Faster RCNN training result problem

[hktxt]

I tried to train Faster-RCNN from scratch. The train.py gives me this result:

{
"main/rpn_loc_loss": 0.045347146689891815,
"main/rpn_cls_loss": 0.17269396781921387,
"main/roi_loc_loss": 0.27309778332710266,
"main/roi_cls_loss": 0.49079766869544983,
"main/loss": 0.9819365739822388,
"lr": 0.0001,
"validation/main/map": 0.12205098147362133,
"validation/main/ap/aeroplane": 0.2037050348560657,
"validation/main/ap/bicycle": 0.15355473865782293,
"validation/main/ap/bird": 0.1069814741360702,
"validation/main/ap/boat": 0.08437066002057338,
"validation/main/ap/bottle": 0.0303030303030303,
"validation/main/ap/bus": 0.13091526733063397,
"validation/main/ap/car": 0.2179439900145522,
"validation/main/ap/cat": 0.1788983192876631,
"validation/main/ap/chair": 0.027534965034965036,
"validation/main/ap/cow": 0.10386202324629096,
"validation/main/ap/diningtable": 0.1376211067035571,
"validation/main/ap/dog": 0.16679011471811683,
"validation/main/ap/horse": 0.18906283713577696,
"validation/main/ap/motorbike": 0.22127384038737857,
"validation/main/ap/person": 0.20588034722737675,
"validation/main/ap/pottedplant": 0.0101010101010101,
"validation/main/ap/sheep": 0.006993006993006993,
"validation/main/ap/sofa": 0.03462976657260593,
"validation/main/ap/train": 0.13475834327905686,
"validation/main/ap/tvmonitor": 0.09583975346687211,
"epoch": 13,
"iteration": 70000,
"elapsed_time": 43275.40101284254
}

the mAP is extremely low... 0.12...
I changed nothing, only the path of VOC dataset instead of download.
I'am sure the test dataset returns use_difficult=True, return_difficult=True.
and DetectionVOCEvaluator() uses use_07_metric=True.
These are default setting in train.py.
I check some issues, some had a lower mAP than reported. However, my result is wrong...
I did not use the pretrained model. To confirm evaluate code, I downloaded the pretrained model from: 'https://chainercv-models.preferred.jp/faster_rcnn_vgg16_voc07_trained_2018_06_01.npz' and tested it, had a result of 0.706.
So, it seems like training porblem. Is the epoch of 13 not enough? should I train more epoches?

[Hakuyume]

Could you share the version of ChainerCV and the commit id of the training code?

[hktxt]

@Hakuyume
The version is v6.
I cloned ChainerCV repo yesterday. So it's the latest commit.
c22c451