Awesome

Structured Sparse R-CNN for Direct Scene Graph Generation

Our paper Structured Sparse R-CNN for Direct Scene Graph Generation has been accepted by CVPR 2022.

Requirements

Environments: python 3.8, cuda 10.1, pytorch 1.7.1

To install requirements:

conda create --name scene_graph_benchmark
conda activate scene_graph_benchmark

pip install --user ipython
pip install --user scipy
pip install --user h5py
pip install --user pyyaml
pip install --user yacs
pip install --user scipy
pip install --user h5py
pip install --user tqdm
pip install --user opencv-python

pip install --user ninja yacs cython matplotlib tqdm opencv-python overrides


conda install pytorch==1.7.1 torchvision==0.8.2 cudatoolkit=10.1 -c pytorch

git clone https://github.com/cocodataset/cocoapi.git
cd cocoapi/PythonAPI
python setup.py build_ext install

git clone https://github.com/NVIDIA/apex.git
cd apex
python setup.py install --cuda_ext --cpp_ext

# put our code SGGbench into your directory, such as /home/username
cd /home/username/SGGbench 
python setup.py build develop

Datasets

VG

For the dataset and pretrained backbone weights preparation, please follow: Scene-Graph-Benchmark.pytorch

OI v4, v6

For the datasets and pretrained backbone weights preparation, please follow: BGNN-SGG. Actually, BGNN-SGG is also compatible with Scene-Graph-Benchmark.pytorch.

Training

VG

On VG, we notice ~0.2 R@100 noise for our model with 300 queries.

To train the model with 300 triplet queries, run this command. The results are R@100: 36.9; zR@100: 3.7; mR@100: 10.0: (8 RTX 2080ti 11G)

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m torch.distributed.launch --master_port 10020 --nproc_per_node=8 tools/relation_train_net.py --config-file "configs/simrel_e2e_relation_X_101_32_8_FPN_1x.yaml" MODEL.ROI_RELATION_HEAD.USE_GT_BOX False MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False MODEL.ROI_RELATION_HEAD.PREDICTOR CausalAnalysisPredictor MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_TYPE none MODEL.ROI_RELATION_HEAD.CAUSAL.FUSION_TYPE sum MODEL.ROI_RELATION_HEAD.CAUSAL.CONTEXT_LAYER motifs SOLVER.IMS_PER_BATCH 8 TEST.IMS_PER_BATCH 8 SOLVER.VAL_PERIOD 12000 SOLVER.MAX_ITER 80000 SOLVER.STEPS '(47000, 64000)' SOLVER.BASE_LR 0.000008 SOLVER.OPTIMIZER "ADAMW" MODEL.PRETRAINED_DETECTOR_CKPT checkpoints/pretrained_faster_rcnn/model_final.pth OUTPUT_DIR Outputs/real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel MODEL.ROI_BOX_HEAD.MLP_HEAD_DIM 1024 MODEL.SimrelRCNN.ENABLE_BG_OBJ False MODEL.SimrelRCNN.ENABLE_REL_X2Y True MODEL.SimrelRCNN.REL_DIM 256 MODEL.SimrelRCNN.NUM_PROPOSALS 300 MODEL.SimrelRCNN.NUM_HEADS 6 MODEL.SimrelRCNN.REL_STACK_NUM 6 MODEL.SimrelRCNN.TRIPLET_MASK_WEIGHT 1.0 MODEL.SimrelRCNN.FREEZE_BACKBONE True MODEL.SimrelRCNN.CROSS_OBJ_FEAT_FUSION False MODEL.SimrelRCNN.CLASS_WEIGHT 1.333 MODEL.SimrelRCNN.L1_WEIGHT 5.0 MODEL.SimrelRCNN.GIOU_WEIGHT 2.0 GLOVE_DIR '' MODEL.SimrelRCNN.ENABLE_FREQ False MODEL.SimrelRCNN.ENABLE_QUERY_REVERSE False MODEL.SimrelRCNN.USE_REFINE_OBJ_FEATURE True MODEL.SimrelRCNN.FREEZE_PUREE_OBJDET False MODEL.SimrelRCNN.PURE_ENT_NUM_PROPOSALS 100 MODEL.SimrelRCNN.ENABLE_MASK_BRANCH False MODEL.SimrelRCNN.KL_BRANCH_WEIGHT 0.25 MODEL.SimrelRCNN.ENABLE_KL_BRANCH True MODEL.SimrelRCNN.POSI_ENCODE_DIM 64 MODEL.SimrelRCNN.REL_CLASS_WEIGHT 1.334 MODEL.SimrelRCNN.PURE_ENT_CLASS_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_GIOU_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_L1_WEIGHT 5.0 MODEL.SimrelRCNN.AUXILIARY_BRANCH True MODEL.SimrelRCNN.AUXILIARY_BRANCH_SELECT_ENT_MAX_NUM 25 MODEL.SimrelRCNN.AUXILIARY_BRANCH_START 11 MODEL.SimrelRCNN.ENABLE_ENT_PROP True MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_ANALYSIS False MODEL.SimrelRCNN.USE_CROSS_RANK False MODEL.SimrelRCNN.DISABLE_KQ_FUSION_SELFATTEN False MODEL.SimrelRCNN.DIM_ENT_PRE_CLS 256 MODEL.SimrelRCNN.DIM_ENT_PRE_REG 256 MODEL.SimrelRCNN.ONE_REL_CONV True MODEL.SimrelRCNN.ENABLE_BATCH_REDUCTION True MODEL.SimrelRCNN.USE_HARD_LABEL_KLMATCH True MODEL.SimrelRCNN.DISABLE_OBJ2REL_LOSS True

To train the model with 800 triplet queries, run this command. The results are R@100: 38.4; zR@100: 4.0; mR@100: 10.3: (8 V100 32G)

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m torch.distributed.launch --master_port 10020 --nproc_per_node=8 tools/relation_train_net.py --config-file "configs/simrel_e2e_relation_X_101_32_8_FPN_1x.yaml" MODEL.ROI_RELATION_HEAD.USE_GT_BOX False MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False MODEL.ROI_RELATION_HEAD.PREDICTOR CausalAnalysisPredictor MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_TYPE none MODEL.ROI_RELATION_HEAD.CAUSAL.FUSION_TYPE sum MODEL.ROI_RELATION_HEAD.CAUSAL.CONTEXT_LAYER motifs SOLVER.IMS_PER_BATCH 8 TEST.IMS_PER_BATCH 8 SOLVER.VAL_PERIOD 12000 SOLVER.MAX_ITER 80000 SOLVER.STEPS '(47000, 64000)' SOLVER.BASE_LR 0.000008 SOLVER.OPTIMIZER "ADAMW" MODEL.PRETRAINED_DETECTOR_CKPT checkpoints/pretrained_faster_rcnn/model_final.pth OUTPUT_DIR Outputs/real025kl_newe2rposition_fullobjbranch_800q_smalldyconv_prequeryaug9_norel_hardlabel MODEL.ROI_BOX_HEAD.MLP_HEAD_DIM 1024 MODEL.SimrelRCNN.ENABLE_BG_OBJ False MODEL.SimrelRCNN.ENABLE_REL_X2Y True MODEL.SimrelRCNN.REL_DIM 256 MODEL.SimrelRCNN.NUM_PROPOSALS 800 MODEL.SimrelRCNN.NUM_HEADS 6 MODEL.SimrelRCNN.REL_STACK_NUM 6 MODEL.SimrelRCNN.TRIPLET_MASK_WEIGHT 1.0 MODEL.SimrelRCNN.FREEZE_BACKBONE True MODEL.SimrelRCNN.CROSS_OBJ_FEAT_FUSION False MODEL.SimrelRCNN.CLASS_WEIGHT 1.333 MODEL.SimrelRCNN.L1_WEIGHT 5.0 MODEL.SimrelRCNN.GIOU_WEIGHT 2.0 GLOVE_DIR '' MODEL.SimrelRCNN.ENABLE_FREQ False MODEL.SimrelRCNN.ENABLE_QUERY_REVERSE False MODEL.SimrelRCNN.USE_REFINE_OBJ_FEATURE True MODEL.SimrelRCNN.FREEZE_PUREE_OBJDET False MODEL.SimrelRCNN.PURE_ENT_NUM_PROPOSALS 100 MODEL.SimrelRCNN.ENABLE_MASK_BRANCH False MODEL.SimrelRCNN.KL_BRANCH_WEIGHT 0.25 MODEL.SimrelRCNN.ENABLE_KL_BRANCH True MODEL.SimrelRCNN.POSI_ENCODE_DIM 64 MODEL.SimrelRCNN.REL_CLASS_WEIGHT 1.334 MODEL.SimrelRCNN.PURE_ENT_CLASS_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_GIOU_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_L1_WEIGHT 5.0 MODEL.SimrelRCNN.AUXILIARY_BRANCH True MODEL.SimrelRCNN.AUXILIARY_BRANCH_SELECT_ENT_MAX_NUM 25 MODEL.SimrelRCNN.AUXILIARY_BRANCH_START 11 MODEL.SimrelRCNN.ENABLE_ENT_PROP True MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_ANALYSIS False MODEL.SimrelRCNN.USE_CROSS_RANK False MODEL.SimrelRCNN.DISABLE_KQ_FUSION_SELFATTEN False MODEL.SimrelRCNN.DIM_ENT_PRE_CLS 256 MODEL.SimrelRCNN.DIM_ENT_PRE_REG 256 MODEL.SimrelRCNN.ONE_REL_CONV True MODEL.SimrelRCNN.ENABLE_BATCH_REDUCTION True MODEL.SimrelRCNN.USE_HARD_LABEL_KLMATCH True MODEL.SimrelRCNN.DISABLE_OBJ2REL_LOSS True

