I have not been able to reproduce the 35.2 mR@50 result of PCPL. The best validation accuracy for PredCls mR@50 I got was 24.92. From what I understand, nor can anybody else so far. For example, the SHA+GCL paper reports 22-24 in their supplemental materials (page 12). The command I used was based on your README.md:
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 48 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 output/motif-precls-exmp-pcpl LOG_TB True SOLVER.PRE_VAL True MODEL.PCPL_CENTER_LOSS True
My log is attached.