mit-han-lab / bevfusion

[ICRA'23] BEVFusion: Multi-Task Multi-Sensor Fusion with Unified Bird's-Eye View Representation
https://bevfusion.mit.edu
Apache License 2.0
2.37k stars 428 forks source link

training strategy #566

Closed liyih closed 7 months ago

liyih commented 11 months ago

Hi, authors! Thanks for your great work. By the way I change some codes of your work and obverse that the model doesn't converge enough in 6 epochs. So, I want to increase the training time (from 6 epochs to 10 or 12 epochs). I want to know what I should change for the Param scheduler (including LinearLR, CosineAnnealingLR, CosineAnnealingMomentum, CosineAnnealingMomentum). Best wish.

186big commented 11 months ago

Hello,I meet a question at #582 ,I want to ask what is the result of model convergence after training 6 epochs with the author's original code, and can it achieve the effect in the author's paper? Thank you , Best wish.

liyih commented 11 months ago

我复现源码的效果和作者的差不多NDS,mAP +- 0.2。我的不收敛发生在我在原代码中加了模块后,但是效果依然正常,只是我觉得第六个epoch效果增长并没有完全停止。问题582中的效果很奇怪,我觉得你可以试着检测一下你生成的data和pickcle是否正确,还有检查一下在load预训练参数的时候有没有出现通道不匹配的情况。你也可以把训练过程中loss的变化发一下,从中可以看出一点端倪。

Li-Whasaka commented 10 months ago

我复现源码的效果和作者的差不多NDS,mAP +- 0.2。我的不收敛发生在我在原代码中加了模块后,但是效果依然正常,只是我觉得第六个epoch效果增长并没有完全停止。问题582中的效果很奇怪,我觉得你可以试着检测一下你生成的data和pickcle是否正确,还有检查一下在load预训练参数的时候有没有出现通道不匹配的情况。你也可以把训练过程中loss的变化发一下,从中可以看出一点端倪。

Could you please share your train log file with me? What is your loss when you achieve the best performance?

186big commented 10 months ago

我复现源码的效果和作者的差不多NDS,mAP +- 0.2。我的不收敛发生在我在原代码中加了模块后,但是效果依然正常,只是我觉得第六个epoch效果增长并没有完全停止。问题582中的效果很奇怪,我觉得你可以试着检测一下你生成的data和pickcle是否正确,还有检查一下在load预训练参数的时候有没有出现通道不匹配的情况。你也可以把训练过程中loss的变化发一下,从中可以看出一点端倪。

Could you please share your train log file with me? What is your loss when you achieve the best performance?

Ok, thank you , I’ll send you later,email?

Li-Whasaka commented 10 months ago

我复现源码的效果和作者的差不多NDS,mAP +- 0.2。我的不收敛发生在我在原代码中加了模块后,但是效果依然正常,只是我觉得第六个epoch效果增长并没有完全停止。问题582中的效果很奇怪,我觉得你可以试着检测一下你生成的data和pickcle是否正确,还有检查一下在load预训练参数的时候有没有出现通道不匹配的情况。你也可以把训练过程中loss的变化发一下,从中可以看出一点端倪。

Could you please share your train log file with me? What is your loss when you achieve the best performance?

Ok, thank you , I’ll send you later,email?

OK,you can attach file in github

186big commented 10 months ago

我复现源码的效果和作者的差不多NDS,mAP +- 0.2。我的不收敛发生在我在原代码中加了模块后,但是效果依然正常,只是我觉得第六个epoch效果增长并没有完全停止。问题582中的效果很奇怪,我觉得你可以试着检测一下你生成的data和pickcle是否正确,还有检查一下在load预训练参数的时候有没有出现通道不匹配的情况。你也可以把训练过程中loss的变化发一下,从中可以看出一点端倪。

Could you please share your train log file with me? What is your loss when you achieve the best performance?

Ok, thank you , I’ll send you later,email?

OK,you can attach file in github

20231225_095320.log.json