inconsistant anchor reference (may be the cause of the lower accuracy than faster-rcnn)

Hi,

Please notice you have an inconsistant reference to the order of the anchors (lines 86 and 79)

https://github.com/jwyang/fpn.pytorch/blob/23bd1d2fa09fbb9453f11625d758a61b9d600942/lib/model/rpn/rpn_fpn.py#L86

https://github.com/jwyang/fpn.pytorch/blob/23bd1d2fa09fbb9453f11625d758a61b9d600942/lib/model/rpn/rpn_fpn.py#L79

lets say k is the number of anchors

than in line 79 you are doing softmax after reshape which make anchors:
0:k-1 assosiated with proposal = false
k:2k-1 assosiated with proposal = True.

and in line 86 your anchors are arranged in a different way:
i mod 2 == 0 associated with proposal = false
i mod 2 == 1 associated with proposal = True.

I propose to reshape the score in this way to be consistant.
(You will also need to change few other things in proposalLayer for it to work)

`rpn_cls_scores.append(rpn_cls_score_reshape.permute(0, 2, 3, 1).contiguous().view(batch_size, -1, 2))`

Start in lib/model/rpn/rpn_fpn.py at the referenced lines 79 and 86, and compare the anchor ordering before and after the score reshape. Trace proposalLayer to identify the other ordering-dependent code paths mentioned in the issue. Done means the anchor-to-proposal-false/true mapping is consistent throughout the RPN and the related behavior is validated.

Written by the indexing model from the issue text.