Project from https://github.com/daikiclimate/asahi_changedet
I worked on this project for just 6 weeks, so there are a lot of hard encoding, not refactoring, bugs. This repo is not completed.
pip install addict barbar wandb
you need to install other libraryes if you have import error.
number of samples is so many. (100,000~) It takes really much time to dataloading.
In thie repo, I made HDF5 data format for this data. HDF5 is Hierarchical data structure, it uses dict-like data access fastly. HDF5 data stracture is probably faster than normal data loading.
python dataset/mkhdf5.py
it takes about 1 hour for training data parse.
mkhdf5.py generate hdf_dataset.h5 which include add data like image, label, path, data list as dict-like. mkhdf5.py has a lot of hardencoding, so if you want, add argparse or something. Please change some path.
in asahi_changedet/dataset/asahi_dataset.py There is CLASS Hdf5Dataset which uses hdf_dataset.h5.
python train.py config/config_multi.yaml
I didnt search hyper parameters because I didnt have enough time.
Only evaluation and inference. Need model weight so run after training.
python inference config/config_multi.yaml