To train the model with 300 triplet queries and backbone, run this command. The results are R@100: 36.7; zR@100: 3.8; mR@100: 10.1: (8 V100 32G)

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m torch.distributed.launch --master_port 10020 --nproc_per_node=8 tools/relation_train_net.py --config-file "configs/simrel_e2e_relation_X_101_32_8_FPN_1x.yaml" MODEL.ROI_RELATION_HEAD.USE_GT_BOX False MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False MODEL.ROI_RELATION_HEAD.PREDICTOR CausalAnalysisPredictor MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_TYPE none MODEL.ROI_RELATION_HEAD.CAUSAL.FUSION_TYPE sum MODEL.ROI_RELATION_HEAD.CAUSAL.CONTEXT_LAYER motifs SOLVER.IMS_PER_BATCH 8 TEST.IMS_PER_BATCH 8 SOLVER.VAL_PERIOD 12000 SOLVER.MAX_ITER 80000 SOLVER.STEPS '(47000, 64000)' SOLVER.BASE_LR 0.000008 SOLVER.OPTIMIZER "ADAMW" SOLVER.BACKBONE_MULTIPLIER 0.1 MODEL.PRETRAINED_DETECTOR_CKPT checkpoints/pretrained_faster_rcnn/model_final.pth OUTPUT_DIR Outputs/real025kl_newe2rposition_fullobjbranch_800q_smalldyconv_prequeryaug9_norel_hardlabel_bkb MODEL.ROI_BOX_HEAD.MLP_HEAD_DIM 1024 MODEL.SimrelRCNN.ENABLE_BG_OBJ False MODEL.SimrelRCNN.ENABLE_REL_X2Y True MODEL.SimrelRCNN.REL_DIM 256 MODEL.SimrelRCNN.NUM_PROPOSALS 300 MODEL.SimrelRCNN.NUM_HEADS 6 MODEL.SimrelRCNN.REL_STACK_NUM 6 MODEL.SimrelRCNN.TRIPLET_MASK_WEIGHT 1.0 MODEL.SimrelRCNN.FREEZE_BACKBONE False MODEL.SimrelRCNN.CROSS_OBJ_FEAT_FUSION False MODEL.SimrelRCNN.CLASS_WEIGHT 1.333 MODEL.SimrelRCNN.L1_WEIGHT 5.0 MODEL.SimrelRCNN.GIOU_WEIGHT 2.0 GLOVE_DIR '' MODEL.SimrelRCNN.ENABLE_FREQ False MODEL.SimrelRCNN.ENABLE_QUERY_REVERSE False MODEL.SimrelRCNN.USE_REFINE_OBJ_FEATURE True MODEL.SimrelRCNN.FREEZE_PUREE_OBJDET False MODEL.SimrelRCNN.PURE_ENT_NUM_PROPOSALS 100 MODEL.SimrelRCNN.ENABLE_MASK_BRANCH False MODEL.SimrelRCNN.KL_BRANCH_WEIGHT 0.25 MODEL.SimrelRCNN.ENABLE_KL_BRANCH True MODEL.SimrelRCNN.POSI_ENCODE_DIM 64 MODEL.SimrelRCNN.REL_CLASS_WEIGHT 1.334 MODEL.SimrelRCNN.PURE_ENT_CLASS_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_GIOU_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_L1_WEIGHT 5.0 MODEL.SimrelRCNN.AUXILIARY_BRANCH True MODEL.SimrelRCNN.AUXILIARY_BRANCH_SELECT_ENT_MAX_NUM 25 MODEL.SimrelRCNN.AUXILIARY_BRANCH_START 11 MODEL.SimrelRCNN.ENABLE_ENT_PROP True MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_ANALYSIS False MODEL.SimrelRCNN.USE_CROSS_RANK False MODEL.SimrelRCNN.DISABLE_KQ_FUSION_SELFATTEN False MODEL.SimrelRCNN.DIM_ENT_PRE_CLS 256 MODEL.SimrelRCNN.DIM_ENT_PRE_REG 256 MODEL.SimrelRCNN.ONE_REL_CONV True MODEL.SimrelRCNN.ENABLE_BATCH_REDUCTION True MODEL.SimrelRCNN.USE_HARD_LABEL_KLMATCH True MODEL.SimrelRCNN.DISABLE_OBJ2REL_LOSS True

OI V4

To train the model with 300 triplet queries in the paper, run this command: (8 RTX 2080ti 11G)

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m torch.distributed.launch --master_port 10020 --nproc_per_node=8 tools/relation_train_net.py --config-file "configs/oiv4_simrel_e2e_relation_X_101_32_8_FPN_1x.yaml" MODEL.ROI_RELATION_HEAD.USE_GT_BOX False MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False MODEL.ROI_RELATION_HEAD.PREDICTOR CausalAnalysisPredictor MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_TYPE none MODEL.ROI_RELATION_HEAD.CAUSAL.FUSION_TYPE sum MODEL.ROI_RELATION_HEAD.CAUSAL.CONTEXT_LAYER motifs SOLVER.IMS_PER_BATCH 8 TEST.IMS_PER_BATCH 8 SOLVER.VAL_PERIOD 12000 SOLVER.MAX_ITER 80000 SOLVER.STEPS '(47000, 64000)' SOLVER.BASE_LR 0.000008 SOLVER.OPTIMIZER "ADAMW" MODEL.PRETRAINED_DETECTOR_CKPT checkpoints/iov4_pretrain_faster_rcnn/model_final.pth OUTPUT_DIR Outputs/oiv4_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel MODEL.ROI_BOX_HEAD.MLP_HEAD_DIM 1024 MODEL.SimrelRCNN.ENABLE_BG_OBJ False MODEL.SimrelRCNN.ENABLE_REL_X2Y True MODEL.SimrelRCNN.REL_DIM 256 MODEL.SimrelRCNN.NUM_PROPOSALS 300 MODEL.SimrelRCNN.NUM_HEADS 6 MODEL.SimrelRCNN.REL_STACK_NUM 6 MODEL.SimrelRCNN.TRIPLET_MASK_WEIGHT 1.0 MODEL.SimrelRCNN.FREEZE_BACKBONE True MODEL.SimrelRCNN.CROSS_OBJ_FEAT_FUSION False MODEL.SimrelRCNN.CLASS_WEIGHT 1.333 MODEL.SimrelRCNN.L1_WEIGHT 5.0 MODEL.SimrelRCNN.GIOU_WEIGHT 2.0 GLOVE_DIR '' MODEL.SimrelRCNN.ENABLE_FREQ False MODEL.SimrelRCNN.ENABLE_QUERY_REVERSE False MODEL.SimrelRCNN.USE_REFINE_OBJ_FEATURE True MODEL.SimrelRCNN.FREEZE_PUREE_OBJDET False MODEL.SimrelRCNN.PURE_ENT_NUM_PROPOSALS 100 MODEL.SimrelRCNN.ENABLE_MASK_BRANCH False MODEL.SimrelRCNN.KL_BRANCH_WEIGHT 0.25 MODEL.SimrelRCNN.ENABLE_KL_BRANCH True MODEL.SimrelRCNN.POSI_ENCODE_DIM 64 MODEL.SimrelRCNN.REL_CLASS_WEIGHT 1.334 MODEL.SimrelRCNN.PURE_ENT_CLASS_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_GIOU_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_L1_WEIGHT 5.0 MODEL.SimrelRCNN.AUXILIARY_BRANCH True MODEL.SimrelRCNN.AUXILIARY_BRANCH_SELECT_ENT_MAX_NUM 25 MODEL.SimrelRCNN.AUXILIARY_BRANCH_START 11 MODEL.SimrelRCNN.ENABLE_ENT_PROP True MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_ANALYSIS False MODEL.SimrelRCNN.USE_CROSS_RANK False MODEL.SimrelRCNN.DISABLE_KQ_FUSION_SELFATTEN False MODEL.SimrelRCNN.DIM_ENT_PRE_CLS 1024 MODEL.SimrelRCNN.DIM_ENT_PRE_REG 256 MODEL.SimrelRCNN.ONE_REL_CONV True MODEL.SimrelRCNN.ENABLE_BATCH_REDUCTION True MODEL.SimrelRCNN.USE_HARD_LABEL_KLMATCH True MODEL.SimrelRCNN.DISABLE_OBJ2REL_LOSS True

