https://www.bilibili.com/video/av43996494/?p=2
0上一节没写完的代码
#!usr/bin/python
# -*- coding: utf-8 -*-
# Creation Date: 2019/7/10
import tensorflow as tf
import numpy as np
import cv2
class ssd(object):
def __init__(self):
pass # 先略过
# ===>l2正则化<===
def l2norm(self, x, scale, trainable=True, scope='L2Normalization'):
n_channels = x.get_shape().as_list()[-1] # 通道数. 得到形状,变成列表,取后一个
l2_norm = tf.nn.l2_normalize(x, dim=[3], epsilon=1e-12) # 只对每个像素点在channels上做归一化
with tf.variable_scope(scope):
gamma = tf.get_variable("gamma", shape=[n_channels, ], dtype=tf.float32,
initializer=tf.constant_initializer(scale),
trainable=trainable)
return l2_norm * gamma
# ===>下面开始定义所需组件<===
# conv2d, max_pool2d, pad2d, dropout
# 定义一个卷积的操作 1输入 2卷积核个数 3卷积核大小 4步长 5padding 6膨胀 7激活函数 8名字
def conv2d(self, x, filter, # 输入x, 卷积核的个数filter
k_size, stride=[1, 1], # k_size卷积核是几*几，步长stride
padding='same', dilation=[1, 1], # padding, 空洞卷积指数这里1代表正常卷积
activation=tf.nn.relu, scope='conv2d'): # 激活函数relu, 名字scope
return tf.layers.conv2d(input=x, filters=filter, kernel_size=k_size,
strides=stride, padding=padding, dilation_rate=dilation,
name=scope, activation=activation)
def max_pool2d(self, x, pool_size, stride, scope='max_pool2d'): # 我猜padding是vaild
return tf.layers.max_pooling2d(inputs=x, pool_size=pool_size, strides=stride, padding='valid', name=scope)
# 用于填充s=2的第8,9层. 从6层往后的卷积层需要自己填充, 不要用它自带的填充.
def pad2d(self, x, pad):
return tf.pad(x, paddings=[[0, 0], [pad, pad], [pad, pad], [0, 0]])
def dropout(self, x, d_rate=0.5):
return tf.layers.dropout(inputs=x, rate=d_rate)
# ===>下面开始写网络架构<===
def set_net(self):
check_points = {} # 装特征层的字典,用于循环迭代
x = tf.placeholder(dtype=tf.float32, shape=[None, 300, 300, 3])
with tf.variable_scope('ssd_300_vgg'):
# ===>VGG前5层<===
# b1
net = self.conv2d(x, filter=64, k_size=[3, 3], scope='conv1_1') # 64个3*3卷积核, s=1 默认，标准卷积
net = self.conv2d(net, 64, [3, 3], scope='conv1_2') # 64个3*3卷积核, s=1默认
net = self.max_pool2d(net, pool_size=[2, 2], stride=[2, 2]) # 池化层2*2卷积核, s=2 默认，池化层一般都是2
# b2
net = self.conv2d(net, filter=128, k_size=[3, 3], scope='conv2_1')
net = self.conv2d(net, 128, [3, 3], scope='conv2_2')
net = self.max_pool2d(net, pool_size=[2, 2], stride=[2, 2], scope='pool2')
# b3
net = self.conv2d(net, filter=256, k_size=[3, 3], scope='conv3_1')
net = self.conv2d(net, 256, [3, 3], scope='conv3_2')
net = self.conv2d(net, 256, [3, 3], scope='conv3_3')
net = self.max_pool2d(net, pool_size=[2, 2], stride=[2, 2], scope='pool3')
# b4 =>第1个检测层
net = self.conv2d(net, filter=512, k_size=[3, 3], scope='conv4_1')
net = self.conv2d(net, 512, [3, 3], scope='conv4_2')
net = self.conv2d(net, 512, [3, 3], scope='conv4_3')
check_points['block4'] = net
net = self.max_pool2d(net, pool_size=[2, 2], stride=[2, 2], scope='pool4')
# b5 关键部分来了，这里与vgg不同了
net = self.conv2d(net, filter=512, k_size=[3, 3], scope='conv5_1')
net = self.conv2d(net, 512, [3, 3], scope='conv5_2')
net = self.conv2d(net, 512, [3, 3], scope='conv5_3')
net = self.max_pool2d(net, pool_size=[3, 3], stride=[1, 1], scope='pool5') # =>池化层3*3核， 步长变成1*1
# ===>卷积层，代替VGG全连接层<===
# b6 conv6: 3x3x1024-d6
net = self.conv2d(net, filter=1024, k_size=[3, 3], dilation=[6, 6], scope='conv6')
# => 个数1024, dilation=[6, 6]
# b7 conv7: 1x1x1024 =>第2个检测层
net = self.conv2d(net, filter=1024, k_size=[3, 3], scope='conv7')
