（二）随便起个名字别人就不能通过关键字搜索到了

这是YOLO V3的代码，只有网络模型的，嘿嘿嘿嘿嘿

1.darknet.py

import tensorflow as tf
from tensorflow.keras import layers, Sequential
from tensorflow.keras.initializers import RandomNormal
from tensorflow.keras.regularizers import l1, l2



def Conv_BN_Leaky(filters, ks, strides=(1,1)):
    lay = Sequential([
                        layers.Conv2D(filters, ks, strides=strides, padding="same",
                                      kernel_initializer=tf.random_normal_initializer(stddev=0.02), kernel_regularizer=l2(5e-4), use_bias=False),
                        layers.BatchNormalization(),
                        layers.LeakyReLU(alpha=0.1)
                    ])
    return lay


def resblock_body(inp, filters, blocks):
    # x = layers.ZeroPadding2D(((1,0), (1,0)))(inp)
    x = Conv_BN_Leaky(filters, (3, 3), strides=(2, 2))(inp)
    for i in range(blocks):
        y = Conv_BN_Leaky(filters//2, (1,1))(x)         # 降维
        y = Conv_BN_Leaky(filters//2, (3, 3))(y)        # 卷积
        y = Conv_BN_Leaky(filters, (1,1))(y)            # 升维
        x = layers.add([x, y])                        # 残差相加
    return x



def darknet_body(x):
    # 1.先进行一次3x3卷积
    x = Conv_BN_Leaky(filters=32, ks=(3,3))(x)
    # 2.resblock_body x 1次
    x = resblock_body(inp=x, filters=64, blocks=1)
    # 3.resblock_body x 2次
    x = resblock_body(inp=x, filters=128, blocks=2)
    # 4.resblock_body x 8次
    x = resblock_body(inp=x, filters=256, blocks=8)
    feat1 = x
    # 5.resblock_body x 8次
    x = resblock_body(inp=x, filters=512, blocks=8)
    feat2 = x
    # 6.resblock_body x 4次
    x = resblock_body(inp=x, filters=1024, blocks=4)
    feat3 = x
    return feat1, feat2, feat3


if __name__ == '__main__':
    from tensorflow.keras.layers import Input
    inputs = Input(shape=(416, 416, 3))
    feat1, feat2, feat3 = darknet_body(inputs)
    print(feat3.shape)

2:   yolo.py

import tensorflow as tf
from tensorflow.keras import Sequential,layers
from darknet import *

def make_five_conv(x, filters):
    x= Conv_BN_Leaky(filters, (1,1))(x)
    x= Conv_BN_Leaky(filters*2, (3,3))(x)
    x= Conv_BN_Leaky(filters, (1,1))(x)
    x= Conv_BN_Leaky(filters*2, (3,3))(x)
    x= Conv_BN_Leaky(filters, (1,1))(x)
    return x


def last_conv(x, out_filters):
    x = Conv_BN_Leaky(out_filters, (3,3))(x)
    y = layers.Conv2D(out_filters, (1, 1), padding="same", kernel_initializer=tf.random_normal_initializer(stddev=0.02), kernel_regularizer=l2(5e-4), use_bias=False)(x)
    return y

def YOLO_V3(inputs, nc):
    feat_big, feat_middle, feat_small = darknet_body(inputs)
    # 第一部分：
    feat_small = make_five_conv(feat_small, 512)
    print("feat small shape:", feat_small)
    out_small = last_conv(feat_small, nc*3)

    # 第二部分：
    feat_small = Conv_BN_Leaky(256, (1,1))(feat_small)
    up_middle = layers.UpSampling2D()(feat_small)
    middle_cat = layers.Concatenate()([up_middle, feat_middle])
    middle_set = make_five_conv(middle_cat, 256)
    out_middle = last_conv(middle_set, nc*3)

    # 第三部分：
    feat_middle = Conv_BN_Leaky(128, (1,1))(middle_set)
    up_big = layers.UpSampling2D()(feat_middle)
    big_cat = layers.Concatenate()([up_big, feat_big])
    big_set = make_five_conv(big_cat, 128)
    out_big = last_conv(big_set, nc*3)
    return out_small, out_middle, out_big


if __name__ == '__main__':
    from tensorflow.keras.layers import Input
    inputs = Input(shape=(416, 416, 3))
    feat1, feat2, feat3 = YOLO_V3(inputs, nc=20)
    print(feat3.shape)