OI V6

To train the model with 300 triplet queries in the paper, run this command: (8 RTX 2080ti 11G)

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m torch.distributed.launch --master_port 10020 --nproc_per_node=8 tools/relation_train_net.py --config-file "configs/oiv6_simrel_e2e_relation_X_101_32_8_FPN_1x.yaml" MODEL.ROI_RELATION_HEAD.USE_GT_BOX False MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False MODEL.ROI_RELATION_HEAD.PREDICTOR CausalAnalysisPredictor MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_TYPE none MODEL.ROI_RELATION_HEAD.CAUSAL.FUSION_TYPE sum MODEL.ROI_RELATION_HEAD.CAUSAL.CONTEXT_LAYER motifs SOLVER.IMS_PER_BATCH 8 TEST.IMS_PER_BATCH 8 SOLVER.VAL_PERIOD 12000 SOLVER.MAX_ITER 80000 SOLVER.STEPS '(47000, 64000)' SOLVER.BASE_LR 0.000008 SOLVER.OPTIMIZER "ADAMW" MODEL.PRETRAINED_DETECTOR_CKPT checkpoints/iov6_pretrain_faster_rcnn/model_final.pth OUTPUT_DIR Outputs/oiv6_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel MODEL.ROI_BOX_HEAD.MLP_HEAD_DIM 1024 MODEL.SimrelRCNN.ENABLE_BG_OBJ False MODEL.SimrelRCNN.ENABLE_REL_X2Y True MODEL.SimrelRCNN.REL_DIM 256 MODEL.SimrelRCNN.NUM_PROPOSALS 300 MODEL.SimrelRCNN.NUM_HEADS 6 MODEL.SimrelRCNN.REL_STACK_NUM 6 MODEL.SimrelRCNN.TRIPLET_MASK_WEIGHT 1.0 MODEL.SimrelRCNN.FREEZE_BACKBONE True MODEL.SimrelRCNN.CROSS_OBJ_FEAT_FUSION False MODEL.SimrelRCNN.CLASS_WEIGHT 1.333 MODEL.SimrelRCNN.L1_WEIGHT 5.0 MODEL.SimrelRCNN.GIOU_WEIGHT 2.0 GLOVE_DIR '' MODEL.SimrelRCNN.ENABLE_FREQ False MODEL.SimrelRCNN.ENABLE_QUERY_REVERSE False MODEL.SimrelRCNN.USE_REFINE_OBJ_FEATURE True MODEL.SimrelRCNN.FREEZE_PUREE_OBJDET False MODEL.SimrelRCNN.PURE_ENT_NUM_PROPOSALS 100 MODEL.SimrelRCNN.ENABLE_MASK_BRANCH False MODEL.SimrelRCNN.KL_BRANCH_WEIGHT 0.25 MODEL.SimrelRCNN.ENABLE_KL_BRANCH True MODEL.SimrelRCNN.POSI_ENCODE_DIM 64 MODEL.SimrelRCNN.REL_CLASS_WEIGHT 1.334 MODEL.SimrelRCNN.PURE_ENT_CLASS_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_GIOU_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_L1_WEIGHT 5.0 MODEL.SimrelRCNN.AUXILIARY_BRANCH True MODEL.SimrelRCNN.AUXILIARY_BRANCH_SELECT_ENT_MAX_NUM 25 MODEL.SimrelRCNN.AUXILIARY_BRANCH_START 11 MODEL.SimrelRCNN.ENABLE_ENT_PROP True MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_ANALYSIS False MODEL.SimrelRCNN.USE_CROSS_RANK False MODEL.SimrelRCNN.DISABLE_KQ_FUSION_SELFATTEN False MODEL.SimrelRCNN.DIM_ENT_PRE_CLS 1024 MODEL.SimrelRCNN.DIM_ENT_PRE_REG 256 MODEL.SimrelRCNN.ONE_REL_CONV True MODEL.SimrelRCNN.ENABLE_BATCH_REDUCTION True MODEL.SimrelRCNN.USE_HARD_LABEL_KLMATCH True MODEL.SimrelRCNN.DISABLE_OBJ2REL_LOSS True

Evaluation

VG

300 queries

To evaluate my model, run:

CUDA_VISIBLE_DEVICES=4,5,6,7 python -m torch.distributed.launch --master_port 10029 --nproc_per_node=4 tools/relation_test_net.py --config-file "configs/simrel_e2e_relation_X_101_32_8_FPN_1x.yaml" MODEL.ROI_RELATION_HEAD.USE_GT_BOX False MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False MODEL.ROI_RELATION_HEAD.PREDICTOR CausalAnalysisPredictor MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_TYPE none MODEL.ROI_RELATION_HEAD.CAUSAL.FUSION_TYPE sum MODEL.ROI_RELATION_HEAD.CAUSAL.CONTEXT_LAYER motifs SOLVER.IMS_PER_BATCH 8 TEST.IMS_PER_BATCH 4 SOLVER.VAL_PERIOD 12000 SOLVER.MAX_ITER 80000 SOLVER.STEPS '(47000, 64000)' SOLVER.BASE_LR 0.000008 SOLVER.OPTIMIZER "ADAMW" MODEL.WEIGHT Outputs/real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel/model_final.pth OUTPUT_DIR Outputs/eva_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel MODEL.ROI_BOX_HEAD.MLP_HEAD_DIM 1024 MODEL.SimrelRCNN.ENABLE_BG_OBJ False MODEL.SimrelRCNN.ENABLE_REL_X2Y True MODEL.SimrelRCNN.REL_DIM 256 MODEL.SimrelRCNN.NUM_PROPOSALS 300 MODEL.SimrelRCNN.NUM_HEADS 6 MODEL.SimrelRCNN.REL_STACK_NUM 6 MODEL.SimrelRCNN.TRIPLET_MASK_WEIGHT 1.0 MODEL.SimrelRCNN.FREEZE_BACKBONE True MODEL.SimrelRCNN.CROSS_OBJ_FEAT_FUSION False MODEL.SimrelRCNN.CLASS_WEIGHT 1.333 MODEL.SimrelRCNN.L1_WEIGHT 5.0 MODEL.SimrelRCNN.GIOU_WEIGHT 2.0 GLOVE_DIR '' MODEL.SimrelRCNN.ENABLE_FREQ False MODEL.SimrelRCNN.ENABLE_QUERY_REVERSE False MODEL.SimrelRCNN.USE_REFINE_OBJ_FEATURE True MODEL.SimrelRCNN.FREEZE_PUREE_OBJDET False MODEL.SimrelRCNN.PURE_ENT_NUM_PROPOSALS 100 MODEL.SimrelRCNN.ENABLE_MASK_BRANCH False MODEL.SimrelRCNN.KL_BRANCH_WEIGHT 0.25 MODEL.SimrelRCNN.ENABLE_KL_BRANCH True MODEL.SimrelRCNN.POSI_ENCODE_DIM 64 MODEL.SimrelRCNN.REL_CLASS_WEIGHT 1.334 MODEL.SimrelRCNN.PURE_ENT_CLASS_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_GIOU_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_L1_WEIGHT 5.0 MODEL.SimrelRCNN.AUXILIARY_BRANCH True MODEL.SimrelRCNN.AUXILIARY_BRANCH_SELECT_ENT_MAX_NUM 25 MODEL.SimrelRCNN.AUXILIARY_BRANCH_START 11 MODEL.SimrelRCNN.ENABLE_ENT_PROP True MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_ANALYSIS False MODEL.SimrelRCNN.USE_CROSS_RANK False MODEL.SimrelRCNN.DISABLE_KQ_FUSION_SELFATTEN False MODEL.SimrelRCNN.DIM_ENT_PRE_CLS 256 MODEL.SimrelRCNN.DIM_ENT_PRE_REG 256 MODEL.SimrelRCNN.ONE_REL_CONV True MODEL.SimrelRCNN.ENABLE_BATCH_REDUCTION True MODEL.SimrelRCNN.USE_HARD_LABEL_KLMATCH True MODEL.SimrelRCNN.DISABLE_OBJ2REL_LOSS True

