A级课程从零开始（七）ssd24训练部分收尾补充（IOU,NMS）

需要学的python命令:

1 np.argsort()

2 np.transpose()

3 np.ones(scores.shape, dtype&＃61;np.bool)

>>> a &＃61; np.array([[1,2],[3,4]])
>>> a
array([[1, 2],[3, 4]])
>>> b &＃61; np.ones(a.shape, dtype&＃61;np.bool)
>>> b
array([[ True, True],[ True, True]], dtype&＃61;bool)
>>> a.size
4

4 np.logical_or(overlap

5 keep_bboxes[(i&＃43;1):] &＃61; np.logical_and(keep_bboxes[(i&＃43;1):], keep_overlap)

>>> np.logical_and(True, False)
False
>>> np.logical_and([True, False], [False, False])
array([False, False], dtype&＃61;bool)

>>> x &＃61; np.arange(5)
>>> x
array([0, 1, 2, 3, 4])
>>> np.logical_and(x>1, x<4)
array([False, False, True, True, False], dtype&＃61;bool)

6 idxes &＃61; np.where(keep_bboxes)

 

当前最终版&＃xff1a;

#!usr/bin/python
# -*- coding: utf-8 -*-
# Creation Date: 2019/7/10
import tensorflow as tf
import numpy as np
import cv2
&＃39;&＃39;&＃39; 注释说明
# 1 /&＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61;>xxxx<&＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61;\　　为一级标题-start|&＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61;>xxxx 为一级标题的补充说明\&＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61;>xxxx<&＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61;/　　为一级标题-end# 2 /&＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61;>xxxx<&＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61;\　　为二级标题-start|&＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61;>xxxx 为二级标题的补充说明\&＃61;&＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61;>xxxx<&＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61;/　　为二级标题-end# 3 /&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;>xxxx<&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;\　　为三级标题-start|&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;>xxxx 为三级标题的补充说明\&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;>xxxx<&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;/　　 为三级标题-end# 4 /&＃61;&＃61;&＃61;>xxxx<&＃61;&＃61;&＃61;\　　为四级标题 &＃61;&＃61;&＃61;>xxxx<&＃61;&＃61;&＃61; 为四级标题的简化形式|&＃61;&＃61;&＃61;> xxx 为四级标题的补充说明\&＃61;&＃61;&＃61;>xxxx<&＃61;&＃61;&＃61;/ 为四级标题-end# 5 &＃61;&＃61;>　or &＃61;>　　为重点-特殊情况# 6 ?&＃61;?　　存在疑惑&＃39;&＃39;&＃39;class ssd(object):def __init__(self):# &＃61;&＃61;&＃61;>完善&＃xff1a;构造函数的参数<&＃61;&＃61;&＃61;self.num_boxes &＃61; [] # 统计锚点框的个数self.feaeture_map_size &＃61; [(38, 38), (19, 19), (10, 10), (5, 5), (3, 3), (1, 1)] # 特征图的大小self.classes &＃61; ["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&＃xff0c;加上背景是21个类别self.feature_layers &＃61; [&＃39;block4&＃39;, &＃39;block7&＃39;, &＃39;block8&＃39;, &＃39;block9&＃39;, &＃39;block10&＃39;, &＃39;block11&＃39;] # 