search

Cuogeihong / CEASC

The official implementation of CEASC
Apache License 2.0
99 stars 12 forks source link

How to set the lr? #13

Closed hefangnan closed 11 months ago

hefangnan commented 12 months ago

great work! Iwant to know If I had 2 gpu,each gpu batch size is 8,then how should I set the lr?

Cuogeihong commented 12 months ago

According to our experimnets, keeping original lr 0.01 is ok. you can also try 0.005, 0.0025 or 0.00125 for instance

hefangnan commented 12 months ago

According to our experimnets, keeping original lr 0.01 is ok. you can also try 0.005, 0.0025 or 0.00125 for instance

感谢你的快速回复，我将学习率设置为了0.02，但在训练的时候出现了mmdet - ERROR - The testing results of the whole dataset is empty.请问这是不是和学习率设置过大有关？ Snipaste_2023-09-10_21-19-45

Cuogeihong commented 12 months ago

是的，0.02有一点大了。但这个错误也可能是数据集的问题。你可以使用论文的设置，将每个gpu的batchsize改成2，用0.01的lr训练，避免学习率问题

hefangnan commented 12 months ago

是的，0.02有一点大了。但这个错误也可能是数据集的问题。你可以使用论文的设置，将每个gpu的batchsize改成2，用0.01的lr训练，避免学习率问题

谢谢您的建议，我把学习率调为了0.005就正常了