📋 Before testing, create a new directory, named 'eva_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel', in ./SGGbench/Outputs/ .

To evaluate my model with TDE, run:

CUDA_VISIBLE_DEVICES=2,3,4,5 python -m torch.distributed.launch --master_port 10029 --nproc_per_node=4 tools/relation_test_net.py --config-file "configs/simrel_e2e_relation_X_101_32_8_FPN_1x.yaml" MODEL.ROI_RELATION_HEAD.USE_GT_BOX False MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False MODEL.ROI_RELATION_HEAD.PREDICTOR CausalAnalysisPredictor MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_TYPE none MODEL.ROI_RELATION_HEAD.CAUSAL.FUSION_TYPE sum MODEL.ROI_RELATION_HEAD.CAUSAL.CONTEXT_LAYER motifs SOLVER.IMS_PER_BATCH 8 TEST.IMS_PER_BATCH 4 SOLVER.VAL_PERIOD 12000 SOLVER.MAX_ITER 80000 SOLVER.STEPS '(47000, 64000)' SOLVER.BASE_LR 0.000008 SOLVER.OPTIMIZER "ADAMW" MODEL.WEIGHT Outputs/real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel/model_final.pth OUTPUT_DIR Outputs/tde_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel MODEL.ROI_BOX_HEAD.MLP_HEAD_DIM 1024 MODEL.SimrelRCNN.ENABLE_BG_OBJ False MODEL.SimrelRCNN.ENABLE_REL_X2Y True MODEL.SimrelRCNN.REL_DIM 256 MODEL.SimrelRCNN.NUM_PROPOSALS 300 MODEL.SimrelRCNN.NUM_HEADS 6 MODEL.SimrelRCNN.REL_STACK_NUM 6 MODEL.SimrelRCNN.TRIPLET_MASK_WEIGHT 1.0 MODEL.SimrelRCNN.FREEZE_BACKBONE True MODEL.SimrelRCNN.CROSS_OBJ_FEAT_FUSION False MODEL.SimrelRCNN.CLASS_WEIGHT 1.333 MODEL.SimrelRCNN.L1_WEIGHT 5.0 MODEL.SimrelRCNN.GIOU_WEIGHT 2.0 GLOVE_DIR '' MODEL.SimrelRCNN.ENABLE_FREQ False MODEL.SimrelRCNN.ENABLE_QUERY_REVERSE False MODEL.SimrelRCNN.USE_REFINE_OBJ_FEATURE True MODEL.SimrelRCNN.FREEZE_PUREE_OBJDET False MODEL.SimrelRCNN.PURE_ENT_NUM_PROPOSALS 100 MODEL.SimrelRCNN.ENABLE_MASK_BRANCH False MODEL.SimrelRCNN.KL_BRANCH_WEIGHT 0.25 MODEL.SimrelRCNN.ENABLE_KL_BRANCH True MODEL.SimrelRCNN.POSI_ENCODE_DIM 64 MODEL.SimrelRCNN.REL_CLASS_WEIGHT 1.334 MODEL.SimrelRCNN.PURE_ENT_CLASS_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_GIOU_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_L1_WEIGHT 5.0 MODEL.SimrelRCNN.AUXILIARY_BRANCH True MODEL.SimrelRCNN.AUXILIARY_BRANCH_SELECT_ENT_MAX_NUM 25 MODEL.SimrelRCNN.AUXILIARY_BRANCH_START 11 MODEL.SimrelRCNN.ENABLE_ENT_PROP True MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_ANALYSIS True MODEL.SimrelRCNN.USE_CROSS_RANK False MODEL.SimrelRCNN.DISABLE_KQ_FUSION_SELFATTEN False MODEL.SimrelRCNN.DIM_ENT_PRE_CLS 256 MODEL.SimrelRCNN.DIM_ENT_PRE_REG 256 MODEL.SimrelRCNN.ONE_REL_CONV True MODEL.SimrelRCNN.ENABLE_BATCH_REDUCTION True MODEL.SimrelRCNN.USE_HARD_LABEL_KLMATCH True MODEL.SimrelRCNN.DISABLE_OBJ2REL_LOSS True

📋 Before testing, create a new directory, named 'tde_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel', in ./SGGbench/Outputs/ .

To evaluate my model with our LA, run:

CUDA_VISIBLE_DEVICES=4,5,6,7 python -m torch.distributed.launch --master_port 10029 --nproc_per_node=4 tools/relation_test_net.py --config-file "configs/simrel_e2e_relation_X_101_32_8_FPN_1x.yaml" MODEL.ROI_RELATION_HEAD.USE_GT_BOX False MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False MODEL.ROI_RELATION_HEAD.PREDICTOR CausalAnalysisPredictor MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_TYPE none MODEL.ROI_RELATION_HEAD.CAUSAL.FUSION_TYPE sum MODEL.ROI_RELATION_HEAD.CAUSAL.CONTEXT_LAYER motifs SOLVER.IMS_PER_BATCH 8 TEST.IMS_PER_BATCH 4 SOLVER.VAL_PERIOD 12000 SOLVER.MAX_ITER 80000 SOLVER.STEPS '(47000, 64000)' SOLVER.BASE_LR 0.000008 SOLVER.OPTIMIZER "ADAMW" MODEL.WEIGHT Outputs/real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel/model_final.pth OUTPUT_DIR Outputs/la03_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel MODEL.ROI_BOX_HEAD.MLP_HEAD_DIM 1024 MODEL.SimrelRCNN.ENABLE_BG_OBJ False MODEL.SimrelRCNN.ENABLE_REL_X2Y True MODEL.SimrelRCNN.REL_DIM 256 MODEL.SimrelRCNN.NUM_PROPOSALS 300 MODEL.SimrelRCNN.NUM_HEADS 6 MODEL.SimrelRCNN.REL_STACK_NUM 6 MODEL.SimrelRCNN.TRIPLET_MASK_WEIGHT 1.0 MODEL.SimrelRCNN.FREEZE_BACKBONE True MODEL.SimrelRCNN.CROSS_OBJ_FEAT_FUSION False MODEL.SimrelRCNN.CLASS_WEIGHT 1.333 MODEL.SimrelRCNN.L1_WEIGHT 5.0 MODEL.SimrelRCNN.GIOU_WEIGHT 2.0 GLOVE_DIR '' MODEL.SimrelRCNN.ENABLE_FREQ False MODEL.SimrelRCNN.ENABLE_QUERY_REVERSE False MODEL.SimrelRCNN.USE_REFINE_OBJ_FEATURE True MODEL.SimrelRCNN.FREEZE_PUREE_OBJDET False MODEL.SimrelRCNN.PURE_ENT_NUM_PROPOSALS 100 MODEL.SimrelRCNN.ENABLE_MASK_BRANCH False MODEL.SimrelRCNN.KL_BRANCH_WEIGHT 0.25 MODEL.SimrelRCNN.ENABLE_KL_BRANCH True MODEL.SimrelRCNN.POSI_ENCODE_DIM 64 MODEL.SimrelRCNN.REL_CLASS_WEIGHT 1.334 MODEL.SimrelRCNN.PURE_ENT_CLASS_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_GIOU_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_L1_WEIGHT 5.0 MODEL.SimrelRCNN.AUXILIARY_BRANCH True MODEL.SimrelRCNN.AUXILIARY_BRANCH_SELECT_ENT_MAX_NUM 25 MODEL.SimrelRCNN.AUXILIARY_BRANCH_START 11 MODEL.SimrelRCNN.ENABLE_ENT_PROP True MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_ANALYSIS False MODEL.SimrelRCNN.USE_CROSS_RANK False MODEL.SimrelRCNN.DISABLE_KQ_FUSION_SELFATTEN False MODEL.SimrelRCNN.DIM_ENT_PRE_CLS 256 MODEL.SimrelRCNN.DIM_ENT_PRE_REG 256 MODEL.SimrelRCNN.ONE_REL_CONV True MODEL.SimrelRCNN.ENABLE_BATCH_REDUCTION True MODEL.SimrelRCNN.USE_HARD_LABEL_KLMATCH True MODEL.SimrelRCNN.DISABLE_OBJ2REL_LOSS True MODEL.SimrelRCNN.REL_LOGITS_ADJUSTMENT True MODEL.SimrelRCNN.LOGIT_ADJ_TAU 0.3

📋 Before testing, create a new directory, named 'la03_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel', in ./SGGbench/Outputs/ .