# => 个数1024, 卷积核不是[1, 1] ?=?
check_points['block7'] = net
# b8 conv8_1: 1x1x256; conv8_2: 3x3x512-s2-vaild =>第3个检测层
net = self.conv2d(net, 256, [1, 1], scope='conv8_1x1') # =>个数256，卷积核1x1
net = self.conv2d(self.pad2d(net, 1), 512, [3, 3], [2, 2], scope='conv8_3x3', padding='valid')
# =>个数512, 卷积核3x3, 步长2, 'valid'
check_points['block8'] = net
# b9 conv9_1: 1x1x128 conv8_2: 3x3x256-s2-vaild =>第4个检测层
net = self.conv2d(net, 128, [1, 1], scope='conv9_1x1') # =>个数128,卷积核1x1
net = self.conv2d(self.pad2d(net, 1), 256, [3, 3], [2, 2], scope='conv9_3x3', padding='valid')
# =>个数256,卷积核3x3，步长2x2, valid
check_points['block9'] = net
# b10 conv10_1: 1x1x128 conv10_2: 3x3x256-s1-valid =>第5个检测层
net = self.conv2d(net, 128, [1, 1], scope='conv10_1x1') # =>个数128,卷积核1x1
net = self.conv2d(net, 256, [3, 3], scope='conv10_3x3', padding='valid')
# =>个数256，valid
check_points['block10'] = net
# b11 conv11_1: 1x1x128 conv11_2: 3x3x256-s1-valid =>第6检测层
net = self.conv2d(net, 128, [1, 1], scope='conv11_1x1') # =>个数128,卷积核1x1
net = self.conv2d(net, 256, [3, 3], scope='conv11_3x3', padding='valid')
# =>个数256， valid
check_points['block11'] = net
print(check_points)
if __name__ == '__main__':
sd = ssd()
sd.set_net()
1 完善网络
需要用到别人训练好的权重文件，已经下载好了
1 添加构造函数
一共8732个锚点框，我算的是8722
block4因为太大需要L2norm
2 最后加了两个卷积，来实现分类和锚点框定位
完善后的代码
#!usr/bin/python
# -*- coding: utf-8 -*-
# Creation Date: 2019/7/10
import tensorflow as tf
import numpy as np
import cv2
class ssd(object):
def __init__(self):
# ===>完善：构造函数的参数<===
self.num_boxes = [] # 统计锚点框的个数
self.feaeture_map_size = [(38, 38), (19, 19), (10, 10), (5, 5), (3, 3), (1, 1)] # 特征图的大小
self.classes = ["aeroplane", "bicycle", "bird", "boat", "bottle",
"bus", "car", "cat", "chair", "cow",
"diningtable", "dog", "horse", "motorbike", "person",
"pottedplant", "sheep", "sofa", "train", "tvmonitor"] # 这里有20个，加上背景是21个类别
self.feature_layers = ['block4', 'block7', 'block8', 'block9', 'block10', 'block11'] # 用于检测的特征层的名字
self.img_size = (300, 300) # 图片的大小
self.num_classes = 21 # 类别的个数,背景也算一类, 第一个类似是:'bg'背景
self.boxes_len = [4, 6, 6, 6, 4, 4] # 6个特征图生成的一组锚点框的框的个数，4 10 11层是4个一组其它的是6个一组
# block4: 38x38大小的特征图就会生成 38x38x4 个锚点框 5766
# block7: 19x19大小的特征图就会生成 19x19x6 个锚点框 2166
# block8: 10x10大小的特征图就会生成 10x10x6 个锚点框 600
# block9: 5x5大小的特征图就会生成 5x5x6 个锚点框 150
# block10: 3x3大小的特征图就会生成 3x3x4 个锚点框 36
# block11: 1x1大小的特征图就会生成 1x1x4 个锚点框 4
# 一共8732个锚点框 ？=？我算出来的是8722个
self.isL2norm = [True, False, False, False, False, False] # block4比较靠前, 因为norm太大需要L2norm
