YOLO V1网络框架

任务描述

本关任务：阅读下面的相关知识，理解YOLO V1的网络框架。

相关知识

网络设计

网络结构借鉴了 GoogLeNet 。24个卷积层，2个全链接层。（用1×1 reduction layers 紧跟 3×3 convolutional layers 取代Goolenet的 inception modules ）

网络结构

YOLO检测网络包含24个卷积层（用来提取特征）和2个全联接层（用来预测图像位置和类别置信度），并且使用了大量的1×1的卷积用来降低上一层的layer到下一层的特征空间。并且在paper中，作者还给出了fast-YOLO的构架，即：9个卷积层和2个全连接层。使用titan x GPU，fast YOLO可以达到155fps的检测速度，但是mAP值也从YOLO的63.4% 降到了 52.7% ，但却仍然远高于以往的实时物体检测方法（DPM）的mAP值。

（这里一开始我数不出来为什么有24个卷积层，其实要仔细看图，你会发现有一些“×4”“×2”，把这些考虑进去你就会发现确实是24层）

最后得到的7×7×30代表的是最后的输出，代表一共49个网格，每个网格拥有30个值，其中有20个值为类别概率值，即该网格检测出来的属于某类物体的概率。而剩下的10个值可以分成两部分，分表代表网格两个Bounding Box各自的参数部分。 我们取一个cell来看，一个cell有30个元素，如图所示：

网络参数展示图：

使用keras实现网络结构：

1. # 卷积层
2. # 根据卷积计算，输出形状：(55, 55, 64)
3. tf.keras.layers.Conv2D(64,(6, 6), strides=4, input_shape=(227, 227, 3), activation=tf.keras.layers.LeakyReLU(0.1)),
5. # 扁平化操作（`Flatten`）：传递给全连接层的卷积层输出必须先扁平化，然后才能全连接层接受输入。
6. # 输出形状：(193600, )
7. tf.keras.layers.Flatten(),
9. # 全连接层
10. # 输出形状：(4096, )
11. tf.keras.layers.Dense(4096, activation=tf.keras.layers.LeakyReLU(0.1)),
13. # Dropout防止过拟合
14. tf.keras.layers.Dropout(0.6),
16. # 全连接层
17. # 输出形状：(30, )
18. tf.keras.layers.Dense(30, activation=tf.keras.layers.LeakyReLU(0.1))

编程要求

根据左侧内容提示，在右侧编辑器补充YOLO V1网络框架代码。

测试说明

平台会对你编写的代码进行测试：

测试输入： 略

预期输出： 网络结构：

1. _________________________________________________________________
2. Layer (type) Output Shape Param #
3. =================================================================
4. conv2d_1 (Conv2D) (None, 224, 224, 64) 1792
5. _________________________________________________________________
6. max_pooling2d_1 (MaxPooling2 (None, 112, 112, 64) 0
7. _________________________________________________________________
8. conv2d_2 (Conv2D) (None, 112, 112, 192) 110784
9. _________________________________________________________________
10. max_pooling2d_2 (MaxPooling2 (None, 56, 56, 192) 0
11. _________________________________________________________________
12. conv2d_3 (Conv2D) (None, 56, 56, 128) 24704
13. _________________________________________________________________
14. conv2d_4 (Conv2D) (None, 56, 56, 256) 295168
15. _________________________________________________________________
16. conv2d_5 (Conv2D) (None, 56, 56, 256) 65792
17. _________________________________________________________________
18. conv2d_6 (Conv2D) (None, 56, 56, 512) 1180160
19. _________________________________________________________________
20. max_pooling2d_3 (MaxPooling2 (None, 28, 28, 512) 0
21. _________________________________________________________________
22. conv2d_7 (Conv2D) (None, 28, 28, 256) 131328
23. _________________________________________________________________
24. conv2d_8 (Conv2D) (None, 28, 28, 512) 1180160
25. _________________________________________________________________
26. conv2d_9 (Conv2D) (None, 28, 28, 256) 131328
27. _________________________________________________________________
28. conv2d_10 (Conv2D) (None, 28, 28, 512) 1180160
29. _________________________________________________________________
30. conv2d_11 (Conv2D) (None, 28, 28, 256) 131328
31. _________________________________________________________________
32. conv2d_12 (Conv2D) (None, 28, 28, 512) 1180160
33. _________________________________________________________________
34. conv2d_13 (Conv2D) (None, 28, 28, 256) 131328
35. _________________________________________________________________
36. conv2d_14 (Conv2D) (None, 28, 28, 512) 1180160
37. _________________________________________________________________
38. conv2d_15 (Conv2D) (None, 28, 28, 512) 262656
39. _________________________________________________________________
40. conv2d_16 (Conv2D) (None, 28, 28, 1024) 525312
41. _________________________________________________________________
42. max_pooling2d_4 (MaxPooling2 (None, 14, 14, 1024) 0
43. _________________________________________________________________
44. conv2d_17 (Conv2D) (None, 14, 14, 512) 524800
45. _________________________________________________________________
46. conv2d_18 (Conv2D) (None, 14, 14, 1024) 4719616
47. _________________________________________________________________
48. conv2d_19 (Conv2D) (None, 14, 14, 512) 524800
49. _________________________________________________________________
50. conv2d_20 (Conv2D) (None, 14, 14, 1024) 4719616
51. _________________________________________________________________
52. conv2d_21 (Conv2D) (None, 14, 14, 1024) 9438208
53. _________________________________________________________________
54. conv2d_22 (Conv2D) (None, 7, 7, 1024) 9438208
55. _________________________________________________________________
56. conv2d_23 (Conv2D) (None, 7, 7, 1024) 9438208
57. _________________________________________________________________
58. conv2d_24 (Conv2D) (None, 7, 7, 1024) 9438208
59. _________________________________________________________________
60. flatten_1 (Flatten) (None, 50176) 0
61. _________________________________________________________________
62. dense_1 (Dense) (None, 4096) 205524992
63. _________________________________________________________________
64. dropout_1 (Dropout) (None, 4096) 0
65. _________________________________________________________________
66. dense_2 (Dense) (None, 1470) 6022590
67. =================================================================
68. Total params: 267,501,566
69. Trainable params: 267,501,566
70. Non-trainable params: 0