800 queries

To evaluate my model, run:

CUDA_VISIBLE_DEVICES=3,5,6,7 python -m torch.distributed.launch --master_port 10029 --nproc_per_node=4 tools/relation_test_net.py --config-file "configs/simrel_e2e_relation_X_101_32_8_FPN_1x.yaml" MODEL.ROI_RELATION_HEAD.USE_GT_BOX False MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False MODEL.ROI_RELATION_HEAD.PREDICTOR CausalAnalysisPredictor MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_TYPE none MODEL.ROI_RELATION_HEAD.CAUSAL.FUSION_TYPE sum MODEL.ROI_RELATION_HEAD.CAUSAL.CONTEXT_LAYER motifs SOLVER.IMS_PER_BATCH 8 TEST.IMS_PER_BATCH 4 SOLVER.VAL_PERIOD 12000 SOLVER.MAX_ITER 80000 SOLVER.STEPS '(47000, 64000)' SOLVER.BASE_LR 0.000008 SOLVER.OPTIMIZER "ADAMW" MODEL.WEIGHT Outputs/real025kl_newe2rposition_fullobjbranch_800q_smalldyconv_prequeryaug9_norel_hardlabel/model_final.pth OUTPUT_DIR Outputs/eva_real025kl_newe2rposition_fullobjbranch_800q_smalldyconv_prequeryaug9_norel_hardlabel MODEL.ROI_BOX_HEAD.MLP_HEAD_DIM 1024 MODEL.SimrelRCNN.ENABLE_BG_OBJ False MODEL.SimrelRCNN.ENABLE_REL_X2Y True MODEL.SimrelRCNN.REL_DIM 256 MODEL.SimrelRCNN.NUM_PROPOSALS 800 MODEL.SimrelRCNN.NUM_HEADS 6 MODEL.SimrelRCNN.REL_STACK_NUM 6 MODEL.SimrelRCNN.TRIPLET_MASK_WEIGHT 1.0 MODEL.SimrelRCNN.FREEZE_BACKBONE True MODEL.SimrelRCNN.CROSS_OBJ_FEAT_FUSION False MODEL.SimrelRCNN.CLASS_WEIGHT 1.333 MODEL.SimrelRCNN.L1_WEIGHT 5.0 MODEL.SimrelRCNN.GIOU_WEIGHT 2.0 GLOVE_DIR '' MODEL.SimrelRCNN.ENABLE_FREQ False MODEL.SimrelRCNN.ENABLE_QUERY_REVERSE False MODEL.SimrelRCNN.USE_REFINE_OBJ_FEATURE True MODEL.SimrelRCNN.FREEZE_PUREE_OBJDET False MODEL.SimrelRCNN.PURE_ENT_NUM_PROPOSALS 100 MODEL.SimrelRCNN.ENABLE_MASK_BRANCH False MODEL.SimrelRCNN.KL_BRANCH_WEIGHT 0.25 MODEL.SimrelRCNN.ENABLE_KL_BRANCH True MODEL.SimrelRCNN.POSI_ENCODE_DIM 64 MODEL.SimrelRCNN.REL_CLASS_WEIGHT 1.334 MODEL.SimrelRCNN.PURE_ENT_CLASS_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_GIOU_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_L1_WEIGHT 5.0 MODEL.SimrelRCNN.AUXILIARY_BRANCH True MODEL.SimrelRCNN.AUXILIARY_BRANCH_SELECT_ENT_MAX_NUM 25 MODEL.SimrelRCNN.AUXILIARY_BRANCH_START 11 MODEL.SimrelRCNN.ENABLE_ENT_PROP True MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_ANALYSIS False MODEL.SimrelRCNN.USE_CROSS_RANK False MODEL.SimrelRCNN.DISABLE_KQ_FUSION_SELFATTEN False MODEL.SimrelRCNN.DIM_ENT_PRE_CLS 256 MODEL.SimrelRCNN.DIM_ENT_PRE_REG 256 MODEL.SimrelRCNN.ONE_REL_CONV True MODEL.SimrelRCNN.ENABLE_BATCH_REDUCTION True MODEL.SimrelRCNN.USE_HARD_LABEL_KLMATCH True MODEL.SimrelRCNN.DISABLE_OBJ2REL_LOSS True

📋 Before testing, create a new directory, named 'eva_real025kl_newe2rposition_fullobjbranch_800q_smalldyconv_prequeryaug9_norel_hardlabel', in ./SGGbench/Outputs/ .

To evaluate my model with TDE, run:

CUDA_VISIBLE_DEVICES=2,3,4,5 python -m torch.distributed.launch --master_port 10029 --nproc_per_node=4 tools/relation_test_net.py --config-file "configs/simrel_e2e_relation_X_101_32_8_FPN_1x.yaml" MODEL.ROI_RELATION_HEAD.USE_GT_BOX False MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False MODEL.ROI_RELATION_HEAD.PREDICTOR CausalAnalysisPredictor MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_TYPE none MODEL.ROI_RELATION_HEAD.CAUSAL.FUSION_TYPE sum MODEL.ROI_RELATION_HEAD.CAUSAL.CONTEXT_LAYER motifs SOLVER.IMS_PER_BATCH 8 TEST.IMS_PER_BATCH 4 SOLVER.VAL_PERIOD 12000 SOLVER.MAX_ITER 80000 SOLVER.STEPS '(47000, 64000)' SOLVER.BASE_LR 0.000008 SOLVER.OPTIMIZER "ADAMW" MODEL.WEIGHT Outputs/real025kl_newe2rposition_fullobjbranch_800q_smalldyconv_prequeryaug9_norel_hardlabel/model_final.pth OUTPUT_DIR Outputs/tde_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel MODEL.ROI_BOX_HEAD.MLP_HEAD_DIM 1024 MODEL.SimrelRCNN.ENABLE_BG_OBJ False MODEL.SimrelRCNN.ENABLE_REL_X2Y True MODEL.SimrelRCNN.REL_DIM 256 MODEL.SimrelRCNN.NUM_PROPOSALS 800 MODEL.SimrelRCNN.NUM_HEADS 6 MODEL.SimrelRCNN.REL_STACK_NUM 6 MODEL.SimrelRCNN.TRIPLET_MASK_WEIGHT 1.0 MODEL.SimrelRCNN.FREEZE_BACKBONE True MODEL.SimrelRCNN.CROSS_OBJ_FEAT_FUSION False MODEL.SimrelRCNN.CLASS_WEIGHT 1.333 MODEL.SimrelRCNN.L1_WEIGHT 5.0 MODEL.SimrelRCNN.GIOU_WEIGHT 2.0 GLOVE_DIR '' MODEL.SimrelRCNN.ENABLE_FREQ False MODEL.SimrelRCNN.ENABLE_QUERY_REVERSE False MODEL.SimrelRCNN.USE_REFINE_OBJ_FEATURE True MODEL.SimrelRCNN.FREEZE_PUREE_OBJDET False MODEL.SimrelRCNN.PURE_ENT_NUM_PROPOSALS 100 MODEL.SimrelRCNN.ENABLE_MASK_BRANCH False MODEL.SimrelRCNN.KL_BRANCH_WEIGHT 0.25 MODEL.SimrelRCNN.ENABLE_KL_BRANCH True MODEL.SimrelRCNN.POSI_ENCODE_DIM 64 MODEL.SimrelRCNN.REL_CLASS_WEIGHT 1.334 MODEL.SimrelRCNN.PURE_ENT_CLASS_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_GIOU_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_L1_WEIGHT 5.0 MODEL.SimrelRCNN.AUXILIARY_BRANCH True MODEL.SimrelRCNN.AUXILIARY_BRANCH_SELECT_ENT_MAX_NUM 25 MODEL.SimrelRCNN.AUXILIARY_BRANCH_START 11 MODEL.SimrelRCNN.ENABLE_ENT_PROP True MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_ANALYSIS True MODEL.SimrelRCNN.USE_CROSS_RANK False MODEL.SimrelRCNN.DISABLE_KQ_FUSION_SELFATTEN False MODEL.SimrelRCNN.DIM_ENT_PRE_CLS 256 MODEL.SimrelRCNN.DIM_ENT_PRE_REG 256 MODEL.SimrelRCNN.ONE_REL_CONV True MODEL.SimrelRCNN.ENABLE_BATCH_REDUCTION True MODEL.SimrelRCNN.USE_HARD_LABEL_KLMATCH True MODEL.SimrelRCNN.DISABLE_OBJ2REL_LOSS True

