TF项目实战（SSD目标检测）-VOC2007

理论详解：

训练好的模型和代码会公布在网上（含 VOC数据集  vgg16 模型  以及训练好的模型）：

 待续

步骤：

1.代码地址：https://github.com/balancap/SSD-Tensorflow

2.解压ssd_300_vgg.ckpt.zip 到checkpoint文件夹下（另外将vgg16模型放在本路径下）

3.测试一下看看，在notebooks文件夹下创建demo_test.py,其实就是复制ssd_notebook.ipynb中的代码，该py文件是完成对于单张图片的测试。

1 import os 2 import math 3 import random 4  5 import numpy as np 6 import tensorflow as tf 7 import cv2 8  9 slim = tf.contrib.slim10 import matplotlib.pyplot as plt11 import matplotlib.image as mpimg12 import sys13 14 sys.path.append('../')15 from nets import ssd_vgg_300, ssd_common, np_methods16 from preprocessing import ssd_vgg_preprocessing17 from notebooks import visualization18 19 # TensorFlow session: grow memory when needed. TF, DO NOT USE ALL MY GPU MEMORY!!!20 gpu_options = tf.GPUOptions(allow_growth=True)21 config = tf.ConfigProto(log_device_placement=False, gpu_options=gpu_options)22 isess = tf.InteractiveSession(config=config)23 # Input placeholder.24 net_shape = (300, 300)25 data_format = 'NHWC'26 img_input = tf.placeholder(tf.uint8, shape=(None, None, 3))27 # Evaluation pre-processing: resize to SSD net shape.28 image_pre, labels_pre, bboxes_pre, bbox_img = ssd_vgg_preprocessing.preprocess_for_eval(29     img_input, None, None, net_shape, data_format, resize=ssd_vgg_preprocessing.Resize.WARP_RESIZE)30 image_4d = tf.expand_dims(image_pre, 0)31 32 # Define the SSD model.33 reuse = True if 'ssd_net' in locals() else None34 ssd_net = ssd_vgg_300.SSDNet()35 with slim.arg_scope(ssd_net.arg_scope(data_format=data_format)):36     predictions, localisations, _, _ = ssd_net.net(image_4d, is_training=False, reuse=reuse)37 38 # Restore SSD model.39 ckpt_filename = '../checkpoints/ssd_300_vgg.ckpt'40 # ckpt_filename = '../checkpoints/VGG_VOC0712_SSD_300x300_ft_iter_120000.ckpt'41 isess.run(tf.global_variables_initializer())42 saver = tf.train.Saver()43 saver.restore(isess, ckpt_filename)44 45 # SSD default anchor boxes.46 ssd_anchors = ssd_net.anchors(net_shape)47 48 49 # Main image processing routine.50 def process_image(img, select_threshold=0.5, nms_threshold=.45, net_shape=(300, 300)):51     # Run SSD network.52     rimg, rpredictions, rlocalisations, rbbox_img = isess.run([image_4d, predictions, localisations, bbox_img],53                                                               feed_dict={img_input: img})54 55     # Get classes and bboxes from the net outputs.56     rclasses, rscores, rbboxes = np_methods.ssd_bboxes_select(57         rpredictions, rlocalisations, ssd_anchors,58         select_threshold=select_threshold, img_shape=net_shape, num_classes=21, decode=True)59 60     rbboxes = np_methods.bboxes_clip(rbbox_img, rbboxes)61     rclasses, rscores, rbboxes = np_methods.bboxes_sort(rclasses, rscores, rbboxes, top_k=400)62     rclasses, rscores, rbboxes = np_methods.bboxes_nms(rclasses, rscores, rbboxes, nms_threshold=nms_threshold)63     # Resize bboxes to original image shape. Note: useless for Resize.WARP!64     rbboxes = np_methods.bboxes_resize(rbbox_img, rbboxes)65     return rclasses, rscores, rbboxes66 67 68 # Test on some demo image and visualize output.69 # 测试的文件夹70 path = '../demo/'71 image_names = sorted(os.listdir(path))72 # 文件夹中的第几张图，-1代表最后一张73 img = mpimg.imread(path + image_names[-1])74 rclasses, rscores, rbboxes = process_image(img)75 76 # visualization.bboxes_draw_on_img(img, rclasses, rscores, rbboxes, visualization.colors_plasma)77 visualization.plt_bboxes(img, rclasses, rscores, rbboxes)

4. 将自己的数据集或者 VOC2007直接放在工程目录下

5. 修改datasets文件夹中pascalvoc_common.py文件，将训练类修改别成自己的（这里如果自己的类别）  本文以两类为例子

1 VOC_LABELS = { 2     'none': (0, 'Background'), 3     'aeroplane': (1, 'Vehicle'), 4     'bicycle': (2, 'Vehicle'), 5     'bird': (3, 'Animal'), 6     'boat': (4, 'Vehicle'), 7     'bottle': (5, 'Indoor'), 8     'bus': (6, 'Vehicle'), 9     'car': (7, 'Vehicle'),10     'cat': (8, 'Animal'),11     'chair': (9, 'Indoor'),12     'cow': (10, 'Animal'),13     'diningtable': (11, 'Indoor'),14     'dog': (12, 'Animal'),15     'horse': (13, 'Animal'),16     'motorbike': (14, 'Vehicle'),17     'person': (15, 'Person'),18     'pottedplant': (16, 'Indoor'),19     'sheep': (17, 'Animal'),20     'sofa': (18, 'Indoor'),21     'train': (19, 'Vehicle'),22     'tvmonitor': (20, 'Indoor'),23 }24 #自己的数据25 # VOC_LABELS = {
    26 #     'none': (0, 'Background'),27 #     'aeroplane': (1, 'Vehicle'),28 # }

