yolov3--进行darknet_53输出以后从特征获取预测结果

对darknet_53输出的特征层中取预测结果

这是上一篇文章，darknet_53的主干网络：https://blog.csdn.net/weixin_44403602/article/details/116241877

接着上文，在完成darknet_53以后呢，会得到三个输出out,out2,out3，这是主干网络提取的特征，需要对这些特征进行预测处理。主要流程图就是下方所示啦：

import torch
import torch.nn as nn
from collections import OrderedDict
import Darknet_53
import numpy as np

"""
这个文件主要用来处理，Darknet_53输出的三个out  ：  
out1 ---> b,256,52*52
out2 ---> b,512,26*26
out3 ---> b,1024,13*13

对out3进行五次卷积，将一部分return回去做上采样([b,256,26,26])，与out2进行cat，一部分进行预测输出[b,75,13,13]
对out2_new进行五次卷积，将一部分return回去做上采样([b,128,52,52])，与out1进行cat，一部分进行预测输出[b,75,26,26]
对out1_new进行五次卷积，进行预测输出[b,75,52,52]
"""


# def my_conv2d(in_channels, out_channels, kernel_size):
#     "将五次卷积中用到的卷积层封装成一个函数"
#     pad = (kernel_size - 1) // 2 if kernel_size else 0
#     conv2d = nn.Sequential(OrderedDict(
#         [
#             ("conv2d", nn.Conv2d(in_channels, out_channels, kernel_size, stride=1, padding=pad, bias=False)),
#             ("bn", nn.BatchNorm2d(out_channels)),
#             ("relu", nn.LeakyReLU(0.1))
#         ]))
#
#     return conv2d
# 这两个方法都是可行的
class my_conv2d(nn.Module):
	"""
	将五次卷积中用到的卷积层封装成一个类
	"""
    def __init__(self,in_channels, out_channels, kernel_size):
        super(my_conv2d, self).__init__()
        pad = (kernel_size - 1) // 2 if kernel_size else 0
        self.layer = conv2d = nn.Sequential(OrderedDict(
        [
            ("conv2d", nn.Conv2d(in_channels, out_channels, kernel_size, stride=1, padding=pad, bias=False