# ?=?如何起作用的呢
# ===>l2正则化<===
def l2norm(self, x, scale, trainable=True, scope='L2Normalization'):
n_channels = x.get_shape().as_list()[-1] # 通道数. 得到形状,变成列表,取后一个
l2_norm = tf.nn.l2_normalize(x, dim=[3], epsilon=1e-12) # 只对每个像素点在channels上做归一化
with tf.variable_scope(scope):
gamma = tf.get_variable("gamma", shape=[n_channels, ], dtype=tf.float32,
initializer=tf.constant_initializer(scale),
trainable=trainable)
return l2_norm * gamma
# ===>下面开始定义所需组件<===
# conv2d, max_pool2d, pad2d, dropout
# 定义一个卷积的操作 1输入 2卷积核个数 3卷积核大小 4步长 5padding 6膨胀 7激活函数 8名字
def conv2d(self, x, filter, # 输入x, 卷积核的个数filter
k_size, stride=[1, 1], # k_size卷积核是几*几，步长stride
padding='same', dilation=[1, 1], # padding, 空洞卷积指数这里1代表正常卷积
activation=tf.nn.relu, scope='conv2d'): # 激活函数relu, 名字scope
return tf.layers.conv2d(input=x, filters=filter, kernel_size=k_size,
strides=stride, padding=padding, dilation_rate=dilation,
name=scope, activation=activation)
def max_pool2d(self, x, pool_size, stride, scope='max_pool2d'): # 我猜padding是vaild
return tf.layers.max_pooling2d(inputs=x, pool_size=pool_size, strides=stride, padding='valid', name=scope)
# 用于填充s=2的第8,9层. 从6层往后的卷积层需要自己填充, 不要用它自带的填充.
def pad2d(self, x, pad):
return tf.pad(x, paddings=[[0, 0], [pad, pad], [pad, pad], [0, 0]])
def dropout(self, x, d_rate=0.5):
return tf.layers.dropout(inputs=x, rate=d_rate)
def ssd_prediction(self, x, num_classes, box_len, isL2norm, scope='multibox'):
reshape = [-1] + x.getshape().as_list()[1:-1] # 去除1 4维度,拿到2 3维度,变成列表.
# block8为例：shape = (?, 10, 10, 512)需要把23维度拿出来
# 前面的-1表示batch, 因为不知道是多少在这里tf一般写-1
with tf.variable_scope(scope): # 开始进行卷积
if isL2norm:
x = self.l2norm(x) # 先判断是否需要归一化
# ==>预测位置:坐标和大小，回归问题:不需softmax
location_pred = self.conv2d(x, filter=box_len * 4, k_size=[3 * 3], activation=None, scope='conv_loc')
'''filter:卷积核的个数=一个锚点多少框 x 一个框对应的4个数据xywh, 卷积核3x3,不需要激活函数，默认def conv2d有激活函数的'''
location_pred = tf.reshape(location_pred, reshape + [box_len, 4]) # 每个中心点生成一个锚点框?=?
# ==>预测类别:分类问题:需要softmax
class_pred = self.conv2d(x, filter=box_len * num_classes, k_size=3 * 3, activation=None,
scope='conv_cls')
'''filter:卷积核的个数=一个锚点多少框 x 一个框对应的21个类别, 卷积核3x3,不需要激活函数，默认def conv2d是有激活函数的'''
class_pred = tf.shape(class_pred, reshape + [box_len, num_classes]) # ?=?
