DucTeacher/
└── datasets/
└── haitian/
├── train/
├── unlabel/
├── val/
└── annotations/
├── instances_train.json
├── instances_unlabel_0.json
└── instances_val.json
install detectron2 in ./detectron2
bash.sh for DucTeacher
get the pre-train model in ./output
mv 'merge_domain_8/merge_0.json' '/cache/data/haitian/annotations/instance_unlabel_0.json'
python train_net.py --num-gpus 8 --config configs/haitian_supervision/faster_rcnn_R_50_FPN_sup_run1.yaml SOLVER.IMG_PER_BATCH_LABEL 16 SOLVER.IMG_PER_BATCH_UNLABEL 16 SEMISUPNET.PARA_MU 0.1 SEMISUPNET.PARA_T 0.7
get the Evaluation results in ./output/infrence/coco_instances_results.json
python train_net.py --num-gpus 8 --eval-only --config configs/haitian_supervision/faster_rcnn_R_50_FPN_sup_run1.yaml SOLVER.IMG_PER_BATCH_LABEL 16 SOLVER.IMG_PER_BATCH_UNLABEL 16
input : ./output/infrence/coco_instances_results.json output : Domain Similarity & Estimated Class Distribution
python get_domain_similarity_class_distribution.py
after getting the domain similarity and class distribution, we can train the DucTeacher.
python DucTeacher_domain_evolve_sh_0_10.py