用于检测的特征层的名字self.img_size &＃61; (300, 300) # 图片的大小self.num_classes &＃61; 21 # 类别的个数,背景也算一类, 第一个类似是:&＃39;bg&＃39;背景self.boxes_len &＃61; [4, 6, 6, 6, 4, 4] # 6个特征图生成的一组锚点框的框的个数&＃xff0c;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 &＃xff1f;&＃61;&＃xff1f;我算出来的是8722个self.isL2norm &＃61; [True, False, False, False, False, False] # block4比较靠前, 因为norm太大需要L2normself.anchor_size &＃61; [(21., 45.), (45., 99.), (99., 153.), (153., 207.), (207., 261.), (261., 315.)]self.anchor_ratios &＃61; [[2, .5], [2, .5, 3, 1./3], [2, .5, 3, 1./3], [2, .5, 3, 1./3], [2, .5], [2, .5]]self.anchor_steps &＃61; [8, 16, 32, 64, 100, 300]self.prior_scaling &＃61; [0.1, 0.1, 0.2, 0.2] # 特征先验框缩放比例: 0.1:xy坐标的缩放比, 0.2:wy坐标的缩放比self.n_boxes &＃61; [5776, 2166, 600, 150, 36, 4]# 一共8732个 # ?&＃61;?怎么计算的, 每个featuremap(特征图)先验框的个数# 4层: 38x38x4# 7层: 19x19x6# 8层: 10x10x6# 9层: 5x5x6# 10层: 3x3x4# 11层: 1x1x4self.threshold &＃61; 0.2 # 原文中是0.5, 为了检测到更多的物体设置为0.2# /&＃61;&＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61;> ssd网络架构部分 <&＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61;&＃61;\# &＃61;&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;>l2正则化<&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;&＃61;def l2norm(self, x, scale, trainable&＃61;True, scope&＃61;&＃39;L2Normalization&＃39;):n_channels &＃61; x.get_shape().as_list()[-1] # 通道数. 得到形状,变成列表,取后一个l2_norm &＃61; tf.nn.l2_normalize(x, dim&＃61;[3], epsilon&＃61;1e-12) # 只对每个像素点在channels上做归一化with tf.variable_scope(scope):gamma &＃61; tf.get_variable("gamma", shape&＃61;[n_channels, ], dtype&＃61;tf.float32,initializer&＃61;tf.constant_initializer(scale), # ?&＃61;?为何作者没有这步骤trainable&＃61;trainable)return l2_norm * gamma# /&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;>下面:定义cnn所需组件<&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;&＃61;\# |&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;conv2d, max_pool2d, pad2d, dropout# |&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;tf.layers.conv2d(inputs&＃61;xxx, filters&＃61;xxx, kernel_size&＃61;xxx,# stride&＃61;xxx, padding&＃61;xxx, dilation_rate&＃61;xxx,# name&＃61;xxx, activation&＃61;xxx)# |&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;tf.layers.max_pooling2d(inputs&＃61;xxx, pool_size&＃61;xxx,# strides&＃61;xxx, padding&＃61;xxx,# name&＃61;xxx)# |&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;tf.pad(x, paddings&＃61;xxx)# |&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;tf.layers.dropout(inputs&＃61;xxx, rate&＃61;xxx)# |&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;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, k_size, # 输入x, 