​论文名称：YOLOv3: An Incremental Improvement
论文下载地址: https://arxiv.org/abs/1804.02767
代码仓库地址：YOLOv3 SPP网络代码仓库(Pytorch版)
参考仓库地址：keras-yolo3

参考视频：
\quad YOLO系列理论视频合集
\quad YOLOv3 SPP网络源码讲解(Pytorch版)

目录

• 1、数据文件结构 （yolo格式）
• 2、网络结构
• 3、配置文件 .cfg 解读
• • 1）convolutional层
  • 2）shortcut层
  • 3）maxpool层
  • 4）route层
  • 5）upsample层
  • 6）yolo层

1、数据文件结构 （yolo格式）

yolo 格式：

• 标注数据以 .txt 文件进行存储，
• 每个.txt 文件对应一张图像
• .txt 文件中的每一行对应该图像中一个 object
• 每行的格式是：class， x_center， y_center， width， height
• 方框的坐标（ x_center， y_center， width， height）必须要归一化到 0 ～ 1，如果方框以像素为单位，请将x_center和width除以图像宽度，将y_center 和 height除以图像高度。
• 类别class 必须从0开始
  

2、网络结构

我用 netron 可视化了网络，图像太长，我放在了文章的最后 （最最后，还附有用 pycharm 输出的网络模型）

网络最后的输出 （以 pascal voc 为例， 有20个类别）：
1）若原图像尺寸为 512×512512 \times 512512×512 ，网络的3个尺度输出分别为（16， 16， 75）， （32， 32， 75）， （64， 64， 75）
2）若原图像尺寸为 416×416416 \times 416416×416 ，网络的3个尺度输出分别为（13， 13， 75）， （26， 26， 75）， （52， 52， 75）

3个尺度表示 原图被缩减了 8倍， 16倍， 32倍
75 表示 4个坐标信息(center_x, center_y, weight, height) + 1个置信度信息 + 20个类别的概率

3、配置文件 .cfg 解读

图片来源于：霹雳吧啦Wz 的bilibili视频讲解

1）convolutional层

注意 pad 的取值规则：pad=1 表示 padding模式为same， 即 padding=kernel_size//2 ；； pad=0 表示不添加padding

2）shortcut层

将上一层的输出与 第 -3 层的输出进行 跨层连接，也就是 add 操作，需要保证 上一层的输出与 第-3层的输出的 尺寸完全一样。

3）maxpool层

默认: padding = (kernel_size) // 2

4）route层

a、layers 只有一个值， 比如 layers=-2。  表示返回到 第-2层，或者说指向 第-2层， 后面的操作会在 第-2层的基础上继续
b、layers 有多个值， 比如 layer = -1， -3， -5， -6。  表示 拼接(concatnate) 第 -1， -3， -5， -6 层

5）upsample层

6）yolo层

classes 表示类别数， COCO 数据集的类别数为80， PASCAL VOC 类别为20
mask 是 anchors 的索引，指明用哪几个 anchors， 比如 mask=0， 1， 2， 表示使用 前3个anchors
anchors 表示 anchor 的尺寸，一共生成3个尺度的检测输出， 每个尺度的分别有 3个尺寸的 anchor。

