Thank you for your contribution. After 20 rounds of training on scenefrlow, and then fine-tuning 20 rounds, is it still not convergence? I added some extra epochs, and the EPE will be lower than 0.64

gangweiX / CGI-Stereo

A novel neural network architecture that can concurrently achieve real-time performance, competitive accuracy, and strong generalization ability.

MIT License

307 stars 28 forks source link

Closed 1434205229 closed 1 year ago

gangweiX commented 1 year ago

Yes, more training epochs can bring better results.

1434205229 commented 1 year ago

thank you

flyxuexi commented 3 months ago

thank you

哥，请问你训练的环境和作者完全一样吗，谢谢

1434205229 commented 3 months ago

差不多

flyxuexi commented 3 months ago

差不多

请问是在3090上训练的吗，epe能够到0.64或更低吗？我在4090上训练，epe收敛在了0.67....

1434205229 commented 3 months ago

差不多

请问是在3090上训练的吗，epe能够到0.64或更低吗？我在4090上训练，epe收敛在了0.67.... 我的是A6000 我当时可能也没到达0.64，具体不太记得了。没有到0.64你就多训练几回合就行。作者的意思是在sceneflow上训练20+20个epoch

flyxuexi commented 3 months ago

差不多

请问是在3090上训练的吗，epe能够到0.64或更低吗？我在4090上训练，epe收敛在了0.67.... 我的是A6000 我当时可能也没到达0.64，具体不太记得了。没有到0.64你就多训练几回合就行。作者的意思是在sceneflow上训练20+20个epoch

好的谢谢