卷积核的个数filter, k_size卷积核是几*几stride&＃61;[1, 1], padding&＃61;&＃39;same&＃39;, dilation&＃61;[1, 1], # # 步长stride, padding, 空洞卷积指数这里1代表正常卷积activation&＃61;tf.nn.relu, scope&＃61;&＃39;conv2d&＃39;): # 激活函数relu, 名字scopereturn tf.layers.conv2d(inputs&＃61;x, filters&＃61;filter, kernel_size&＃61;k_size,strides&＃61;stride, dilation_rate&＃61;dilation, padding&＃61;padding,name&＃61;scope, activation&＃61;activation)def max_pool2d(self, x, pool_size,stride,scope&＃61;&＃39;max_pool2d&＃39;): # 我猜padding是vaildreturn tf.layers.max_pooling2d(inputs&＃61;x, pool_size&＃61;pool_size, strides&＃61;stride, padding&＃61;&＃39;valid&＃39;, name&＃61;scope)# 用于填充s&＃61;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&＃61;[[0, 0], [pad, pad], [pad, pad], [0, 0]])def dropout(self, x, d_rate&＃61;0.5):return tf.layers.dropout(inputs&＃61;x, rate&＃61;d_rate)def ssd_prediction(self, x, num_classes, box_num, isL2norm, scope&＃61;&＃39;multibox&＃39;):reshape &＃61; [-1] &＃43; x.getshape().as_list()[1:-1] # 去除第1,4数,拿到2,3个数,变成列表.即去除第一个和最后一个得到shape# python中 a &＃61; [1,2,3,4], b &＃61; a[1:3]&＃61;a[1:-1], c &＃61; [-1] &＃43; b# print(b) &＃61; [2,3], print(c) &＃61; [-1, 1, 2, 3, 4]# block8为例&＃xff1a;shape &＃61; (?, 10, 10, 512)需要把第2,3个数拿出来# 前面的-1表示batch, 因为不知道是多少在这里tf一般写-1# reshape &＃61; [-1, 10, 10]with tf.variable_scope(scope): # 开始进行卷积if isL2norm:x &＃61; self.l2norm(x) # 先判断是否需要归一化# &＃61;&＃61;>预测位置:坐标和大小&＃xff0c;回归问题:不需softmaxlocation_pred &＃61; self.conv2d(x, filter&＃61;box_num * 4, k_size&＃61;[3 * 3],activation&＃61;None, scope&＃61;&＃39;conv_loc&＃39;)&＃39;&＃39;&＃39;filter:卷积核的个数&＃61;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&＃39;&＃39;&＃39;location_pred &＃61; tf.reshape(location_pred, reshape &＃43; [box_num, 4]) # 每个中心点生成一个锚点框?&＃61;?# reshape &＃43; [box_num, 4] &＃61; [-1, 10, 10, box_num, 4]# &＃61;&＃61;>预测类别:分类问题:需要softmaxclass_pred &＃61; self.conv2d(x, filter&＃61;box_num * num_classes, k_size&＃61;3 * 3,activation&＃61;None, scope&＃61;&＃39;conv_cls&＃39;)&＃39;&＃39;&＃39;filter:卷积核的个数&＃61;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&＃39;&＃39;&＃39;class_pred &＃61; tf.shape(class_pred, reshape &＃43; [box_num, num_classes]) # ?&＃61;?# reshape &＃43; [box_num, num_classes] &＃61; [-1, 10, 10, box_num, num_classes]print(location_pred, class_pred)return location_pred, class_pred# \&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;>上面:定义cnn所需组件<&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;&＃61;/# /&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;>下面:具体网络架构-start<&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;\def set_net(self):check_points &＃61; {} # 装特征层的字典,用于循环迭代predictions &＃61; []locations &＃61; []x &＃61; tf.placeholder(dtype&＃61;tf.float32, shape&＃61;[None, 300, 300, 3])with tf.variable_scope(&＃39;ssd_300_vgg&＃39;):# &＃61;&＃61;&＃61;>VGG前5层<&＃61;&＃61;&＃61;# b1net &＃61; self.conv2d(x, filter&＃61;64, k_size&＃61;[3, 3], scope&＃61;&＃39;conv1_1&＃39;) # 64个3*3卷积核, s&＃61;1 默认&＃xff0c;标准卷积net &＃61; self.conv2d(net, 64, [3, 3], scope&＃61;&＃39;conv1_2&＃39;) # 64个3*3卷积核, s&＃61;1默认net &＃61; self.max_pool2d(net, pool_size&＃61;[2, 2], stride&＃61;[2, 2], scope&＃61;&＃39;pool1&＃39;) # 池化层2*2卷积核, s&＃61;2 默认&＃xff0c;池化层一般都是2# b2net &＃61; self.conv2d(net, filter&＃61;128, k_size&＃61;[3, 3], scope&＃61;&＃39;conv2_1&＃39;)net &＃61; self.conv2d(net, 128, [3, 3], scope&＃61;&＃39;conv2_2&＃39;)net &＃61; self.max_pool2d(net, pool_size&＃61;[2, 2], stride&＃61;[2, 2], scope&＃61;&＃39;pool2&＃39;)# b3net &＃61; self.conv2d(net, filter&＃61;256, k_size&＃61;[3, 3], scope&＃61;&＃39;conv3_1&＃39;)net &＃61; self.conv2d(net, 256, [3, 3], scope&＃61;&＃39;conv3_2&＃39;)net &＃61; self.conv2d(net, 256, [3, 3], scope&＃61;&＃39;conv3_3&＃39;)net &＃61; self.max_pool2d(net, pool_size&＃61;[2, 2], stride&＃61;[2, 2], scope&＃61;&＃39;pool3&＃39;)# b4 &＃61;>第1个检测层net &＃61; self.conv2d(net, filter&＃61;512, k_size&＃61;[3, 3], scope&＃61;&＃39;conv4_1&＃39;)net &＃61; self.conv2d(net, 512, [3, 3], scope&＃61;&＃39;conv4_2&＃39;)net &＃61; self.conv2d(net, 512, [3, 3], scope&＃61;&＃39;conv4_3&＃39;)check_points[&＃39;block4&＃39;] &＃61; netnet &＃61; self.max_pool2d(net, pool_size&＃61;[2, 2], stride&＃61;[2, 2], scope&＃61;&＃39;pool4&＃39;)# b5 关键部分来了&＃xff0c;这里与vgg不同了net &＃61; self.conv2d(net, filter&＃61;512, k_size&＃61;[3, 3], scope&＃61;&＃39;conv5_1&＃39;)net &＃61; self.conv2d(net, 512, [3, 3], scope&＃61;&＃39;conv5_2&＃39;)net &＃61; self.conv2d(net, 512, [3, 3], scope&＃61;&＃39;conv5_3&＃39;)net &＃61; self.max_pool2d(net, pool_size&＃61;[3, 3], stride&＃61;[1, 1], scope&＃61;&＃39;pool5&＃39;) # &＃61;>池化层3*3核&＃xff0c; 步长变成1*1# &＃61;&＃61;&＃61;>卷积层&＃xff0c;代替VGG全连接层<&＃61;&＃61;&＃61;# b6 conv6: 3x3x1024-d6net &＃61; self.conv2d(net, filter&＃61;1024, k_size&＃61;[3, 3], dilation&＃61;[6, 6], scope&＃61;&＃39;conv6&＃39;)# &＃61;> 个数1024, dilation&＃61;[6, 6]# b7 conv7: 1x1x1024 &＃61;>第2个检测层net &＃61; self.conv2d(net, filter&＃61;1024, k_size&＃61;[1, 1], scope&＃61;&＃39;conv7&＃39;)# &＃61;> 个数1024, 卷积核是[1, 1]check_points[&＃39;block7&＃39;] &＃61; net# b8 conv8_1: 1x1x256; conv8_2: 3x3x512-s2-vaild &＃61;>第3个检测层net &＃61; self.conv2d(net, 256, [1, 1], scope&＃61;&＃39;conv8_1x1&＃39;) # &＃61;>个数256&＃xff0c;卷积核1x1net &＃61; self.conv2d(self.pad2d(net, 1), 512, [3, 3], [2, 2], scope&＃61;&＃39;conv8_3x3&＃39;, padding&＃61;&＃39;valid&＃39;)# &＃61;>个数512, 卷积核3x3, 