CUDA out of memory. 1080Ti 11G how can i less the memory ?

[lightingstars]

[lightingstars]

both train and test , both all out of memory

[cheng peng]

you can set the IMS_PER_BATCH smaller in yaml,
and change BASE_LR linearly according to your modification

[lightingstars]

BASE_LR

thank you very much
do you mean that if i change IMS_PER_BATCH 64 to 16
then BASE_LR should change from 0.2 to 0.05

[cheng peng]

yes！
Absolutely right！

[lightingstars]

yes！
Absolutely right！

thank you very much
have a good day!