Scene Graph Benchmark in Pytorch

Paper Unbiased Scene Graph Generation from Biased Training has been accepted by CVPR 2020 (Oral).

Recent Updates

TODO: Deployment on sagemaker

Quick Start

Check customs.ipynb to quick start. (No need to view the process behind)

Installation

sh gcc.sh to install gcc==7.3.0. Check INSTALL.md for installation instructions.

Dataset

Check DATASET.md for instructions of dataset preprocessing.

Metrics and Results (IMPORTANT)

Explanation of metrics in our toolkit and reported results are given in METRICS.md

Pretrained Models

Please download the Faster R-CNN model, extract all the files to the directory /home/username/checkpoints/pretrained_faster_rcnn. To train your own Faster R-CNN model, please follow the next section.

The above pretrained Faster R-CNN model achives 38.52/26.35/28.14 mAp on VG train/val/test set respectively.

Faster R-CNN pre-training

The following command can be used to train your own Faster R-CNN model:

CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --master_port 10001 --nproc_per_node=4 tools/detector_pretrain_net.py --config-file "configs/e2e_relation_detector_X_101_32_8_FPN_1x.yaml" SOLVER.IMS_PER_BATCH 8 TEST.IMS_PER_BATCH 4 DTYPE "float16" SOLVER.MAX_ITER 50000 SOLVER.STEPS "(30000, 45000)" SOLVER.VAL_PERIOD 2000 SOLVER.CHECKPOINT_PERIOD 2000 MODEL.RELATION_ON False OUTPUT_DIR /home/kaihua/checkpoints/pretrained_faster_rcnn SOLVER.PRE_VAL False

where CUDA_VISIBLE_DEVICES and --nproc_per_node represent the id of GPUs and number of GPUs you use, --config-file means the config we use, where you can change other parameters. SOLVER.IMS_PER_BATCH and TEST.IMS_PER_BATCH are the training and testing batch size respectively, DTYPE "float16" enables Automatic Mixed Precision supported by APEX, SOLVER.MAX_ITER is the maximum iteration, SOLVER.STEPS is the steps where we decay the learning rate, SOLVER.VAL_PERIOD and SOLVER.CHECKPOINT_PERIOD are the periods of conducting val and saving checkpoint, MODEL.RELATION_ON 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 and log (considering /home/username/checkpoints/pretrained_faster_rcnn), SOLVER.PRE_VAL means whether we conduct validation before training or not.

Perform training on Scene Graph Generation

There are three standard protocols: (1) Predicate Classification (PredCls): taking ground truth bounding boxes and labels as inputs, (2) Scene Graph Classification (SGCls) : using ground truth bounding boxes without labels, (3) Scene Graph Detection (SGDet): detecting SGs from scratch. We use two switches MODEL.ROI_RELATION_HEAD.USE_GT_BOX and MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL to select the protocols.

For Predicate Classification (PredCls), we need to set:

MODEL.ROI_RELATION_HEAD.USE_GT_BOX True MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL True

For Scene Graph Classification (SGCls):

MODEL.ROI_RELATION_HEAD.USE_GT_BOX True MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False

For Scene Graph Detection (SGDet):

MODEL.ROI_RELATION_HEAD.USE_GT_BOX False MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False

Predefined Models

We abstract various SGG models to be different relation-head predictors in the file roi_heads/relation_head/roi_relation_predictors.py, which are independent of the Faster R-CNN backbone and relation-head feature extractor. To select our predefined models, you can use MODEL.ROI_RELATION_HEAD.PREDICTOR.

For Neural-MOTIFS Model:

MODEL.ROI_RELATION_HEAD.PREDICTOR MotifPredictor

For Iterative-Message-Passing(IMP) Model (Note that SOLVER.BASE_LR should be changed to 0.001 in SGCls, or the model won't converge):

MODEL.ROI_RELATION_HEAD.PREDICTOR IMPPredictor

For VCTree Model:

MODEL.ROI_RELATION_HEAD.PREDICTOR VCTreePredictor

For our predefined Transformer Model (Note that Transformer Model needs to change SOLVER.BASE_LR to 0.001, SOLVER.SCHEDULE.TYPE to WarmupMultiStepLR, SOLVER.MAX_ITER to 16000, SOLVER.IMS_PER_BATCH to 16, SOLVER.STEPS to (10000, 16000).), which is provided by Jiaxin Shi:

MODEL.ROI_RELATION_HEAD.PREDICTOR TransformerPredictor

For Unbiased-Causal-TDE Model:

