Wayfear / Coacher

Other
22 stars 2 forks source link

COACHER

This is all the code for our ECML-PKDD 2021 Paper: Zero-Shot Scene Graph Relation Prediction through Commonsense Knowledge Integration. We developed this repository based on KaiHua Tang's SGG Benchmark Project.

Installation

Please follow the instructions in INSTALL.md to set up the environment.

Data and Settings

We use the public dataset Visual Genome, with our zero-shot setting stored in last_0.5_biased_zero_shot.pt. After downloading this file, please place it to: Coacher/maskrcnn_benchmark/data/datasets/evaluation/vg/

The neighbor information is stored in file VG_neighbor.npy. After downloadining this file, please place it to: Coacher/maskrcnn_benchmark/datasets/vg

The path information is stored in file sub_graph_1.4.pth. After downloading it, please place it to: Coacher/maskrcnn_benchmark/datasets/vg

Regarding the ConceptNet Embedding, please download the newest version and set the parameters GLOVE_DIR as the path where you store the downloaded embedding file.

Usage

We include running commands for three proposed approached to integrate knowledge graph for scene graph generation on two datsets.

Zero-shot dataset

Neighbor

python -m torch.distributed.launch --master_port 10303 --nproc_per_node=1 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 True MODEL.ROI_RELATION_HEAD.PREDICTOR MotifPredictorWithExternalKnowledge SOLVER.IMS_PER_BATCH 12 TEST.IMS_PER_BATCH 12 DTYPE "float32" SOLVER.MAX_ITER 50000 SOLVER.VAL_PERIOD 2000 SOLVER.CHECKPOINT_PERIOD 2000 GLOVE_DIR /local/home/{Your Name}/glove MODEL.PRETRAINED_DETECTOR_CKPT /local/home/{Your Name}/checkpoints/pretrained_faster_rcnn/model_final.pth OUTPUT_DIR /local/home/{Your Name}/checkpoints/N_last_0.5_without_biased_data_G+centeremb_1_layer MODEL.ROI_RELATION_HEAD.PREDICT_USE_BIAS False MODEL.EXRERNAL_KNOWLEDGE.PATH False MODEL.EXRERNAL_KNOWLEDGE.ZERO_SHOT_PATH last_0.5_biased_zero_shot.pt

Path

python -m torch.distributed.launch --master_port 10302 --nproc_per_node=1 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 True MODEL.ROI_RELATION_HEAD.PREDICTOR GNNPredictorWithExternalKnowledge SOLVER.IMS_PER_BATCH 12 TEST.IMS_PER_BATCH 12 DTYPE "float32" SOLVER.MAX_ITER 50000 SOLVER.VAL_PERIOD 2000 SOLVER.CHECKPOINT_PERIOD 2000 GLOVE_DIR /local/home/{Your Name}/glove MODEL.PRETRAINED_DETECTOR_CKPT /local/home/{Your Name}/checkpoints/pretrained_faster_rcnn/model_final.pth OUTPUT_DIR /local/home/{Your Name}/checkpoints/G_last_0.5_without_biased_data_G+centeremb_1_layer MODEL.ROI_RELATION_HEAD.PREDICT_USE_BIAS False MODEL.EXRERNAL_KNOWLEDGE.GRAPH_WITH_NEIGHBOR False MODEL.EXRERNAL_KNOWLEDGE.GNN_LAYER 1 MODEL.EXRERNAL_KNOWLEDGE.GNN_MODE DGCNN MODEL.EXRERNAL_KNOWLEDGE.ZERO_SHOT_PATH last_0.5_biased_zero_shot.pt