步长2, &＃39;valid&＃39;check_points[&＃39;block8&＃39;] &＃61; net# b9 conv9_1: 1x1x128 conv8_2: 3x3x256-s2-vaild &＃61;>第4个检测层net &＃61; self.conv2d(net, 128, [1, 1], scope&＃61;&＃39;conv9_1x1&＃39;) # &＃61;>个数128,卷积核1x1net &＃61; self.conv2d(self.pad2d(net, 1), 256, [3, 3], [2, 2], scope&＃61;&＃39;conv9_3x3&＃39;, padding&＃61;&＃39;valid&＃39;)# &＃61;>个数256,卷积核3x3&＃xff0c;步长2x2, validcheck_points[&＃39;block9&＃39;] &＃61; net# b10 conv10_1: 1x1x128 conv10_2: 3x3x256-s1-valid &＃61;>第5个检测层net &＃61; self.conv2d(net, 128, [1, 1], scope&＃61;&＃39;conv10_1x1&＃39;) # &＃61;>个数128,卷积核1x1net &＃61; self.conv2d(net, 256, [3, 3], scope&＃61;&＃39;conv10_3x3&＃39;, padding&＃61;&＃39;valid&＃39;)# &＃61;>个数256&＃xff0c;validcheck_points[&＃39;block10&＃39;] &＃61; net# b11 conv11_1: 1x1x128 conv11_2: 3x3x256-s1-valid &＃61;>第6检测层net &＃61; self.conv2d(net, 128, [1, 1], scope&＃61;&＃39;conv11_1x1&＃39;) # &＃61;>个数128,卷积核1x1net &＃61; self.conv2d(net, 256, [3, 3], scope&＃61;&＃39;conv11_3x3&＃39;, padding&＃61;&＃39;valid&＃39;)# &＃61;>个数256&＃xff0c; validcheck_points[&＃39;block11&＃39;] &＃61; netfor i, j in enumerate(self.feature_layers): # 枚举特征层i表示第几个, j是名字如&＃39;block4&＃39;loc, cls &＃61; self.ssd_prediction(x&＃61;check_points[j],num_classes&＃61;self.num_classes,box_num&＃61;self.boxes_len[i],isL2norm&＃61;self.isL2norm[i],scope&＃61;j &＃43; &＃39;_box&＃39;)predictions.append(tf.nn.softmax(cls)) # 需要softmaxlocations.append(loc) # 不需要print(check_points) # 检查网络的结构, eg:block8: (?, 10, 10, 512)print(locations, predictions)return locations, predictions, x# locations是5d张量,最后一个维度是4,里面装着预测出来的 0x 1y 2h 3w# locations是列表, 里面的元素形如&＃xff1a;[-1, 10, 10, box_num, 4]# \&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;>上面:具体网络架构<&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;/
# \&＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61;　&＃61;&＃61;&＃61;> ssd网络架构部分-end <&＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61;/# /&＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61;　&＃61;&＃61;&＃61;> 先验框生成*解码*先验框筛选-start <&＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61;\# /&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;> 1&＃xff1a;先验框生成-start <&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;\# |&＃61;&＃61;&＃61; 以block8为例10x10, 生成先验框# |&＃61;&＃61;&＃61; h0, h1, h2, h3, h4, h5:# |&＃61;&＃61;&＃61; 小正方形 大正方形 1/2长方形 2/1长方形 1/3长方形 3/1长方形def ssd_anchor_layer(self, img_size, feature_map_size,anchor_size, anchor_ratio, anchor_step,box_num, offset&＃61;0.5):# 提取feature map 的每一个坐标y, x &＃61; np.mgrid[0: feature_map_size[0], 0:feature_map_size[1]] # 以block8为例这里是 0:10,0:10# >>> y, x&＃61; np.mgrid[0:10, 0:10]# >>> print(x)# [[0 