P3Former icon indicating copy to clipboard operation
P3Former copied to clipboard

Published 20 hours ago • verified publisher icon OpenRobotLab

nuScenes Dataset

Open aselimc opened this issue 11 months ago • 19 comments

Hello,

Although in #4, this was discussed, I'd like to ask once more about any plans to release nuScenes code instructions?

Furthermore, can you please share more details about the GPU memory requirement (I've seen you have used 8xA100 but do you use all 80gb memory of A100s) and time (in hours/days) to train for nuScenes?

Best

aselimc avatar Mar 01 '24 11:03 aselimc

Hi. thanks for your interest in our methods. Since I'm busy with other topics and not able to catch up upgradation of mmdet3d, I will try to release the nuscenes codebase with the old version of mmdet3d this month.

xizaoqu avatar Mar 04 '24 07:03 xizaoqu

Hi, thanks for the answer and I'm glad you'll be sharing nuScenes codebase. And can you share details about the train time and GPU memory?

aselimc avatar Mar 04 '24 10:03 aselimc

Hi, thanks for the answer and I'm glad you'll be sharing nuScenes codebase. And can you share details about the train time and GPU memory?

It takes ~15 minutes to train one epoch on the 8 GPU * 4 batch size setting. I find that the PQ increases slowly even more than 100 epochs. So I keep training for about 200 epochs before the PQ is more than 75, which takes around 2 days. But I'm not sure about the exact memory for now. Need to check later.

xizaoqu avatar Mar 04 '24 11:03 xizaoqu

20221107_153449.log Here is one training log.

xizaoqu avatar Mar 04 '24 11:03 xizaoqu

I suppose memory:14956 refers to 14.956 GB ?

aselimc avatar Mar 04 '24 11:03 aselimc

I suppose memory:14956 refers to 14.956 GB ?

I don't know what it represents, but it can not be 15GB Cuda memory.

xizaoqu avatar Mar 04 '24 13:03 xizaoqu

Hello,

Although in #4, this was discussed, I'd like to ask once more about any plans to release nuScenes code instructions?

Furthermore, can you please share more details about the GPU memory requirement (I've seen you have used 8xA100 but do you use all 80gb memory of A100s) and time (in hours/days) to train for nuScenes?

Best

Hi, the NuScenes codebase is released, feel free to try it out.

xizaoqu avatar Apr 01 '24 10:04 xizaoqu

Thank you for the release? Do you also plan to share the config file related to nuScenes ?

aselimc avatar Apr 02 '24 11:04 aselimc

Thank you for the release? Do you also plan to share the config file related to nuScenes ?

Please checkout the nuscene_oldversion_det3d branch.

xizaoqu avatar Apr 02 '24 12:04 xizaoqu

Thank you so much for releasing the codebase. I'm training it now using your config, I won't close this issue until I confirm the results.

aselimc avatar Apr 03 '24 11:04 aselimc

Thank you so much for releasing the codebase. I'm training it now using your config, I won't close this issue until I confirm the results.

You may also try to train it based on pretrained semantic models like https://github.com/cardwing/Codes-for-PVKD. It helps to stabilize the results.

xizaoqu avatar Apr 03 '24 12:04 xizaoqu

Thank you for the release? Do you also plan to share the config file related to nuScenes ?

Please checkout the nuscene_oldversion_det3d branch. @xizaoqu I have a question: is the mask_hungarian_assigner the same as the new version of mmdet used in the main branch and Semantickitti?

RONINGOD avatar Apr 23 '24 08:04 RONINGOD

Hi @xizaoqu

Why the PQ results from the pretrained model (75.1%) is little bit smaller than the results reported in the paper (75.9%)?

Thank you

hphnngcquan avatar Jun 25 '24 16:06 hphnngcquan

Hi @aselimc,

Have you trained the nuscenes code? Can you please share with me your train log from the first epoch?

Thank you

hphnngcquan avatar Jun 27 '24 00:06 hphnngcquan

Have you trained the nuscenes code? Can you please share with me your train log from the first epoch?

Hi, it is the full log from the first epoch (note: we use pretrained semantic segmentation models from https://github.com/cardwing/Codes-for-PVKD for initialization).

20221106_001005.log

xizaoqu avatar Jun 27 '24 02:06 xizaoqu

Hi @xizaoqu

Why the PQ results from the pretrained model (75.1%) is little bit smaller than the results reported in the paper (75.9%)?

We have trained a new model implemented the new version of mmdet3d, but accidentally deleted by the cluster server. So maybe I do not have time to recover it.

xizaoqu avatar Jun 27 '24 02:06 xizaoqu

Thank you for the release? Do you also plan to share the config file related to nuScenes ?

Please checkout the nuscene_oldversion_det3d branch. @xizaoqu I have a question: is the mask_hungarian_assigner the same as the new version of mmdet used in the main branch and Semantickitti?

Hi, maybe you can directly compare the file. I'm not sure if the new version has any change.

xizaoqu avatar Jun 27 '24 02:06 xizaoqu

thank you very much @xizaoqu

hphnngcquan avatar Jun 27 '24 03:06 hphnngcquan

Have you trained the nuscenes code? Can you please share with me your train log from the first epoch?

Hi, it is the full log from the first epoch (note: we use pretrained semantic segmentation models from https://github.com/cardwing/Codes-for-PVKD for initialization).

20221106_001005.log

Hi @xizaoqu , As i reckon, this is the log from the previous version of code got deleted in cluster server. There are some parameters i do not see in the current code

hphnngcquan avatar Jul 03 '24 02:07 hphnngcquan