Problem of warmup_ratio

[Icecream-blue-sky]

Prerequisite

• I have searched Issues and Discussions but cannot get the expected help.
• I have read the FAQ documentation but cannot get the expected help.
• The bug has not been fixed in the latest version (master) or latest version (3.x).

Task

I'm using the official example scripts/configs for the officially supported tasks/models/datasets.

Branch

master branch https://github.com/open-mmlab/mmdetection

Environment

python3

Reproduces the problem - code sample

The warmup_ratio means "the ratio of the starting learning rate used for warmup". But I find that if I set the lr to 0.1 and warmup_ratio to 0.001, the actual starting learning rate used for warmup is not 0.001 * 0.1, why? How to caculate the actual starting learning rate?

Reproduces the problem - command or script

None

Reproduces the problem - error message

None

Additional information

None

[ZwwWayne]

This is because the log shows the LR at the 50th iteration rather than the first iteration.

[Icecream-blue-sky]

This is because the log shows the LR at the 50th iteration rather than the first iteration.

Set x1=1, y1=0.001*0.1 (the learning rate of the first iteration), x2=1000, y2=0.01 (the initial learning rate), so the function of linear warmming up is y = 9.91*10^(-6)*x + 9.009*10^(-5). So when x=50 (the 50th iteration), the learning rate y should be 5.8559*10^(-4), which is different from the one in the log, why?

[ZwwWayne]

The calculation is implemented here: https://github.com/open-mmlab/mmcv/blob/master/mmcv/runner/hooks/lr_updater.py#L87