Lower performance for faster-rcnn

[Yun-960]

Thank you for your good work!
But I still can’t reach your score when I trained faster-rcnn. All my configurations are consistent with your default configurations. I don't know why.
My Environment:
CUDA10.0
pytorch1.0.0
python3.6
My configurations with 3 gpu:
DATASET:
NAME: "vg"
MODE: "benchmark"
PATH: "datasets/vg_bm"
TRAIN_BATCH_SIZE: 6
TEST_BATCH_SIZE: 1
MODEL:
WEIGHT_IMG: "catalog://ImageNetPretrained/MSRA/R-101"
WEIGHT_DET: ""
ALGORITHM: "faster_rcnn"
BACKBONE:
CONV_BODY: "R-101-C4"
ROI_BOX_HEAD:
NUM_CLASSES: 151
ROI_RELATION_HEAD:
NUM_CLASSES: 51
SOLVER:
BASE_LR: 5e-3
MAX_ITER: 100000
STEPS: (70000,90000)
CHECKPOINT_PERIOD: 5000
My performance: