The training process of DETACH

Prepare data

Get into the preparation dir

cd prepare_data

Generate a list of png format preprocessed image data. The generation process includes correct degradation, remove limb brightness, and image histogram equalization.

python -u generate_preprocessed_data.py

Train data

Download pretrained data

mim download mmdet --config mask-rcnn_r50-caffe_fpn_ms-poly-3x_coco --dest ./checkpoints

Download the solar_event_dataset_coco_v0.json from release and place it into the annotation dir.

Generate train and val data

python prepare_data/generate_train_val_coco_json.py

Add environment setting

export CUBLAS_WORKSPACE_CONFIG=:16:8

Train:

python tools/train.py configs/solar_event/convnext_detach.py

Check result

check the prediction of one image

python data_analyse/prediction_example.py

Annotate the data

The annotated file in DETACH is labeled manually by ourself. Here's the preparation process with the help of HEK database.

Get into the preparation dir

cd prepare_data

Download coronal hole segmentation from HEK database

python get_hek_data.py

Generate the coco-format json annotation

python -u generate_coco_mask.py

The default annotation boundary is not accurate, so we need to manually correct the annotation.

We annotated the results by using the CVAT tools. The solar_event_dataset_coco_v0.json in release is our annotated results.

Hyperparameter optimization

The hyperparameter in DETACH is selected semi-automatically by the SMAC algorithm with the help of nni toolkit.

Generate training config file:

python prepare_data/generate_training_config_convnext.py

Auto-ML with nni:

python nni-exp/nni_search.py