📋 Before testing, create a new directory, named 'tde_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel', in ./SGGbench/Outputs/ .

To evaluate my model with our LA, run:

CUDA_VISIBLE_DEVICES=3,5,6,7 python -m torch.distributed.launch --master_port 10029 --nproc_per_node=4 tools/relation_test_net.py --config-file "configs/simrel_e2e_relation_X_101_32_8_FPN_1x.yaml" MODEL.ROI_RELATION_HEAD.USE_GT_BOX False MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False MODEL.ROI_RELATION_HEAD.PREDICTOR CausalAnalysisPredictor MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_TYPE none MODEL.ROI_RELATION_HEAD.CAUSAL.FUSION_TYPE sum MODEL.ROI_RELATION_HEAD.CAUSAL.CONTEXT_LAYER motifs SOLVER.IMS_PER_BATCH 8 TEST.IMS_PER_BATCH 4 SOLVER.VAL_PERIOD 12000 SOLVER.MAX_ITER 80000 SOLVER.STEPS '(47000, 64000)' SOLVER.BASE_LR 0.000008 SOLVER.OPTIMIZER "ADAMW" MODEL.WEIGHT Outputs/real025kl_newe2rposition_fullobjbranch_800q_smalldyconv_prequeryaug9_norel_hardlabel/model_final.pth OUTPUT_DIR Outputs/LA03_real025kl_newe2rposition_fullobjbranch_800q_smalldyconv_prequeryaug9_norel_hardlabel MODEL.ROI_BOX_HEAD.MLP_HEAD_DIM 1024 MODEL.SimrelRCNN.ENABLE_BG_OBJ False MODEL.SimrelRCNN.ENABLE_REL_X2Y True MODEL.SimrelRCNN.REL_DIM 256 MODEL.SimrelRCNN.NUM_PROPOSALS 800 MODEL.SimrelRCNN.NUM_HEADS 6 MODEL.SimrelRCNN.REL_STACK_NUM 6 MODEL.SimrelRCNN.TRIPLET_MASK_WEIGHT 1.0 MODEL.SimrelRCNN.FREEZE_BACKBONE True MODEL.SimrelRCNN.CROSS_OBJ_FEAT_FUSION False MODEL.SimrelRCNN.CLASS_WEIGHT 1.333 MODEL.SimrelRCNN.L1_WEIGHT 5.0 MODEL.SimrelRCNN.GIOU_WEIGHT 2.0 GLOVE_DIR '' MODEL.SimrelRCNN.ENABLE_FREQ False MODEL.SimrelRCNN.ENABLE_QUERY_REVERSE False MODEL.SimrelRCNN.USE_REFINE_OBJ_FEATURE True MODEL.SimrelRCNN.FREEZE_PUREE_OBJDET False MODEL.SimrelRCNN.PURE_ENT_NUM_PROPOSALS 100 MODEL.SimrelRCNN.ENABLE_MASK_BRANCH False MODEL.SimrelRCNN.KL_BRANCH_WEIGHT 0.25 MODEL.SimrelRCNN.ENABLE_KL_BRANCH True MODEL.SimrelRCNN.POSI_ENCODE_DIM 64 MODEL.SimrelRCNN.REL_CLASS_WEIGHT 1.334 MODEL.SimrelRCNN.PURE_ENT_CLASS_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_GIOU_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_L1_WEIGHT 5.0 MODEL.SimrelRCNN.AUXILIARY_BRANCH True MODEL.SimrelRCNN.AUXILIARY_BRANCH_SELECT_ENT_MAX_NUM 25 MODEL.SimrelRCNN.AUXILIARY_BRANCH_START 11 MODEL.SimrelRCNN.ENABLE_ENT_PROP True MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_ANALYSIS False MODEL.SimrelRCNN.USE_CROSS_RANK False MODEL.SimrelRCNN.DISABLE_KQ_FUSION_SELFATTEN False MODEL.SimrelRCNN.DIM_ENT_PRE_CLS 256 MODEL.SimrelRCNN.DIM_ENT_PRE_REG 256 MODEL.SimrelRCNN.ONE_REL_CONV True MODEL.SimrelRCNN.ENABLE_BATCH_REDUCTION True MODEL.SimrelRCNN.USE_HARD_LABEL_KLMATCH True MODEL.SimrelRCNN.DISABLE_OBJ2REL_LOSS True MODEL.SimrelRCNN.REL_LOGITS_ADJUSTMENT True MODEL.SimrelRCNN.LOGIT_ADJ_TAU 0.3

📋 Before testing, create a new directory, named 'LA03_real025kl_newe2rposition_fullobjbranch_800q_smalldyconv_prequeryaug9_norel_hardlabel', in ./SGGbench/Outputs/ .

OI V4

300 queries

To evaluate my model, run:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m torch.distributed.launch --master_port 10020 --nproc_per_node=8 tools/relation_test_net.py --config-file "configs/oiv4_simrel_e2e_relation_X_101_32_8_FPN_1x.yaml" MODEL.ROI_RELATION_HEAD.USE_GT_BOX False MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False MODEL.ROI_RELATION_HEAD.PREDICTOR CausalAnalysisPredictor MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_TYPE none MODEL.ROI_RELATION_HEAD.CAUSAL.FUSION_TYPE sum MODEL.ROI_RELATION_HEAD.CAUSAL.CONTEXT_LAYER motifs SOLVER.IMS_PER_BATCH 8 TEST.IMS_PER_BATCH 8 SOLVER.VAL_PERIOD 12000 SOLVER.MAX_ITER 80000 SOLVER.STEPS '(47000, 64000)' SOLVER.BASE_LR 0.000008 SOLVER.OPTIMIZER "ADAMW" MODEL.WEIGHT Outputs/oiv4_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel/model_final.pth OUTPUT_DIR Outputs/eva_oiv4_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel MODEL.ROI_BOX_HEAD.MLP_HEAD_DIM 1024 MODEL.SimrelRCNN.ENABLE_BG_OBJ False MODEL.SimrelRCNN.ENABLE_REL_X2Y True MODEL.SimrelRCNN.REL_DIM 256 MODEL.SimrelRCNN.NUM_PROPOSALS 300 MODEL.SimrelRCNN.NUM_HEADS 6 MODEL.SimrelRCNN.REL_STACK_NUM 6 MODEL.SimrelRCNN.TRIPLET_MASK_WEIGHT 1.0 MODEL.SimrelRCNN.FREEZE_BACKBONE True MODEL.SimrelRCNN.CROSS_OBJ_FEAT_FUSION False MODEL.SimrelRCNN.CLASS_WEIGHT 1.333 MODEL.SimrelRCNN.L1_WEIGHT 5.0 MODEL.SimrelRCNN.GIOU_WEIGHT 2.0 GLOVE_DIR '' MODEL.SimrelRCNN.ENABLE_FREQ False MODEL.SimrelRCNN.ENABLE_QUERY_REVERSE False MODEL.SimrelRCNN.USE_REFINE_OBJ_FEATURE True MODEL.SimrelRCNN.FREEZE_PUREE_OBJDET False MODEL.SimrelRCNN.PURE_ENT_NUM_PROPOSALS 100 MODEL.SimrelRCNN.ENABLE_MASK_BRANCH False MODEL.SimrelRCNN.KL_BRANCH_WEIGHT 0.25 MODEL.SimrelRCNN.ENABLE_KL_BRANCH True MODEL.SimrelRCNN.POSI_ENCODE_DIM 64 MODEL.SimrelRCNN.REL_CLASS_WEIGHT 1.334 MODEL.SimrelRCNN.PURE_ENT_CLASS_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_GIOU_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_L1_WEIGHT 5.0 MODEL.SimrelRCNN.AUXILIARY_BRANCH True MODEL.SimrelRCNN.AUXILIARY_BRANCH_SELECT_ENT_MAX_NUM 25 MODEL.SimrelRCNN.AUXILIARY_BRANCH_START 11 MODEL.SimrelRCNN.ENABLE_ENT_PROP True MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_ANALYSIS False MODEL.SimrelRCNN.USE_CROSS_RANK False MODEL.SimrelRCNN.DISABLE_KQ_FUSION_SELFATTEN False MODEL.SimrelRCNN.DIM_ENT_PRE_CLS 1024 MODEL.SimrelRCNN.DIM_ENT_PRE_REG 256 MODEL.SimrelRCNN.ONE_REL_CONV True MODEL.SimrelRCNN.ENABLE_BATCH_REDUCTION True MODEL.SimrelRCNN.USE_HARD_LABEL_KLMATCH True MODEL.SimrelRCNN.DISABLE_OBJ2REL_LOSS True

