The computation way of 'step_scale' seems to ambiguous?

roytseng-tw / Detectron.pytorch

A pytorch implementation of Detectron. Both training from scratch and inferring directly from pretrained Detectron weights are available.

MIT License

2.82k stars 565 forks source link

The computation way of 'step_scale' seems to ambiguous? #169

Open RanTaimu opened 5 years ago

RanTaimu commented 5 years ago

https://github.com/roytseng-tw/Detectron.pytorch/blob/8315af319cd29b8884a7c0382c4700a96bf35bbc/tools/train_net_step.py#L193

Why you use original_batch_size instead of original_batch_size * original_iter_size ? Is the reason that you ignore the 'original_iter_size' because its default value is 1?