NICE for SGG in Pytorch

Our paper The Devil is in the Labels: Noisy Label Correction for Robust Scene Graph Generation has been accepted by CVPR 2022.

Installation

Check INSTALL.md for installation instructions.

Dataset

Check DATASET.md for instructions of dataset preprocessing.

Training Predictor in NSD

CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --master_port 10025 --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 MotifConfidencePredictor 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 checkpoints/motif-predcls-non-bg-reweight TYPE None ADD_BG False NEW_DATA_PATH new_data SOLVER.SCHEDULE.TYPE WarmupMultiStepLR SOLVER.BASE_LR 0.001

Extract Information from Original Dataset

CUDA_VISIBLE_DEVICES=0 CUDA_LAUNCH_BLOCKING=1 python -m torch.distributed.launch --master_port 10037 --nproc_per_node=1 tools/generate_has_rels_results.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/motif-precls-exmp OUTPUT_DIR checkpoints/motif-precls-exmp GROUPS tail GROUPS_T [0.6] ADD_BG True NEW_DATA_PATH new_data

Completing Missing Annotated Triplets by NSD

CUDA_VISIBLE_DEVICES=0 CUDA_LAUNCH_BLOCKING=1 python -m torch.distributed.launch --master_port 10031 --nproc_per_node=1 tools/generate_nsd_results.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 MotifConfidencePredictor TEST.IMS_PER_BATCH 1 DTYPE "float16" GLOVE_DIR glove MODEL.PRETRAINED_DETECTOR_CKPT checkpoints/motif-predcls-non-bg-reweight OUTPUT_DIR checkpoints/motif-predcls-non-bg-reweight GROUPS tail GROUPS_T [0.6] TYPE complete_bg ADD_BG False  NEW_DATA_PATH new_data

Generating New Dataset by NICE

CUDA_VISIBLE_DEVICES=0 CUDA_LAUNCH_BLOCKING=1 python -m torch.distributed.launch --master_port 10032 --nproc_per_node=1 tools/generate_nice_results.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/motif-precls-exmp OUTPUT_DIR checkpoints/motif-precls-exmp GROUPS tail GROUPS_T [0.6] TYPE extract_pos ADD_BG True NEW_DATA_PATH new_data

Training Models with NICE

CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --master_port 10025 --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 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 checkpoints/model_predcls_nice TYPE nice ADD_BG True NEW_DATA_PATH new_data

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.
GROUPS: It represents the predicate in which group to complete. The options are head, body, tail, or their combinations, separated by commas.
TYPE: The type of dataset loaded. If it set to 'complete_bg', it represents the loading of triplet samples of 'background' predicates. If it set to 'extract_pos', it means the loading of the positive triplet samples. If it set to 'nice', it means the loading of the samples generated by NICE.
GROUPS_T: Thresholds of confidence score of groups in 'GROUPS'.
ADD_BG: It means whether or not the sample loaded contains 'background'.
NEW_DATA_PATH: The path where NICE generates new datasets.

Models and Generated Files

For the Motifs Predictor in NSD and Motifs-NICE, we provide the trained models (checkpoint) for verification purpose. Please download from here* and unzip to checkpoints. Besides, we provide the files generated in new_data, you can download from here*.) The NSD checkpoint can be download from here*.

Citations

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

Credits

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

muktilin / NICE