1 2 3 4 5 6 7 8 9]# [0 1 2 3 4 5 6 7 8 9]# [0 1 2 3 4 5 6 7 8 9]# [0 1 2 3 4 5 6 7 8 9]# [0 1 2 3 4 5 6 7 8 9]# [0 1 2 3 4 5 6 7 8 9]# [0 1 2 3 4 5 6 7 8 9]# [0 1 2 3 4 5 6 7 8 9]# [0 1 2 3 4 5 6 7 8 9]# [0 1 2 3 4 5 6 7 8 9]]# >>> print(y)# [[0 0 0 0 0 0 0 0 0 0]# [1 1 1 1 1 1 1 1 1 1]# [2 2 2 2 2 2 2 2 2 2]# [3 3 3 3 3 3 3 3 3 3]# [4 4 4 4 4 4 4 4 4 4]# [5 5 5 5 5 5 5 5 5 5]# [6 6 6 6 6 6 6 6 6 6]# [7 7 7 7 7 7 7 7 7 7]# [8 8 8 8 8 8 8 8 8 8]# [9 9 9 9 9 9 9 9 9 9]]y &＃61; (y.astype(np.float32) &＃43; offset) * anchor_step / img_size[0]x &＃61; (x.astype(np.float32) &＃43; offset) * anchor_step / img_size[1]# 计算两个长宽比为1的 h, wh &＃61; np.zeros((box_num,), np.float32)w &＃61; np.zeros((box_num,), np.float32)# h >>> array([ 0., 0., 0., 0., 0., 0.], dtype&＃61;float32)# w >>> array([ 0., 0., 0., 0., 0., 0.], dtype&＃61;float32)h[0] &＃61; anchor_size[0] / img_size[0] # 小正方形w[0] &＃61; anchor_size[0] / img_size[0]h[1] &＃61; (anchor_size[0] * anchor_size[1]) ** 0.5 / img_size[0] # 大正方形w[1] &＃61; (anchor_size[0] * anchor_size[1]) ** 0.5 / img_size[0]for i, j in enumerate(anchor_ratio):h[i &＃43; 2] &＃61; anchor_size[0] / img_size[0] / (j ** 0.5)w[i &＃43; 2] &＃61; anchor_size[0] / img_size[0] * (j ** 0.5)return y, x, h, w# h[0]&＃61;99/300 w[0]&＃61;99/300&＃xff1a; 小 正方型 h w# h[1]&＃61;sqrt(99*513)/300 w[1]&＃61;sqrt(99*513)/300&＃xff1a; 大 正方型 h w# h[2]&＃61;99/300/sqrt(2) w[2]&＃61;99/300*sqrt(2)&＃xff1a; 横向 - 长方型2/1 对应 anchor_ratio[0] &＃61; 2# h[3]&＃61;99/300/sqrt(0.5) w[2]&＃61;99/300*sqrt(0.5)&＃xff1a; 纵向 | 长方型1/2 对应 anchor_ratio[1] &＃61; .5# h[4]&＃61;99/300/sqrt(3) w[2]&＃61;99/300*sqrt(3)&＃xff1a; 横向 - 长方型3/1 对应 anchor_ratio[2] &＃61; 3# h[5]&＃61;99/300/sqrt(1/3) w[2]&＃61;99/300*sqrt(1/3)&＃xff1a; 纵向 | 长方型1/3 对应 anchor_ratio[3] &＃61; 1./3# h[0]正方型 h[1]正方型 h[2]长方型2/1 h[3]长方型1/2 h[4]长方型3/1 h[5]长方型1/3# h &＃61; array([ 0.33000001, 0.41024384, 0.23334524, 0.46669048, 0.19052559, 0.57157677], dtype&＃61;float32)# w &＃61; array([ 0.33000001, 0.41024384, 0.46669048, 0.23334524, 0.57157677, 0.19052559], dtype&＃61;float32)# \&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;> 1&＃xff1a;先验框生成-end <&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;/# /&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;> 2&＃xff1a;解码-start <&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;\def ssd_decode(self, location, box, prior_scaling):y_a, x_a, h_a, w_a &＃61; boxcx &＃61; location[:, :, :, :, 0] * w_a * prior_scaling[0] &＃43; x_a # ?&＃61;?