CFA for SGG in Pytorch

Our paper Compositional Feature Augmentation for Unbiased Scene Graph Generation has been accepted by ICCV 2023.

Installation

Check INSTALL.md for installation instructions.

Dataset

Check DATASET.md for instructions of dataset preprocessing.

Extract Features

CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.launch --master_port 10032 --nproc_per_node=1 tools/generate_aug_feature.py --config-file "configs/e2e_relation_X_101_32_8_FPN_1x.yaml" MODEL.ROI_RELATION_HEAD.USE_GT_BOX True MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL True MODEL.ROI_RELATION_HEAD.PREDICTOR MotifPredictor  TEST.IMS_PER_BATCH 1 DTYPE "float16" GLOVE_DIR glove MODEL.PRETRAINED_DETECTOR_CKPT checkpoints/pretrained_faster_rcnn/model_final.pth OUTPUT_DIR exp/motif-precls MIXUP.FEAT_PATH feats TYPE extract_aug

Processing Features

python tools/processing_features.py

Training Models with CFA

CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --master_port 10054 --nproc_per_node=2 tools/relation_train_net.py --config-file "configs/e2e_relation_X_101_32_8_FPN_1x.yaml" MODEL.ROI_RELATION_HEAD.USE_GT_BOX True MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False MODEL.ROI_RELATION_HEAD.PREDICTOR MotifPredictor SOLVER.IMS_PER_BATCH 12 TEST.IMS_PER_BATCH 2 DTYPE "float16" SOLVER.MAX_ITER 50000 SOLVER.VAL_PERIOD 2000 SOLVER.CHECKPOINT_PERIOD 2000 GLOVE_DIR glove MODEL.PRETRAINED_DETECTOR_CKPT checkpoints/pretrained_faster_rcnn/model_final.pth OUTPUT_DIR ./exp/motifs_cfa_sgcls TYPE cfa MIXUP.FEAT_PATH feats MIXUP.MIXUP_BG True MIXUP.MIXUP_FG True MIXUP.BG_LAMBDA 0.5 MIXUP.FG_LAMBDA 0.5 MIXUP.PREDICATE_LOSS_TYPE MIXUP_CE MIXUP.MIXUP_ADD_TAIL True FG_TAIL True FG_BODY True BG_TAIL True CL_TAIL True USE_PREDCLS_FEATURE False CONTRA True PKO False

Test Models with Prior Knowledge

CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.launch --master_port 10054 --nproc_per_node=1 tools/relation_test_net.py --config-file "configs/e2e_relation_X_101_32_8_FPN_1x.yaml" MODEL.ROI_RELATION_HEAD.USE_GT_BOX True MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False MODEL.ROI_RELATION_HEAD.PREDICTOR MotifPredictor SOLVER.IMS_PER_BATCH 12 TEST.IMS_PER_BATCH 1 DTYPE "float16" SOLVER.MAX_ITER 50000 SOLVER.VAL_PERIOD 2000 SOLVER.CHECKPOINT_PERIOD 2000 GLOVE_DIR glove MODEL.PRETRAINED_DETECTOR_CKPT checkpoints/pretrained_faster_rcnn/model_final.pth OUTPUT_DIR ./exp/motifs_cfa_sgcls TYPE cfa MIXUP.FEAT_PATH feats MIXUP.MIXUP_BG True MIXUP.MIXUP_FG True MIXUP.BG_LAMBDA 0.5 MIXUP.FG_LAMBDA 0.5 MIXUP.PREDICATE_LOSS_TYPE MIXUP_CE MIXUP.MIXUP_ADD_TAIL True FG_TAIL True FG_BODY True BG_TAIL True CL_TAIL True USE_PREDCLS_FEATURE False CONTRA True PKO True

Comments for Parameters in Command

To make it easier for you to run our code, the Parameters in the command are explained here:

--master_port: It represents the port on which the command is run.
CUDA_VISIBLE_DEVICES: It means the the GPUs that you are going to use. For example, CUDA_VISIBLE_DEVICES=0,1 use the first two GPUs.
--nproc_per_node: It is the number of GPUs you are going to use.
SOLVER.IMS_PER_BATCH: It is the training batch size.
TEST.IMS_PER_BATCH: It is the testing batch size.
SOLVER.MAX_ITER: It is the maximum iteration.
SOLVER.STEPS: It is the steps where we decay the learning rate
SOLVER.VAL_PERIOD: It is the period of conducting val.
SOLVER.CHECKPOINT_PERIOD: It is the period of saving checkpoint.
MODEL.RELATION_ON It means turning on the relationship head or not (since this is the pretraining phase for Faster R-CNN only, we turn off the relationship head), OUTPUT_DIR is the output directory to save checkpoints.
MODEL.ROI_RELATION_HEAD.USE_GT_BOX and MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL: They used to select the protocols, (1) PredCls: They are all set as True. (2) SGCls: MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL is set to False, while MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL True is set to True. (3) SGDet: They are all set to False.
MODEL.ROI_RELATION_HEAD.PREDICTOR: It is the backbobe you are going to use., and the MOTIFS SGG backbone (MotifPredictor) is used by default.
MIXUP.FEAT_PATH: It refers to the path through which features are extracted and processed.
EXTRACT_GROUP: It represents the predicate in which group to extract. The options are head, body, tail, or their combinations, separated by commas.
TYPE: The type of operation. If it set to 'cfa', it represents training with cfa. If it set to 'extract_aug', it represents feature extraction operation
FG_HEAD/FG_BODY/FG_TAIL: It represents whether the Etrinsic-CFA operation for the group's foreground is performed.
BG_HEAD/BG_BODY/BG_TAIL: It represents whether the Etrinsic-CFA operation for the group's background is performed.
CL_HEAD/CL_BODY/CL_TAIL: It represents whether the Intrinsic-CFA operation for the group is performed.
CONTRA: It implies whether to use the contrastive loss.
PKO: It implies whether to use the prior knowledge during the inference.

Models and Generated Files

For the Motifs-CFA, we provide the trained models (checkpoint) for verification purpose. Please download from here* and unzip to checkpoints. Besides, we provide the extracted feature files, you can download from here*.

Citations

If you find this project helps your research, please kindly consider citing our paper in your publications.

@InProceedings{Li_2023_ICCV,
    author    = {Li, Lin and Chen, Guikun and Xiao, Jun and Yang, Yi and Wang, Chunping and Chen, Long},
    title     = {Compositional Feature Augmentation for Unbiased Scene Graph Generation},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {21685-21695}
}

Credits

Our codebase is based on Scene-Graph-Benchmark.pytorch.

guikunchen/CFA