easy-Fpn源码解读(五)：rpn

easy-Fpn源码解读(五)：rpn

region_proposal_network.py代码解析

from typing import Tuple, List

import numpy as np
import torch
from torch import nn, Tensor
from torch.nn import functional as F

from bbox import BBox
from nms.nms import NMS


class RegionProposalNetwork(nn.Module):

    def __init__(self, num_features_out: int, anchor_ratios: List[Tuple[int, int]], anchor_scales: List[int], pre_nms_top_n: int, post_nms_top_n: int):
        super().__init__()

        self._features = nn.Sequential(
            nn.Conv2d(in_channels=num_features_out, out_channels=512, kernel_size=3, padding=1),
            nn.ReLU()
        )
        # 通道数从256维升至512维

        self._anchor_ratios = anchor_ratios
        self._anchor_scales = anchor_scales

        num_anchor_ratios = len(self._anchor_ratios)
        num_anchor_scales = len(self._anchor_scales)
        num_anchors = num_anchor_ratios * num_anchor_scales

        self._pre_nms_top_n = pre_nms_top_n
        self._post_nms_top_n = post_nms_top_n

        self._objectness = nn.Conv2d(in_channels=512, out_channels=num_anchors * 2, kernel_size=1)
        self._transformer = nn.Conv2d(in_channels=512, out_channels=num_anchors * 4, kernel_size=1)
        # 因为是1*1的kernel，所以不会改变特征图的size，
        # 对一个特征图像上的每一个点来说，会生成num*anchors个anchor，而每个anchor都有两个前景/背景概率
        # 和4个坐标位置，因此对于self._objectness来说两个通道对应一个anchor，而对于self._transformer来说
        # 4个通道对应一个anchor。这4个通道是一个anchor的四个位置的修正偏移量。
        # 所以在forward的里，才会将通道数换至最后一维，再展开。这样每一行就会对应一个anchor

    def forward(self, features: Tensor, image_width: int, image_height: int) -> Tuple[Tensor, Tensor]:
        features = self._features(features)
        objectnesses = self._objectness(features)  # anchor前景/背景修正
        transformers = self._transformer(features)  # anchor位置偏移量修正

        # permute函数将tensor的维度换位:num*CWH -> num*WHC
        #
        # contiguous()一般在permu