how to train from scratch（如何从头开始训练网络）

[PanffeeReal]

Hi，I'd like to ask how to revise the config file when I want to train a detection model such as FCOS from scratch, that is, don't use the pretrained model and initialize the backbone parameters randomly.

你好，我想请教如何配置 config文件，实现从头训练检测模型，即不使用预训练模型参数，而采用例如随即的方式初始化特征提取器backbone的参数？

期待您的回复～

[viktorho]

Not a pro, but in this repository, they trying to make everything in a pipeline, if you tried to create something from the scratch, if your model didn't require to make any customize with these pipelines mentioned in this docs
mmegine: https://mmengine.readthedocs.io/en/latest/design/runner.html#customize-runner
then you need to create step-by-step all of these:
First, create a dataset, dataloader; if it is a detection model, mmdet repository could help
mmdet: https://mmdetection.readthedocs.io/en/latest/
then replicate your model, similar to Pytorch
Ex: https://mmengine.readthedocs.io/en/latest/tutorials/model.html
after that, is optim-wraper and exe,...
make sure to registry all of them on the same framework, sometime when you import mmdet instead of mmengine can caused import issues.
Good luck,