--batch vs --batch-gpu

[nadavpo]

in the configs.md you mentioned the --batch-gpu parameter.
you also show example of running with 1 gpu where you set the --batch to 32 and the --batch-gpu to 16. what effect it have? if the batch size is 32, you restrict the samples per batch per gpu to 16 and you have only one gpu, doesn't that mean that your batch size is 16?

[PDillis]

Hello, I know this is an old question, but what they did is gradient accumulation: say you want to do a backward pass on a batch size of 32, but you can only fit a batch of 16 on your current GPU. So you accumulate two forward passes, then do the backward pass. It's this reason why, if you don't specify --batch-gpu and ony --batch, then this batch size will be divided by the number of GPUs you are using. Hope this helps.