multi-gpu modify batchsize

[shangyihao]

when i use Multi-GPU training in pytorch lightning, like Trainer(gpus=[1,2], accelerator='dp') with

what or where can i do to modify the effective batch size for distributed training?
thx

[Borda]

you have set batch_size already, if you want to adjust the size according to actual GPU, you can use auto_scale_batch_size
Moreover, I see that your train_dataloder returns a sampler instead of dataloader instance (or list f instances)