netron 可视化 网络结构

** 如下，WeightedFeatureFusion 就是 shortcut；\quad FeatureConcat 就是 concatenate

ModuleList((0): Sequential((Conv2d): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(1): Sequential((Conv2d): Conv2d(32, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(2): Sequential((Conv2d): Conv2d(64, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(3): Sequential((Conv2d): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(4): WeightedFeatureFusion()(5): Sequential((Conv2d): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(6): Sequential((Conv2d): Conv2d(128, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(7): Sequential((Conv2d): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(8): WeightedFeatureFusion()(9): Sequential((Conv2d): Conv2d(128, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(10): Sequential((Conv2d): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(11): WeightedFeatureFusion()(12): Sequential((Conv2d): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(13): Sequential((Conv2d): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(14): Sequential((Conv2d): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(15): WeightedFeatureFusion()(16): Sequential((Conv2d): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(17): Sequential((Conv2d): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(18): WeightedFeatureFusion()(19): Sequential((Conv2d): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(20): Sequential((Conv2d): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(21): WeightedFeatureFusion()(22): Sequential((Conv2d): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(23): Sequential((Conv2d): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(24): WeightedFeatureFusion()(25): Sequential((Conv2d): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(26): Sequential((Conv2d): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(27): WeightedFeatureFusion()(28): Sequential((Conv2d): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(29): Sequential((Conv2d): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(30): WeightedFeatureFusion()(31): Sequential((Conv2d): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(32): Sequential((Conv2d): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(33): WeightedFeatureFusion()(34): Sequential((Conv2d): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(35): Sequential((Conv2d): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(36): WeightedFeatureFusion()(37): Sequential((Conv2d): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(38): Sequential((Conv2d): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(39): Sequential((Conv2d): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(40): WeightedFeatureFusion()(41): Sequential((Conv2d): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(42): Sequential((Conv2d): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(43): WeightedFeatureFusion()(44): Sequential((Conv2d): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(45): Sequential((Conv2d): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(46): WeightedFeatureFusion()(47): Sequential((Conv2d): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(48): Sequential((Conv2d): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(49): WeightedFeatureFusion()(50): Sequential((Conv2d): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(51): Sequential((Conv2d): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(52): WeightedFeatureFusion()(53): Sequential((Conv2d): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(54): Sequential((Conv2d): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(55): WeightedFeatureFusion()(56): Sequential((Conv2d): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(57): Sequential((Conv2d): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(58): WeightedFeatureFusion()(59): Sequential((Conv2d): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(60): Sequential((Conv2d): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(61): WeightedFeatureFusion()(62): Sequential((Conv2d): Conv2d(512, 1024, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(63): Sequential((Conv2d): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(64): Sequential((Conv2d): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(65): WeightedFeatureFusion()(66): Sequential((Conv2d): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(67): Sequential((Conv2d): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(68): WeightedFeatureFusion()(69): Sequential((Conv2d): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(70): Sequential((Conv2d): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(71): WeightedFeatureFusion()(72): Sequential((Conv2d): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(73): Sequential((Conv2d): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(74): WeightedFeatureFusion()(75): Sequential((Conv2d): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(76): Sequential((Conv2d): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(77): Sequential((Conv2d): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(78): MaxPool2d(kernel_size=5, stride=1, padding=2, dilation=1, ceil_mode=False)(79): FeatureConcat()(80): MaxPool2d(kernel_size=9, stride=1, padding=4, dilation=1, ceil_mode=False)(81): FeatureConcat()(82): MaxPool2d(kernel_size=13, stride=1, padding=6, dilation=1, ceil_mode=False)(83): FeatureConcat()(84): Sequential((Conv2d): Conv2d(2048, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(85): Sequential((Conv2d): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(86): Sequential((Conv2d): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(87): Sequential((Conv2d): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(88): Sequential((Conv2d): Conv2d(1024, 75, kernel_size=(1, 1), stride=(1, 1)))(89): YOLOLayer()(90): FeatureConcat()(91): Sequential((Conv2d): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(92): Upsample(scale_factor=2.0, mode=nearest)(93): FeatureConcat()(94): Sequential((Conv2d): Conv2d(768, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(95): Sequential((Conv2d): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(96): Sequential((Conv2d): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(97): Sequential((Conv2d): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(98): Sequential((Conv2d): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(99): Sequential((Conv2d): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(100): Sequential((Conv2d): Conv2d(512, 75, kernel_size=(1, 1), stride=(1, 1)))(101): YOLOLayer()(102): FeatureConcat()(103): Sequential((Conv2d): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(104): Upsample(scale_factor=2.0, mode=nearest)(105): FeatureConcat()(106): Sequential((Conv2d): Conv2d(384, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(107): Sequential((Conv2d): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(108): Sequential((Conv2d): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(109): Sequential((Conv2d): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(110): Sequential((Conv2d): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(111): Sequential((Conv2d): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(BatchNorm2d): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activation): LeakyReLU(negative_slope=0.1, inplace=True))(112): Sequential((Conv2d): Conv2d(256, 75, kernel_size=(1, 1), stride=(1, 1)))(113): YOLOLayer()
)