Why the retinaface(mnet backbone) training result is worse than yours?

**this is my result**
INFO:root:Epoch[200] Batch [0-80] Speed: 28.29 samples/sec RPNAcc_s32=0.994834 RPNAcc_s32_BG=0.999217 RPNAcc_s32_FG=0.832069 RPNL1Loss_s32=0.095835 RPNLandMarkL1Loss_s32=0.110469 RPNAcc_s16=0.993185 RPNAcc_s16_BG=0.998837 RPNAcc_s16_FG=0.813887 RPNL1Loss_s16=0.102131 RPNLandMarkL1Loss_s16=0.111187 RPNAcc_s8=0.989686 RPNAcc_s8_BG=0.999001 RPNAcc_s8_FG=0.437806 RPNL1Loss_s8=0.168240 RPNLandMarkL1Loss_s8=0.283297
**this is your result**
INFO:root:Epoch[80] Batch [20] Speed: 22.20 samples/sec RPNAcc_s32=0.996012 RPNAcc_s32_BG=0.999618 RPNAcc_s32_FG=0.860803 RPNL1Loss_s32=0.105018 RPNLandMarkL1Loss_s32=0.117736 RPNAcc_s16=0.994158 RPNAcc_s16_BG=0.998816 RPNAcc_s16_FG=0.832971 RPNL1Loss_s16=0.103016 RPNLandMarkL1Loss_s16=0.104968 RPNAcc_s8=0.990917 RPNAcc_s8_BG=0.998928 RPNAcc_s8_FG=0.497361 RPNL1Loss_s8=0.161664

my config as below:

INFO:root:===================
INFO:root:{'BBOX_MASK_THRESH': 0,
'CASCADE': 0,
'CASCADE_BBOX_STRIDES': [64, 32, 16, 8, 4],
'CASCADE_CLS_STRIDES': [64, 32, 16, 8, 4],
'CASCADE_MODE': 1,
'COLOR_JITTERING': 0.125,
'COLOR_MODE': 1,
'CONTEXT_FILTER_RATIO': 1,
'DENSE_ANCHOR': False,
'FACE_LANDMARK': True,
'FIXED_PARAMS': ['^stage1', '^.*upsampling'],
'HEAD_BOX': False,
'HEAD_FILTER_NUM': 64,
'HEAD_MODULE': 'SSH',
'IMAGE_STRIDE': 0,
'LANDMARK_LR_MULT': 2.5,
'LAYER_FIX': True,
'LR_MODE': 0,
'MIXUP': 0.0,
'MORE_SMALL_BOX': True,
'NET_MODE': 2,
'NUM_ANCHORS': 2,
'NUM_CLASSES': 2,
'NUM_CPU': 4,
'ORIGIN_SCALE': False,
'PIXEL_MEANS': array([0., 0., 0.]),
'PIXEL_SCALE': 1.0,
'PIXEL_STDS': array([1., 1., 1.]),
'PRE_SCALES': [(1200, 1600)],
'RANDOM_FEAT_STRIDE': False,
'RPN_ANCHOR_CFG': {'16': {'ALLOWED_BORDER': 9999,
'BASE_SIZE': 16,
'NUM_ANCHORS': 2,
'RATIOS': [1.0],
'SCALES': [8, 4]},
'32': {'ALLOWED_BORDER': 9999,
'BASE_SIZE': 16,
'NUM_ANCHORS': 2,
'RATIOS': [1.0],
'SCALES': [32, 16]},
'8': {'ALLOWED_BORDER': 9999,
'BASE_SIZE': 16,
'NUM_ANCHORS': 2,
'RATIOS': [1.0],
'SCALES': [2, 1]}},
'RPN_FEAT_STRIDE': [32, 16, 8],
'SCALES': [(640, 640)],
'SHARE_WEIGHT_BBOX': False,
'SHARE_WEIGHT_LANDMARK': False,
'TEST': {'BATCH_IMAGES': 1,
'CXX_PROPOSAL': True,
'HAS_RPN': False,
'IOU_THRESH': 0.5,
'NMS': 0.3,
'RPN_NMS_THRESH': 0.3,
'RPN_POST_NMS_TOP_N': 3000,
'RPN_PRE_NMS_TOP_N': 1000,
'SCORE_THRESH': 0.05},
'TRAIN': {'ASPECT_GROUPING': False,
'BATCH_IMAGES': 16,
'BBOX_STDS': [1.0, 1.0, 1.0, 1.0],
'CASCADE_OVERLAP': [0.4, 0.5],
'END2END': True,
'IMAGE_ALIGN': 0,
'LANDMARK_STD': 1.0,
'MIN_BOX_SIZE': 0,
'OHEM_MODE': 1,
'RPN_BATCH_SIZE': 256,
'RPN_CLOBBER_POSITIVES': False,
'RPN_ENABLE_OHEM': 2,
'RPN_FG_FRACTION': 0.25,
'RPN_FORCE_POSITIVE': False,
'RPN_NEGATIVE_OVERLAP': 0.3,
'RPN_POSITIVE_OVERLAP': 0.5},
'USE_3D': False,
'USE_BLUR': False,
'USE_CROP': True,
'USE_DCN': 0,
'USE_FPN': True,
'USE_MAXOUT': 0,
'USE_OCCLUSION': False,
'dataset': 'retinaface',
'max_feat_channel': 8888,
'network': 'mnet'}
origin image size 12880
retinaface_train gt roidb loaded from data/cache/retinaface_train_train_gt_roidb.pkl
roidb size 12876
INFO:root:retinaface_train append flipped images to roidb
flipped roidb size 25752
INFO:root:loading model/mobilenet,0,batchsize 16
[02:13:22] ../src/nnvm/legacy_json_util.cc:209: Loading symbol saved by previous version v1.3.0. Attempting to upgrade...
[02:13:22] ../src/nnvm/legacy_json_util.cc:217: Symbol successfully upgraded!
[02:13:22] ../src/storage/storage.cc:199: Using Pooled (Naive) StorageManager for CPU
=========lr_iters===========
[1609, 3219, 4828, 6438, 8047, 88522, 109446, 128760, 209235, 321900]
0.001
INFO:root:lr 0.001000 lr_epoch_diff [1, 2, 3, 4, 5, 55, 68, 80, 130, 200] lr_steps [(1609, 0.1), (3219, 0.1), (4828, 0.1), (6438, 0.1), (8047, 0.1), (88522, 0.1), (109446, 0.1), (128760, 0.1), (209235, 0.1), (321900, 0.1)]