Batchsize and lr of HTC+Swin-L

[BinZhu-ece]

Thank you for your outstanding work, I only have 8 A-100 cards, can I train a batchsize each card and adjust the learning rate to 0.0001 / 2? Is there a negative impact on this relatively large model accuracy of HTC+swin-L?

[liming-ai]

Thank you for your outstanding work, I only have 8 A-100 cards, can I train a batchsize each card and adjust the learning rate to 0.0001 / 2? Is there a negative impact on this relatively large model accuracy of HTC+swin-L?

Hi, @962858249

Could you run the code successfully? There are a lot of bugs when I tried to run the config of HTC + Swin-L

[clw5180]

Thank you for your outstanding work, I only have 8 A-100 cards, can I train a batchsize each card and adjust the learning rate to 0.0001 / 2? Is there a negative impact on this relatively large model accuracy of HTC+swin-L?

only用得好