Unofficial application of SOCO:Aligning Pretraining for Detection via Object-Level Contrastive Learning. This application includes all tricks used in SOCO.
Setup Env: requirements.txt.
Use ./selective-search to generate selective-search proposals.
waiting
cd ./model_zoo_exp/imagenet1k_scratch_sgd_224_p4_r50_512/
./slurm_train.sh spring_scheduler 32
First, use soco2detectron.py to convert the checkpoint to detectron.
Then, train faster-rcnn-fpn on detectron2 with the same setting with detco.
| Method | Setting | VOC-AP50 | VOC-AP50 |
|---|---|---|---|
| SOCO | bs512, syncBN, sgd, lr0.4, epoch100, weight decay$1e^{-4}$, v3:224 | 55.7656 | 82.03 |
| SOCO | bs2048, syncBN, lars, lr8, epoch400, weight decay$1e^{-5}$, v3:96 | 56.50 | 82.51 |