MinWang1997 Just to clarify, 20% data for ablation is only valid for Waymo Open Dataset, since 20% of Waymo Open Dataset is still several times larger than KITTI/nuScenes. That said, using 20% data is only applicable and adopted for Waymo Open Dataset and not others.

[minwang-ai]

Hi Ziqi, Thank you for your clarification ☺️ Do you think it is also applicable to use 20% Waymo data for normal training models and tune hyper parameters?

…

On Mon 14. Feb 2022 at 22:54, Ziqi Pang ***@***.***> wrote: @MinWang1997 <https://github.com/MinWang1997> Just to clarify, 20% data for ablation is only valid for Waymo Open Dataset, since 20% of Waymo Open Dataset is still several times larger than KITTI/nuScenes. That said, using 20% data is only applicable and adopted for Waymo Open Dataset and not others. — Reply to this email directly, view it on GitHub <#10 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AR7Z7VINYSXPYL4XMSOJO5TU3F2Z5ANCNFSM5N4R4URA> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>. You are receiving this because you were mentioned.Message ID: ***@***.***>

[ziqipang]

@MinWang1997 Yes, it is valid to tune hyper-parameters with 20% waymo dataset.

[minwang-ai]

Hi Ziqi, Thank you for your reply!

…

On Tue 15. Feb 2022 at 22:03, Ziqi Pang ***@***.***> wrote: @MinWang1997 <https://github.com/MinWang1997> Yes, it is valid to tune hyper-parameters with 20% waymo dataset. — Reply to this email directly, view it on GitHub <#10 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AR7Z7VPQRAEUSTEUS3MOYBTU3K5QJANCNFSM5N4R4URA> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>. You are receiving this because you were mentioned.Message ID: ***@***.***>