这部分应该是w hcy &＃61; location[:, :, :, :, 1] * h_a * prior_scaling[1] &＃43; y_a# locations是set_net网络的返回值# locations是5d张量,最后一个维度是4,里面装着预测出来的 0x 1y 2h 3w# locations是列表, 里面的元素形如&＃xff1a;[-1, 10, 10, box_num, 4]# 在这里location被带入了locations[2], 也就是block8层w &＃61; w_a * tf.exp(location[:, :, :, :, 2] * prior_scaling[2])h &＃61; h_a * tf.exp(location[:, :, :, :, 3] * prior_scaling[3]) # 实际格子的高度print(cx, cy, w, h)bboxes &＃61; tf.stack([cy - h/2.0, cx - w/2.0, cy &＃43; h/2.0, cx &＃43; w/2.0], axis&＃61;-1)# 特征图比较多&＃xff0c;需要叠加起来# 左上角点的y坐标 cy-h/2, x坐标cx-w/2. 右下角点的y坐标 cy&＃43;h/2, x坐标cx&＃43;w/2print(bboxes)return bboxes# \&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;> 2&＃xff1a;解码-end <&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;/# /&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;> 3:先验框筛选 <&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;\def choose_anchor_boxes(self, predictions, anchor_box, n_box):# predictions列表里面的元素表示&＃xff1a;类别预测的置信度, shape &＃61; [-1, 10, 10, box_num, num_classes]anchor_box &＃61; tf.reshape(anchor_box, [n_box, 4]) # 5d张量改为2d张量 ?&＃61;? n_box是所有的锚点框的总数量&＃61;批数x10x10x6x4prediction &＃61; tf.reshape(predictions, [n_box, 21])prediction &＃61; prediction[:, 1:] # 第一个0是背景的置信度, 我们不需要,从1往后取classes &＃61; tf.argmax(prediction, axis&＃61;1) &＃43; 1 # 得到preditions的概率最大类别的索引值, 1表示按行找最大&＃43;1: 是因为代码里是0开始scores &＃61; tf.reduce_max(prediction, axis&＃61;1) # 得到最大类别的得分, 当大于阈值就保留下来(下面后话)filter_mask &＃61; scores > self.thresholdclasses &＃61; tf.boolean_mask(classes, filter_mask) # 前面放筛选目标, 后面放筛选条件scores &＃61; tf.boolean_mask(scores, filter_mask)anchor_box &＃61; tf.boolean_mask(anchor_box, filter_mask)return classes, scores, anchor_box# 需要学习的指令&＃xff1a;tf.reshape() tf.reduce_max() tf.boolean_mask()# \&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;> 3:先验框筛选 <&＃61;&＃61;&＃61; &＃61;&＃61;&＃61;/
# \&＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61;　&＃61;&＃61;&＃61;> 先验框生成*解码*先验框筛选-start <&＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61; &＃61;&＃61;&＃61;/if __name__ &＃61;&＃61; &＃39;__main__&＃39;:sd &＃61; ssd()locations, predictions, x &＃61; sd.set_net()box &＃61; sd.ssd_anchor_layer(sd.img_size, (10, 10), (99., 153.), [2., .5, 3., 1/3], 32, 6)boex &＃61; sd.ssd_decode(locations[2], box, sd.prior_scaling)print(boex) # shape &＃61; (?, 10, 10, 6, 4)# 这里以block8为例&＃xff1a;的输出结果为Tensor("stacck:0", shape&＃61;(?, 10, 10, 6, 4), dtype&＃61;float32)# 10, 10表示的是第三个特征层为10x10, 因为是locations[2] ?&＃61;?# 6表示六个特征图 ?&＃61;?# 4表示　左上角&＃xff06;右下角坐标　max_x max_y min_x min_y# locations[0]是38x38 locations[1]是19x19 locations[2]是10x10 locations[3]是5x5 [4]是3x3 [5]是2x2 [6]是1x1cls, sco, a_box &＃61; sd.choose_anchor_boxes(predictions[2], boex, sd.n_boxes[2])print(&＃39;----------------------------&＃39;)print(cls, sco, a_box)