Question about the number of training And the result of mAP

[xunhen]

When I try to reproduce the experiment, I find some difference between the code and paper.
If you can answer my questions, I will thank you very much.

1. the lr in code is 0.00025, but 0.001 in paper?
2. "In SGD training, 60K iterations are performed on 8 GPUs", but when I reproduce the experiment, I find the number of roidb entries is 219630.
   So the total numbers of iterations is 440k not 60k*8GPU=480k?
   
3. After eval my model retrained followed by the original set, the mAP is just 66.57%; however the mAP is 72.93% with the official trained model.
   

Is there anything wrong with my training?

The config is below:
training config:{'CLASS_AGNOSTIC': True,
'MXNET_VERSION': 'mxnet',
'SCALES': [(600, 1000)],
'TEST': {'BATCH_IMAGES': 1,
'CXX_PROPOSAL': True,
'EVAL_NUM_BATCH': 1000,
'HAS_RPN': True,
'KEY_FRAME_INTERVAL': 10,
'NMS': 0.3,
'RPN_MIN_SIZE': 0,
'RPN_NMS_THRESH': 0.7,
'RPN_POST_NMS_TOP_N': 300,
'RPN_PRE_NMS_TOP_N': 6000,
'max_per_image': 300,
'test_epoch': 2},
'TRAIN': {'ASPECT_GROUPING': True,
'BATCH_IMAGES': 1,
'BATCH_ROIS': -1,
'BATCH_ROIS_OHEM': 128,
'BBOX_MEANS': [0.0, 0.0, 0.0, 0.0],
'BBOX_NORMALIZATION_PRECOMPUTED': True,
'BBOX_REGRESSION_THRESH': 0.5,
'BBOX_STDS': [0.1, 0.1, 0.2, 0.2],
'BBOX_WEIGHTS': array([1., 1., 1., 1.]),
'BG_THRESH_HI': 0.5,
'BG_THRESH_LO': 0.0,
'CXX_PROPOSAL': True,
'ENABLE_OHEM': True,
'END2END': True,
'FG_FRACTION': 0.25,
'FG_THRESH': 0.5,
'FLIP': True,
'MAX_OFFSET': 0,
'MIN_OFFSET': -9,
'RESUME': False,
'RPN_BATCH_SIZE': 256,
'RPN_BBOX_WEIGHTS': [1.0, 1.0, 1.0, 1.0],
'RPN_CLOBBER_POSITIVES': False,
'RPN_FG_FRACTION': 0.5,
'RPN_MIN_SIZE': 0,
'RPN_NEGATIVE_OVERLAP': 0.3,
'RPN_NMS_THRESH': 0.7,
'RPN_POSITIVE_OVERLAP': 0.7,
'RPN_POSITIVE_WEIGHT': -1.0,
'RPN_POST_NMS_TOP_N': 300,
'RPN_PRE_NMS_TOP_N': 6000,
'SHUFFLE': True,
'begin_epoch': 0,
'end_epoch': 2,
'lr': 0.00025,
'lr_factor': 0.1,
'lr_step': '1.333',
'model_prefix': 'dff_rfcn_vid',
'momentum': 0.9,
'warmup': False,
'warmup_lr': 0,
'warmup_step': 0,
'wd': 0.0005},
'dataset': {'NUM_CLASSES': 31,
'dataset': 'ImageNetVID',
'dataset_path': 'E:\DataSet\ImageNetVID\ILSVRC2015',
'image_set': 'DET_train_30classes+VID_train_15frames',
'proposal': 'rpn',
'root_path': 'E:\DataSet\ImageNetVID',
'test_image_set': 'VID_val_videos',
'val_image_set': 'VID_val_videos'},
'default': {'frequent': 100, 'kvstore': 'device'},
'gpus': '0',
'network': {'ANCHOR_MEANS': [0.0, 0.0, 0.0, 0.0],
'ANCHOR_RATIOS': [0.5, 1, 2],
'ANCHOR_SCALES': [8, 16, 32],
'ANCHOR_STDS': [0.1, 0.1, 0.4, 0.4],
'DFF_FEAT_DIM': 1024,
'FIXED_PARAMS': ['conv1',
'bn_conv1',
'res2',
'bn2',
'gamma',
'beta'],
'IMAGE_STRIDE': 0,
'NORMALIZE_RPN': True,
'NUM_ANCHORS': 9,
'PIXEL_MEANS': array([103.06, 115.9 , 123.15]),
'RCNN_FEAT_STRIDE': 16,
'RPN_FEAT_STRIDE': 16,
'pretrained': '../../model/pretrained_model/resnet_v1_101',
'pretrained_epoch': 0,
'pretrained_flow': '../../model/pretrained_model/flownet'},
'output_path': './output/dff_rfcn/imagenet_vid',
'symbol': 'resnet_v1_101_flownet_rfcn'}

Note: When reproduce the experiment, I use only one gpu. Does it affect the experimental results?