using pytorch lightning for training object detection models in mmdet.
- mmdet implementation of datasets and network is cool but going through their training/inference code is a mess (atleast for me). so this repo tries to decorate code using pytorch lightning.
- This is still experimental and contains lot of bugs. For instance. mmdet uses an iterbased or epochbased learner and they have optimizer and schedular defined accordingly. I have made some changes here as I didn't find a way to embed an iter+epoch (steplr(epoch) + warmup(iter)) based schedular in pytorch lightning yet.
- clone the repo
- do
poetry install - To contribute, also do
pre-commit install
- dataset and dataloader
- define model
- optimizers setup
- training_step and logger
- validation_step and logger
- validation_epoch_end (coco utils checks)
- train the model for few epochs
- write inference pipeline
- write test pipeline to calculate the stats.