MODEL.ROI_RELATION_HEAD.PREDICTOR CausalAnalysisPredictor

The default settings are under configs/e2e_relation_X_101_32_8_FPN_1x.yaml and maskrcnn_benchmark/config/defaults.py. The priority is command > yaml > defaults.py

Customize Your Own Model

If you want to customize your own model, you can refer maskrcnn-benchmark/modeling/roi_heads/relation_head/model_XXXXX.py and maskrcnn-benchmark/modeling/roi_heads/relation_head/utils_XXXXX.py. You also need to add corresponding nn.Module in maskrcnn-benchmark/modeling/roi_heads/relation_head/roi_relation_predictors.py. Sometimes you may also need to change the inputs & outputs of the module through maskrcnn-benchmark/modeling/roi_heads/relation_head/relation_head.py.

The proposed Causal TDE on Unbiased Scene Graph Generation from Biased Training

As to the Unbiased-Causal-TDE, there are some additional parameters you need to know. MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_TYPE is used to select the causal effect analysis type during inference(test), where "none" is original likelihood, "TDE" is total direct effect, "NIE" is natural indirect effect, "TE" is total effect. MODEL.ROI_RELATION_HEAD.CAUSAL.FUSION_TYPE has two choice "sum" or "gate". Since Unbiased Causal TDE Analysis is model-agnostic, we support Neural-MOTIFS, VCTree and VTransE. MODEL.ROI_RELATION_HEAD.CAUSAL.CONTEXT_LAYER is used to select these models for Unbiased Causal Analysis, which has three choices: motifs, vctree, vtranse.

Note that during training, we always set MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_TYPE to be 'none', because causal effect analysis is only applicable to the inference/test phase.

Examples of the Training Command

Training Example 1 : (PreCls, Motif Model)

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 /home/kaihua/glove MODEL.PRETRAINED_DETECTOR_CKPT /home/kaihua/checkpoints/pretrained_faster_rcnn/model_final.pth OUTPUT_DIR /home/kaihua/checkpoints/motif-precls-exmp

where GLOVE_DIR is the directory used to save glove initializations, MODEL.PRETRAINED_DETECTOR_CKPT is the pretrained Faster R-CNN model you want to load, OUTPUT_DIR is the output directory used to save checkpoints and the log. Since we use the WarmupReduceLROnPlateau as the learning scheduler for SGG, SOLVER.STEPS is not required anymore.

Training Example 2 : (SGCls, Causal, TDE, SUM Fusion, MOTIFS Model)

CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --master_port 10026 --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 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 12 TEST.IMS_PER_BATCH 2 DTYPE "float16" SOLVER.MAX_ITER 50000 SOLVER.VAL_PERIOD 2000 SOLVER.CHECKPOINT_PERIOD 2000 GLOVE_DIR /home/kaihua/glove MODEL.PRETRAINED_DETECTOR_CKPT /home/kaihua/checkpoints/pretrained_faster_rcnn/model_final.pth OUTPUT_DIR /home/kaihua/checkpoints/causal-motifs-sgcls-exmp

Evaluation

Examples of the Test Command

Test Example 1 : (PreCls, Motif Model)

CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.launch --master_port 10027 --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 True MODEL.ROI_RELATION_HEAD.PREDICTOR MotifPredictor TEST.IMS_PER_BATCH 1 DTYPE "float16" GLOVE_DIR /home/kaihua/glove MODEL.PRETRAINED_DETECTOR_CKPT /home/kaihua/checkpoints/motif-precls-exmp OUTPUT_DIR /home/kaihua/checkpoints/motif-precls-exmp

Test Example 2 : (SGCls, Causal, TDE, SUM Fusion, MOTIFS Model)

CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.launch --master_port 10028 --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 CausalAnalysisPredictor MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_TYPE TDE MODEL.ROI_RELATION_HEAD.CAUSAL.FUSION_TYPE sum MODEL.ROI_RELATION_HEAD.CAUSAL.CONTEXT_LAYER motifs  TEST.IMS_PER_BATCH 1 DTYPE "float16" GLOVE_DIR /home/kaihua/glove MODEL.PRETRAINED_DETECTOR_CKPT /home/kaihua/checkpoints/causal-motifs-sgcls-exmp OUTPUT_DIR /home/kaihua/checkpoints/causal-motifs-sgcls-exmp

