How to set the base_learning_rate in finetune autoencoderKL?

[haooxia]

How to set the base_learning_rate in finetune autoencoder? Should I modify the default base_learning_rate: 1.6e-7

It can be seen that the learning rate changes with num_gpus, batchsize and base_learning_rate:

[OwalnutO]

Set --scale_lr=False, then the learning rate will not be modified.

[haooxia]

Set --scale_lr=False, then the learning rate will not be modified.

Thanks for the reply, the base_learning_rate given by the repository is 4.5e-6, so what value should I shrink to when finetune on my own dataset?