📋 Before testing, create a new directory, named 'eva_oiv4_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel', in ./SGGbench/Outputs/ .

To evaluate my model with our LA, run:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m torch.distributed.launch --master_port 10020 --nproc_per_node=8 tools/relation_test_net.py --config-file "configs/oiv4_simrel_e2e_relation_X_101_32_8_FPN_1x.yaml" MODEL.ROI_RELATION_HEAD.USE_GT_BOX False MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False MODEL.ROI_RELATION_HEAD.PREDICTOR CausalAnalysisPredictor MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_TYPE none MODEL.ROI_RELATION_HEAD.CAUSAL.FUSION_TYPE sum MODEL.ROI_RELATION_HEAD.CAUSAL.CONTEXT_LAYER motifs SOLVER.IMS_PER_BATCH 8 TEST.IMS_PER_BATCH 8 SOLVER.VAL_PERIOD 12000 SOLVER.MAX_ITER 80000 SOLVER.STEPS '(47000, 64000)' SOLVER.BASE_LR 0.000008 SOLVER.OPTIMIZER "ADAMW" MODEL.WEIGHT Outputs/oiv4_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel/model_final.pth OUTPUT_DIR Outputs/la_oiv4_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel MODEL.ROI_BOX_HEAD.MLP_HEAD_DIM 1024 MODEL.SimrelRCNN.ENABLE_BG_OBJ False MODEL.SimrelRCNN.ENABLE_REL_X2Y True MODEL.SimrelRCNN.REL_DIM 256 MODEL.SimrelRCNN.NUM_PROPOSALS 300 MODEL.SimrelRCNN.NUM_HEADS 6 MODEL.SimrelRCNN.REL_STACK_NUM 6 MODEL.SimrelRCNN.TRIPLET_MASK_WEIGHT 1.0 MODEL.SimrelRCNN.FREEZE_BACKBONE True MODEL.SimrelRCNN.CROSS_OBJ_FEAT_FUSION False MODEL.SimrelRCNN.CLASS_WEIGHT 1.333 MODEL.SimrelRCNN.L1_WEIGHT 5.0 MODEL.SimrelRCNN.GIOU_WEIGHT 2.0 GLOVE_DIR '' MODEL.SimrelRCNN.ENABLE_FREQ False MODEL.SimrelRCNN.ENABLE_QUERY_REVERSE False MODEL.SimrelRCNN.USE_REFINE_OBJ_FEATURE True MODEL.SimrelRCNN.FREEZE_PUREE_OBJDET False MODEL.SimrelRCNN.PURE_ENT_NUM_PROPOSALS 100 MODEL.SimrelRCNN.ENABLE_MASK_BRANCH False MODEL.SimrelRCNN.KL_BRANCH_WEIGHT 0.25 MODEL.SimrelRCNN.ENABLE_KL_BRANCH True MODEL.SimrelRCNN.POSI_ENCODE_DIM 64 MODEL.SimrelRCNN.REL_CLASS_WEIGHT 1.334 MODEL.SimrelRCNN.PURE_ENT_CLASS_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_GIOU_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_L1_WEIGHT 5.0 MODEL.SimrelRCNN.AUXILIARY_BRANCH True MODEL.SimrelRCNN.AUXILIARY_BRANCH_SELECT_ENT_MAX_NUM 25 MODEL.SimrelRCNN.AUXILIARY_BRANCH_START 11 MODEL.SimrelRCNN.ENABLE_ENT_PROP True MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_ANALYSIS False MODEL.SimrelRCNN.USE_CROSS_RANK False MODEL.SimrelRCNN.DISABLE_KQ_FUSION_SELFATTEN False MODEL.SimrelRCNN.DIM_ENT_PRE_CLS 1024 MODEL.SimrelRCNN.DIM_ENT_PRE_REG 256 MODEL.SimrelRCNN.ONE_REL_CONV True MODEL.SimrelRCNN.ENABLE_BATCH_REDUCTION True MODEL.SimrelRCNN.USE_HARD_LABEL_KLMATCH True MODEL.SimrelRCNN.DISABLE_OBJ2REL_LOSS True MODEL.SimrelRCNN.REL_LOGITS_ADJUSTMENT True MODEL.SimrelRCNN.LOGIT_ADJ_TAU 0.3

📋 Before testing, create a new directory, named 'la_oiv4_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel', in ./SGGbench/Outputs/ .

OI V6

300 queries

To evaluate my model, run:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m torch.distributed.launch --master_port 10020 --nproc_per_node=8 tools/relation_test_net.py --config-file "configs/oiv6_simrel_e2e_relation_X_101_32_8_FPN_1x.yaml" MODEL.ROI_RELATION_HEAD.USE_GT_BOX False MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False MODEL.ROI_RELATION_HEAD.PREDICTOR CausalAnalysisPredictor MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_TYPE none MODEL.ROI_RELATION_HEAD.CAUSAL.FUSION_TYPE sum MODEL.ROI_RELATION_HEAD.CAUSAL.CONTEXT_LAYER motifs SOLVER.IMS_PER_BATCH 8 TEST.IMS_PER_BATCH 8 SOLVER.VAL_PERIOD 12000 SOLVER.MAX_ITER 80000 SOLVER.STEPS '(47000, 64000)' SOLVER.BASE_LR 0.000008 SOLVER.OPTIMIZER "ADAMW" MODEL.WEIGHT Outputs/oiv6_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel/model_final.pth OUTPUT_DIR Outputs/eva_oiv6_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel MODEL.ROI_BOX_HEAD.MLP_HEAD_DIM 1024 MODEL.SimrelRCNN.ENABLE_BG_OBJ False MODEL.SimrelRCNN.ENABLE_REL_X2Y True MODEL.SimrelRCNN.REL_DIM 256 MODEL.SimrelRCNN.NUM_PROPOSALS 300 MODEL.SimrelRCNN.NUM_HEADS 6 MODEL.SimrelRCNN.REL_STACK_NUM 6 MODEL.SimrelRCNN.TRIPLET_MASK_WEIGHT 1.0 MODEL.SimrelRCNN.FREEZE_BACKBONE True MODEL.SimrelRCNN.CROSS_OBJ_FEAT_FUSION False MODEL.SimrelRCNN.CLASS_WEIGHT 1.333 MODEL.SimrelRCNN.L1_WEIGHT 5.0 MODEL.SimrelRCNN.GIOU_WEIGHT 2.0 GLOVE_DIR '' MODEL.SimrelRCNN.ENABLE_FREQ False MODEL.SimrelRCNN.ENABLE_QUERY_REVERSE False MODEL.SimrelRCNN.USE_REFINE_OBJ_FEATURE True MODEL.SimrelRCNN.FREEZE_PUREE_OBJDET False MODEL.SimrelRCNN.PURE_ENT_NUM_PROPOSALS 100 MODEL.SimrelRCNN.ENABLE_MASK_BRANCH False MODEL.SimrelRCNN.KL_BRANCH_WEIGHT 0.25 MODEL.SimrelRCNN.ENABLE_KL_BRANCH True MODEL.SimrelRCNN.POSI_ENCODE_DIM 64 MODEL.SimrelRCNN.REL_CLASS_WEIGHT 1.334 MODEL.SimrelRCNN.PURE_ENT_CLASS_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_GIOU_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_L1_WEIGHT 5.0 MODEL.SimrelRCNN.AUXILIARY_BRANCH True MODEL.SimrelRCNN.AUXILIARY_BRANCH_SELECT_ENT_MAX_NUM 25 MODEL.SimrelRCNN.AUXILIARY_BRANCH_START 11 MODEL.SimrelRCNN.ENABLE_ENT_PROP True MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_ANALYSIS False MODEL.SimrelRCNN.USE_CROSS_RANK False MODEL.SimrelRCNN.DISABLE_KQ_FUSION_SELFATTEN False MODEL.SimrelRCNN.DIM_ENT_PRE_CLS 1024 MODEL.SimrelRCNN.DIM_ENT_PRE_REG 256 MODEL.SimrelRCNN.ONE_REL_CONV True MODEL.SimrelRCNN.ENABLE_BATCH_REDUCTION True MODEL.SimrelRCNN.USE_HARD_LABEL_KLMATCH True MODEL.SimrelRCNN.DISABLE_OBJ2REL_LOSS True