Examples of Pretrained Causal MOTIFS-SUM models

Examples of Pretrained Causal MOTIFS-SUM models on SGDet/SGCls/PredCls (batch size 12): (SGDet Download), (SGCls Download), (PredCls Download)

Corresponding Results (The original models used in the paper are lost. These are the fresh ones, so there are some fluctuations on the results. More results can be found in Reported Results):

Models	R@20	R@50	R@100	mR@20	mR@50	mR@100	zR@20	zR@50	zR@100
MOTIFS-SGDet-none	25.42	32.45	37.26	4.36	5.83	7.08	0.02	0.08	0.24
MOTIFS-SGDet-TDE	11.92	16.56	20.15	6.58	8.94	10.99	1.54	2.33	3.03
MOTIFS-SGCls-none	36.02	39.25	40.07	6.50	8.02	8.51	1.06	2.18	3.07
MOTIFS-SGCls-TDE	20.47	26.31	28.79	9.80	13.21	15.06	1.91	2.95	4.10
MOTIFS-PredCls-none	59.64	66.11	67.96	11.46	14.60	15.84	5.79	11.02	14.74
MOTIFS-PredCls-TDE	33.38	45.88	51.25	17.85	24.75	28.70	8.28	14.31	18.04

SGDet on Custom Images

Note that evaluation on custum images is only applicable for SGDet model, because PredCls and SGCls model requires additional ground-truth bounding boxes information. To detect scene graphs into a json file on your own images, you need to turn on the switch TEST.CUSTUM_EVAL and give a folder path that contains the custom images to TEST.CUSTUM_PATH. Only JPG files are allowed. The output will be saved as custom_prediction.json in the given DETECTED_SGG_DIR.

Test Example 1 : (SGDet, Causal TDE, MOTIFS Model, SUM Fusion) (checkpoint)

CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.launch --master_port 10027 --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 False MODEL.ROI_RELATION_HEAD.USE_GT_OBJECT_LABEL False MODEL.ROI_RELATION_HEAD.PREDICTOR CausalAnalysisPredictor MODEL.ROI_RELATION_HEAD.CAUSAL.EFFECT_TYPE TDE MODEL.ROI_RELATION_HEAD.CAUSAL.FUSION_TYPE sum MODEL.ROI_RELATION_HEAD.CAUSAL.CONTEXT_LAYER motifs TEST.IMS_PER_BATCH 1 DTYPE "float16" GLOVE_DIR /home/kaihua/glove MODEL.PRETRAINED_DETECTOR_CKPT /home/kaihua/checkpoints/causal-motifs-sgdet OUTPUT_DIR /home/kaihua/checkpoints/causal-motifs-sgdet TEST.CUSTUM_EVAL True TEST.CUSTUM_PATH /home/kaihua/checkpoints/custom_images DETECTED_SGG_DIR /home/kaihua/checkpoints/your_output_path

Test Example 2 : (SGDet, Original, MOTIFS Model, SUM Fusion) (same checkpoint)

CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.launch --master_port 10027 --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 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 TEST.IMS_PER_BATCH 1 DTYPE "float16" GLOVE_DIR /home/kaihua/glove MODEL.PRETRAINED_DETECTOR_CKPT /home/kaihua/checkpoints/causal-motifs-sgdet OUTPUT_DIR /home/kaihua/checkpoints/causal-motifs-sgdet TEST.CUSTUM_EVAL True TEST.CUSTUM_PATH /home/kaihua/checkpoints/custom_images DETECTED_SGG_DIR /home/kaihua/checkpoints/your_output_path

The output is a json file. For each image, the scene graph information is saved as a dictionary containing bbox(sorted), bbox_labels(sorted), bbox_scores(sorted), rel_pairs(sorted), rel_labels(sorted), rel_scores(sorted), rel_all_scores(sorted), where the last rel_all_scores give all 51 predicates probability for each pair of objects. The dataset information is saved as custom_data_info.json in the same DETECTED_SGG_DIR.

Visualize Detected SGs of Custom Images

To visualize the detected scene graphs of custom images, you can follow the jupyter note: visualization/3.visualize_custom_SGDet.jpynb. The inputs of our visualization code are custom_prediction.json and custom_data_info.json in DETECTED_SGG_DIR. They will be automatically generated if you run the above custom SGDet instruction successfully. Note that there may be too much trivial bounding boxes and relationships, so you can select top-k bbox and predicates for better scene graphs by change parameters box_topk and rel_topk.

HaoranLv/SGG