Mmdetection learning rate and batch_ Size relationship

commonly mmdetection The learning rate of the reproduced articles is corresponding config Under the document , yes 0.01 perhaps 0.02.

If you want to reproduce your thesis , however gpu inadequate , For example, there is only one piece gpu, How to set the learning rate

But the learning rate of the original thesis is 0.01,samples_per_gpu=2,gpu_num = 8

Original thesis batch_size = samples_per_gpu * gpu_num =16

that batch_size*LR This value is constant

If you only have one piece gpu,samples_per_gpu=8,batch_size=8, Your new learning rate lr = 0.005

If you have two pieces gpu,sample_per_gpu=2,batch_size=4, So the new learning rate lr=0.025

If you have 4 block gpu,sample_per_gpu=4,batch_size=16, Constant learning rate ,lr=0.01