Neighbor+Path

python -m torch.distributed.launch --master_port 10301 --nproc_per_node=1 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 True MODEL.ROI_RELATION_HEAD.PREDICTOR GNNPredictorWithExternalKnowledge SOLVER.IMS_PER_BATCH 12 TEST.IMS_PER_BATCH 12 DTYPE "float32" SOLVER.MAX_ITER 50000 SOLVER.VAL_PERIOD 2000 SOLVER.CHECKPOINT_PERIOD 2000 GLOVE_DIR /local/home/{Your Name}/glove MODEL.PRETRAINED_DETECTOR_CKPT /local/home/{Your Name}/checkpoints/pretrained_faster_rcnn/model_final.pth OUTPUT_DIR /local/home/{Your Name}/checkpoints/N+G_last_0.5_without_biased_data_G+centeremb_1_layer MODEL.ROI_RELATION_HEAD.PREDICT_USE_BIAS False MODEL.EXRERNAL_KNOWLEDGE.GRAPH_WITH_NEIGHBOR True MODEL.EXRERNAL_KNOWLEDGE.GNN_LAYER 1 MODEL.EXRERNAL_KNOWLEDGE.GNN_MODE DGCNN MODEL.EXRERNAL_KNOWLEDGE.ZERO_SHOT_PATH last_0.5_biased_zero_shot.pt

The original dataset

Neighbor


python -m torch.distributed.launch --master_port 10303 --nproc_per_node=1 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 True MODEL.ROI_RELATION_HEAD.PREDICTOR MotifPredictorWithExternalKnowledge SOLVER.IMS_PER_BATCH 12 TEST.IMS_PER_BATCH 12 DTYPE "float32" SOLVER.MAX_ITER 50000 SOLVER.VAL_PERIOD 2000 SOLVER.CHECKPOINT_PERIOD 2000 GLOVE_DIR /local/home/{Your Name}/glove MODEL.PRETRAINED_DETECTOR_CKPT /local/home/{Your Name}/checkpoints/pretrained_faster_rcnn/model_final.pth OUTPUT_DIR /local/home/{Your Name}/checkpoints/N+G_original_without_biased_data MODEL.ROI_RELATION_HEAD.PREDICT_USE_BIAS False MODEL.EXRERNAL_KNOWLEDGE.PATH False 

Path

python -m torch.distributed.launch --master_port 10304 --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 True MODEL.ROI_RELATION_HEAD.PREDICTOR GNNPredictorWithExternalKnowledge SOLVER.IMS_PER_BATCH 12 TEST.IMS_PER_BATCH 12 DTYPE "float32" SOLVER.MAX_ITER 50000 SOLVER.VAL_PERIOD 2000 SOLVER.CHECKPOINT_PERIOD 2000 GLOVE_DIR /local/home/{Your Name}/glove MODEL.PRETRAINED_DETECTOR_CKPT /local/home/{Your Name}/checkpoints/pretrained_faster_rcnn/model_final.pth OUTPUT_DIR /local/home/{Your Name}/checkpoints/G_original_without_biased_data_layer_1 MODEL.ROI_RELATION_HEAD.PREDICT_USE_BIAS False MODEL.EXRERNAL_KNOWLEDGE.GRAPH_WITH_NEIGHBOR False 

Neighbor+Path

python -m torch.distributed.launch --master_port 10301 --nproc_per_node=1 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 True MODEL.ROI_RELATION_HEAD.PREDICTOR GNNPredictorWithExternalKnowledge SOLVER.IMS_PER_BATCH 12 TEST.IMS_PER_BATCH 12 DTYPE "float32" SOLVER.MAX_ITER 50000 SOLVER.VAL_PERIOD 2000 SOLVER.CHECKPOINT_PERIOD 2000 GLOVE_DIR /local/home/{Your Name}/glove MODEL.PRETRAINED_DETECTOR_CKPT /local/home/{Your Name}/checkpoints/pretrained_faster_rcnn/model_final.pth OUTPUT_DIR /local/home/{Your Name}/checkpoints/N_original_without_biased_data_G+centeremb_1_layer MODEL.ROI_RELATION_HEAD.PREDICT_USE_BIAS False MODEL.EXRERNAL_KNOWLEDGE.GRAPH_WITH_NEIGHBOR True 

Citation

Please cite our paper if you find this code useful:

@misc{kan2021zeroshot,
      title={Zero-Shot Scene Graph Relation Prediction through Commonsense Knowledge Integration}, 
      author={Xuan Kan and Hejie Cui and Carl Yang},
      year={2021},
      eprint={2107.05080},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}