Detectron2: When the model is not trained on 8 GPUS but on fewer devices, what should we be careful of?

[ChenJoya]

❓ Questions and Help

Thank you for such a wonderful job!

My question is:
When the model is not trained on 8 GPUS but on fewer devices, such as 4 GPUS and 2 GPUS, what should we be careful of?

e.g., In the original maskrcnn-benchmark, the POST_NMS_TOPK_TRAIN should be set as 1000 * batch_per_gpu.

Thanks for your attention! Look forward to your reply.

[ppwwyyxx]

Just learning rate and batch size. (and of course the steps if batch size is changed

[ChenJoya]

Okay, this simplifies the training. Thanks! @ppwwyyxx