📋 Before testing, create a new directory, named 'eva_oiv6_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel', in ./SGGbench/Outputs/ .

To evaluate my model with our LA, run:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m torch.distributed.launch --master_port 10020 --nproc_per_node=8 tools/relation_test_net.py --config-file "configs/oiv6_simrel_e2e_relation_X_101_32_8_FPN_1x.yaml" MODEL.ROI_RELATION_HEAD.USE_GT_BOX False MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False MODEL.ROI_RELATION_HEAD.PREDICTOR CausalAnalysisPredictor MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_TYPE none MODEL.ROI_RELATION_HEAD.CAUSAL.FUSION_TYPE sum MODEL.ROI_RELATION_HEAD.CAUSAL.CONTEXT_LAYER motifs SOLVER.IMS_PER_BATCH 8 TEST.IMS_PER_BATCH 8 SOLVER.VAL_PERIOD 12000 SOLVER.MAX_ITER 80000 SOLVER.STEPS '(47000, 64000)' SOLVER.BASE_LR 0.000008 SOLVER.OPTIMIZER "ADAMW" MODEL.WEIGHT Outputs/oiv6_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel/model_final.pth OUTPUT_DIR Outputs/la_oiv6_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel MODEL.ROI_BOX_HEAD.MLP_HEAD_DIM 1024 MODEL.SimrelRCNN.ENABLE_BG_OBJ False MODEL.SimrelRCNN.ENABLE_REL_X2Y True MODEL.SimrelRCNN.REL_DIM 256 MODEL.SimrelRCNN.NUM_PROPOSALS 300 MODEL.SimrelRCNN.NUM_HEADS 6 MODEL.SimrelRCNN.REL_STACK_NUM 6 MODEL.SimrelRCNN.TRIPLET_MASK_WEIGHT 1.0 MODEL.SimrelRCNN.FREEZE_BACKBONE True MODEL.SimrelRCNN.CROSS_OBJ_FEAT_FUSION False MODEL.SimrelRCNN.CLASS_WEIGHT 1.333 MODEL.SimrelRCNN.L1_WEIGHT 5.0 MODEL.SimrelRCNN.GIOU_WEIGHT 2.0 GLOVE_DIR '' MODEL.SimrelRCNN.ENABLE_FREQ False MODEL.SimrelRCNN.ENABLE_QUERY_REVERSE False MODEL.SimrelRCNN.USE_REFINE_OBJ_FEATURE True MODEL.SimrelRCNN.FREEZE_PUREE_OBJDET False MODEL.SimrelRCNN.PURE_ENT_NUM_PROPOSALS 100 MODEL.SimrelRCNN.ENABLE_MASK_BRANCH False MODEL.SimrelRCNN.KL_BRANCH_WEIGHT 0.25 MODEL.SimrelRCNN.ENABLE_KL_BRANCH True MODEL.SimrelRCNN.POSI_ENCODE_DIM 64 MODEL.SimrelRCNN.REL_CLASS_WEIGHT 1.334 MODEL.SimrelRCNN.PURE_ENT_CLASS_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_GIOU_WEIGHT 2.0 MODEL.SimrelRCNN.PURE_ENT_L1_WEIGHT 5.0 MODEL.SimrelRCNN.AUXILIARY_BRANCH True MODEL.SimrelRCNN.AUXILIARY_BRANCH_SELECT_ENT_MAX_NUM 25 MODEL.SimrelRCNN.AUXILIARY_BRANCH_START 11 MODEL.SimrelRCNN.ENABLE_ENT_PROP True MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_ANALYSIS False MODEL.SimrelRCNN.USE_CROSS_RANK False MODEL.SimrelRCNN.DISABLE_KQ_FUSION_SELFATTEN False MODEL.SimrelRCNN.DIM_ENT_PRE_CLS 1024 MODEL.SimrelRCNN.DIM_ENT_PRE_REG 256 MODEL.SimrelRCNN.ONE_REL_CONV True MODEL.SimrelRCNN.ENABLE_BATCH_REDUCTION True MODEL.SimrelRCNN.USE_HARD_LABEL_KLMATCH True MODEL.SimrelRCNN.DISABLE_OBJ2REL_LOSS True MODEL.SimrelRCNN.REL_LOGITS_ADJUSTMENT True MODEL.SimrelRCNN.LOGIT_ADJ_TAU 0.3

📋 Before testing, create a new directory, named 'la_oiv6_real025kl_newe2rposition_fullobjbranch_300q_smalldyconv_prequeryaug9_norel_hardlabel', in ./SGGbench/Outputs/ .

Pre-trained Models

The pre-trained backbone weights for our models are aligned with the ones for faster r-cnn. This can ensure a fair comparison of backbone to other SGG methods. These weights can be found in: BGNN-SGG (OI) and Scene-Graph-Benchmark.pytorch (VG).

You can download our pretrained models here:

VG

300 queries

800 queries

OI V4

300 queries

OI V6

300 queries

Results

Our model achieves the following performance on :

Visual Genome

* means 800 queries.

Models	SGGen R@20	SGGen R@50	SGGen R@100	SGGen zR@20	SGGen zR@50	SGGen zR@100	SGGen mR@20	SGGen mR@50	SGGen mR@100
Our model	25.8	32.7	36.9	1.5	2.7	3.7	6.1	8.4	10.0
Our model*	26.1	33.5	38.4	1.5	2.7	4.0	6.2	8.6	10.3
Our model+TDE	14.5	18.3	21.0	1.8	2.7	3.6	10.8	15.0	18.5
Our model*+TDE	15.0	19.7	22.9	1.6	2.7	3.8	9.8	14.6	18.0
Our model+LA	18.4	23.3	26.5	1.9	2.9	4.0	13.5	17.9	21.4
Our model*+LA	18.2	23.7	27.3	2.0	3.1	4.5	13.7	18.6	22.5

Hint

The mAP evaluator of object detection in our model: Because our model is directly predicting the object pairs, the AP evaluator in Cocoapi is actually not comparable it, so its mAP cannot be properly shown in log.txt. This mAP indicators of objects is not useful for our paradigm. If you want to obtain the exact mAP of individual object predictions from these predicted object pairs, you need to modify the codes of Cocoapi to select the individual objects from those pairs.

Acknowledgement

Our code is mainly based on: Scene-Graph-Benchmark.pytorch, SparseR-CNN and BGNN-SGG.

For this paper, I'm extremely grateful to my advisor Prof. Limin Wang. I should also appreciate my group members for discussing with me: Jing Tan, Ziteng Gao and Jiaqi Tang.

Citations

@InProceedings{Teng_2022_CVPR,
    author    = {Teng, Yao and Wang, Limin},
    title     = {Structured Sparse R-CNN for Direct Scene Graph Generation},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {19437-19446}
}