---

开始你的任务吧，祝你成功！

参考代码：

import tensorflow as tf


def createYOLO_v1_Model(tiny=True):
    if tiny:
        # 序贯模型（Sequential):单输入单输出，一条路通到底，层与层之间只有相邻关系，没有跨层连接。
        models = tf.keras.Sequential([
            # 卷积层相关参数:卷积核64个，卷积核大小为3*3，步长为2，填充为same，输入图片尺寸为448x448x3，而激活函数为LeakyReLU，超参数为0.1
            tf.keras.layers.Conv2D(64, (3, 3), strides=2, padding='same', input_shape=(448, 448, 3), activation=tf.keras.layers.LeakyReLU(0.1)),
            # 池化层相关参数：池化大小为2*2，步长为2
            tf.keras.layers.MaxPooling2D((2, 2), strides=2),

            tf.keras.layers.Conv2D(192, (3, 3), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            tf.keras.layers.MaxPooling2D((2, 2), strides=2),

            tf.keras.layers.Conv2D(128, (1, 1), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            tf.keras.layers.Conv2D(256, (3, 3), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            tf.keras.layers.Conv2D(256, (1, 1), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            tf.keras.layers.Conv2D(512, (3, 3), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            # 请在此添加代码，根据左侧展示图补充池化层代码
            ########## Begin ##########
            tf.keras.layers.MaxPooling2D((2, 2), strides=2),
            ########## End ##########    



            tf.keras.layers.Conv2D(256, (1, 1), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            tf.keras.layers.Conv2D(512, (3, 3), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            tf.keras.layers.Conv2D(256, (1, 1), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            tf.keras.layers.Conv2D(512, (3, 3), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            tf.keras.layers.Conv2D(256, (1, 1), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            tf.keras.layers.Conv2D(512, (3, 3), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            tf.keras.layers.Conv2D(256, (1, 1), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            tf.keras.layers.Conv2D(512, (3, 3), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            tf.keras.layers.Conv2D(512, (1, 1), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            tf.keras.layers.Conv2D(1024, (1, 1), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            tf.keras.layers.MaxPooling2D((2, 2), strides=2),


            tf.keras.layers.Conv2D(512, (1, 1), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            tf.keras.layers.Conv2D(1024, (3, 3), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            tf.keras.layers.Conv2D(512, (1, 1), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            tf.keras.layers.Conv2D(1024, (3, 3), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            tf.keras.layers.Conv2D(1024, (3, 3), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            # 请在此添加代码，根据左侧展示图补充卷积层代码
            ########## Begin ##########
            tf.keras.layers.Conv2D(1024, (3, 3), strides=2, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            tf.keras.layers.Conv2D(1024, (3, 3), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            tf.keras.layers.Conv2D(1024, (3, 3), strides=1, padding='same', activation=tf.keras.layers.LeakyReLU(0.1)),
            ########## End ##########   

            # 请在此添加代码，根据左侧内容提示补充扁平层代码
            ########## Begin ##########
            tf.keras.layers.Flatten(),
            ########## End ##########    

            # 全连接层
            tf.keras.layers.Dense(4096, activation=tf.keras.layers.LeakyReLU(0.1)),
            # 请在此添加代码，根据左侧内容提示补充Dropout层代码，超参数是0.5，可以缓解过拟合操作
            ########## Begin ##########
            tf.keras.layers.Dropout(0.5),
            ########## End ##########    

            # 输出一个7*7*30的张量
            tf.keras.layers.Dense(7*7*30)])
        return models


model = createYOLO_v1_Model()
print('网络结构：')
model.summary()