Skip to content

Latest commit

 

History

History
43 lines (30 loc) · 1.8 KB

README.md

File metadata and controls

43 lines (30 loc) · 1.8 KB
Raw

Transferring to Object Detection

We use detectron2 to train the object detection models.

  1. Install detectron2.

  2. Put dataset under "benchmarks/detection/datasets" directory, following the directory structure requried by detectron2.

  3. Convert the pre-trained backbone weights to detectron2's format:

    cd benchmarks/detection
WEIGHT_FILE=../../work_dirs/selfsup/densecl/densecl_coco_800ep/extracted_densecl_coco_800ep.pth
OUTPUT_FILE=extracted_densecl_coco_800ep.pkl
python convert-pretrain-to-detectron2.py ${WEIGHT_FILE} ${OUTPUT_FILE}

  4. Start training:

    DET_CFG=configs/pascal_voc_R_50_C4_24k_moco.yaml
bash run.sh ${DET_CFG} ${OUTPUT_FILE} output/run_1_voc_R_50_C4_24k_densecl_coco_800ep

Transferring to Semantic Segmentation

We use mmsegmentation to train the semantic segmentation models.

  1. Clone and install our modified mmsegmentation.

  2. Prepare the datasets according to dataset_prepare.md.

  3. Modify the pretrained model path in config.

  4. Start training (here we use Cityscapes for example):

    CONFIG_FILE=configs/densecl/fcn_r50-d8_769x769_40k_cityscapes.py
GPUS=4
OUTPUT_DIR=models/fcn_r50-d8_769x769_40k_cityscapes_densecl_coco_800ep
./tools/dist_train.sh ${CONFIG_FILE} ${GPUS} --work-dir ${OUTPUT_DIR}