How to train on 4 3090 GPUs?

[atti0127]

For 'Restormer/train.sh', I simply changed --nproc_per_node to 4

In case of 'RealDenoising_Restormer.yml', how should I change? Can simply change num_gpu to 4, but have no idea how to handle 'batch_size_per_gpu' and 'mini_batch_sizes'. Default is 8 and [8,5,4,2,1,1], respectively.

Could anyone share the experimental results using 4 gpus?

[Mytttttttt]

这是来自QQ邮箱的假期自动回复邮件。你好，我最近正在休假中，无法亲自回复你的邮件。我将在假期结束后，尽快给你回复。

[atti0127]

这是来自QQ邮箱的假期自动回复邮件。你好，我最近正在休假中，无法亲自回复你的邮件。我将在假期结束后，尽快给你回复。

Still on vacation?

[Mytttttttt]

这是来自QQ邮箱的假期自动回复邮件。你好，我最近正在休假中，无法亲自回复你的邮件。我将在假期结束后，尽快给你回复。