print(location_pred, class_pred)
return location_pred, class_pred
# ===>下面开始写网络架构<===
def set_net(self):
check_points = {} # 装特征层的字典,用于循环迭代
predictions = []
locations = []
x = tf.placeholder(dtype=tf.float32, shape=[None, 300, 300, 3])
with tf.variable_scope('ssd_300_vgg'):
# ===>VGG前5层<===
# b1
net = self.conv2d(x, filter=64, k_size=[3, 3], scope='conv1_1') # 64个3*3卷积核, s=1 默认，标准卷积
net = self.conv2d(net, 64, [3, 3], scope='conv1_2') # 64个3*3卷积核, s=1默认
net = self.max_pool2d(net, pool_size=[2, 2], stride=[2, 2]) # 池化层2*2卷积核, s=2 默认，池化层一般都是2
# b2
net = self.conv2d(net, filter=128, k_size=[3, 3], scope='conv2_1')
net = self.conv2d(net, 128, [3, 3], scope='conv2_2')
net = self.max_pool2d(net, pool_size=[2, 2], stride=[2, 2], scope='pool2')
# b3
net = self.conv2d(net, filter=256, k_size=[3, 3], scope='conv3_1')
net = self.conv2d(net, 256, [3, 3], scope='conv3_2')
net = self.conv2d(net, 256, [3, 3], scope='conv3_3')
net = self.max_pool2d(net, pool_size=[2, 2], stride=[2, 2], scope='pool3')
# b4 =>第1个检测层
net = self.conv2d(net, filter=512, k_size=[3, 3], scope='conv4_1')
net = self.conv2d(net, 512, [3, 3], scope='conv4_2')
net = self.conv2d(net, 512, [3, 3], scope='conv4_3')
check_points['block4'] = net
net = self.max_pool2d(net, pool_size=[2, 2], stride=[2, 2], scope='pool4')
# b5 关键部分来了，这里与vgg不同了
net = self.conv2d(net, filter=512, k_size=[3, 3], scope='conv5_1')
net = self.conv2d(net, 512, [3, 3], scope='conv5_2')
net = self.conv2d(net, 512, [3, 3], scope='conv5_3')
net = self.max_pool2d(net, pool_size=[3, 3], stride=[1, 1], scope='pool5') # =>池化层3*3核， 步长变成1*1
# ===>卷积层，代替VGG全连接层<===
# b6 conv6: 3x3x1024-d6
net = self.conv2d(net, filter=1024, k_size=[3, 3], dilation=[6, 6], scope='conv6')
# => 个数1024, dilation=[6, 6]
# b7 conv7: 1x1x1024 =>第2个检测层
net = self.conv2d(net, filter=1024, k_size=[3, 3], scope='conv7')
# => 个数1024, 卷积核不是[1, 1] ?=?
check_points['block7'] = net
# b8 conv8_1: 1x1x256; conv8_2: 3x3x512-s2-vaild =>第3个检测层
net = self.conv2d(net, 256, [1, 1], scope='conv8_1x1') # =>个数256，卷积核1x1
net = self.conv2d(self.pad2d(net, 1), 512, [3, 3], [2, 2], scope='conv8_3x3', padding='valid')
# =>个数512, 卷积核3x3, 步长2, 'valid'
check_points['block8'] = net
# b9 conv9_1: 1x1x128 conv8_2: 3x3x256-s2-vaild =>第4个检测层
net = self.conv2d(net, 128, [1, 1], scope='conv9_1x1') # =>个数128,卷积核1x1
net = self.conv2d(self.pad2d(net, 1), 256, [3, 3], [2, 2], scope='conv9_3x3', padding='valid')
# =>个数256,卷积核3x3，步长2x2, valid
check_points['block9'] = net
# b10 conv10_1: 1x1x128 conv10_2: 3x3x256-s1-valid =>第5个检测层
net = self.conv2d(net, 128, [1, 1], scope='conv10_1x1') # =>个数128,卷积核1x1
net = self.conv2d(net, 256, [3, 3], scope='conv10_3x3', padding='valid')
# =>个数256，valid
check_points['block10'] = net
# b11 conv11_1: 1x1x128 conv11_2: 3x3x256-s1-valid =>第6检测层
net = self.conv2d(net, 128, [1, 1], scope='conv11_1x1') # =>个数128,卷积核1x1
net = self.conv2d(net, 256, [3, 3], scope='conv11_3x3', padding='valid')
# =>个数256， valid
check_points['block11'] = net
for i, j in enumerate(self.feature_layers): # 枚举特征层 i是个数, j是名字 blockx
loc, cls = self.ssd_prediction(x=check_points[j],
num_classes=self.classes,
box_len=self.boxes_len[i],
isL2norm=self.isL2norm[i],
scope=j + '_box'
)
predictions.append(tf.nn.softmax(cls)) # 需要softmax
locations.append(loc) # 不需要
print(check_points) # 检查网络的结构, eg:block8: (?, 10, 10, 512)
print(locations, predictions)
if __name__ == '__main__':
sd = ssd()
sd.set_net()
问题是我的还是运行不了：
报错如下
打印结果分析：
左边的输出结果
（？， 38， 38， 4，4）
38x38的特征图, 4个框, 4个预测值xywh（？， 19， 19， 6，4）
19x19的特征图, 6个框, 4个预测值xywh右边的输出结果
（？， 38， 38， 4，21）
38x38的特征图, 4个框, 21个类别（类别预测个数）（？， 19， 19， 6，21）
19x19的特征图, 6个框, 21个类别（类别预测个数）
扫一扫
1万+
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