6.  将图像数据转换为tfrecods格式，修改datasets文件夹中的pascalvoc_to_tfrecords.py文件，然后更改文件的83行读取方式为’rb‘，如果你的文件不是.jpg格式，也可以修改图片的类型。

另外这个修改

7.运行tf_convert_data.py文件，但是需要传给它一些参数：  这个文件生成TFrecords文件的代码

但是该文件需要像类似于linux 命令那样传入参数。  pycharm中如何解决呢？？？

假设我们需要执行：python ./tf_convert_data.py   --dataset_name=pascalvoc  --dataset_dir=./VOC2007/  --output_name=voc_2007_train  --output_dir=./tfrecords_怎么办呢？

 

我们可以在  run中的 Edit ... 进入

一、

二、

三、

参数：--dataset_name=pascalvoc --dataset_dir=./VOC2007/ --output_name=voc_2007_train --output_dir=./tfrecords_

然后执行该py文件就ok。

 

如果出现错误（文件夹相关的错误），则在工程下建立一个文件夹就可以了。

 

8.训练模型train_ssd_network.py文件中修改

 

None代表一直训练。

 

 

其它需要修改的文件：

① nets/ssd_vgg_300.py  (因为使用此网络结构) ，修改87 和88行的类别

② train_ssd_network.py,修改类别120行，GPU占用量，学习率，batch_size等

③ eval_ssd_network.py 修改类别，66行

④ datasets/pascalvoc_2007.py 根据自己的训练数据修改整个文件

 

9.开始训练

 类似于第7步中的 三

训练的主文件为  train_ssd_network.py

参数为：

--train_dir=./train_model/ --dataset_dir=./tfrecords_/ --dataset_name=pascalvoc_2007 --dataset_split_name=train --model_name=ssd_300_vgg --checkpoint_path=./checkpoints/vgg_16.ckpt --checkpoint_model_scope=vgg_16 --checkpoint_exclude_scopes=ssd_300_vgg/conv6,ssd_300_vgg/conv7,ssd_300_vgg/block8,ssd_300_vgg/block9,ssd_300_vgg/block10,ssd_300_vgg/block11,ssd_300_vgg/block4_box,ssd_300_vgg/block7_box,ssd_300_vgg/block8_box,ssd_300_vgg/block9_box,ssd_300_vgg/block10_box,ssd_300_vgg/block11_box --trainable_scopes=ssd_300_vgg/conv6,ssd_300_vgg/conv7,ssd_300_vgg/block8,ssd_300_vgg/block9,ssd_300_vgg/block10,ssd_300_vgg/block11,ssd_300_vgg/block4_box,ssd_300_vgg/block7_box,ssd_300_vgg/block8_box,ssd_300_vgg/block9_box,ssd_300_vgg/block10_box,ssd_300_vgg/block11_box --save_summaries_secs=60 --save_interval_secs=600 --weight_decay=0.0005 --optimizer=adam --learning_rate=0.001 --learning_rate_decay_factor=0.94 --batch_size=24 --gpu_memory_fraction=0.9

 

训练过程

：

 

 

10 测试：

先看效果

 

另外我修改了demo_test文件 调取电脑摄像投来执行代码。

 

如果单独看一张图的效果则执行函数：

代码如下：

1 __author__ = "WSX"  2 import os  3 import math  4 import random  5   6 import numpy as np  7 import tensorflow as tf  8 import cv2  9  10 slim = tf.contrib.slim 11 import matplotlib.pyplot as plt 12 import matplotlib.image as mpimg 13 import sys 14  15 sys.path.append('../') 16 from nets import ssd_vgg_300, ssd_common, np_methods 17 from preprocessing import ssd_vgg_preprocessing 18 from notebooks import visualization 19  20 # TensorFlow session: grow memory when needed. TF, DO NOT USE ALL MY GPU MEMORY!!! 21 gpu_options = tf.GPUOptions(allow_growth=True) 22 config = tf.ConfigProto(log_device_placement=False, gpu_options=gpu_options) 23 isess = tf.InteractiveSession(config=config) 24 # Input placeholder. 25 net_shape = (300, 300) 26 data_format = 'NHWC' 27 img_input = tf.placeholder(tf.uint8, shape=(None, None, 3)) 28 # Evaluation pre-processing: resize to SSD net shape. 29 image_pre, labels_pre, bboxes_pre, bbox_img = ssd_vgg_preprocessing.preprocess_for_eval( 30     img_input, None, None, net_shape, data_format, resize=ssd_vgg_preprocessing.Resize.WARP_RESIZE) 31 image_4d = tf.expand_dims(image_pre, 0) 32  33 # Define the SSD model. 34 reuse = True if 'ssd_net' in locals() else None 35 ssd_net = ssd_vgg_300.SSDNet() 36 with slim.arg_scope(ssd_net.arg_scope(data_format=data_format)): 37     predictions, localisations, _, _ = ssd_net.net(image_4d, is_training=False, reuse=reuse) 38  39 # Restore SSD model. 40 ckpt_filename = '../checkpoints/ssd_300_vgg.ckpt' 41 # ckpt_filename = '../checkpoints/VGG_VOC0712_SSD_300x300_ft_iter_120000.ckpt' 42 isess.run(tf.global_variables_initializer()) 43 saver = tf.train.Saver() 44 saver.restore(isess, ckpt_filename) 45  46 # SSD default anchor boxes. 47 ssd_anchors = ssd_net.anchors(net_shape) 48  49  50 # Main image processing routine. 51 def process_image(img, select_threshold=0.5, nms_threshold=.45, net_shape=(300, 300)): 52     # Run SSD network. 53     rimg, rpredictions, rlocalisations, rbbox_img = isess.run([image_4d, predictions, localisations, bbox_img], 54                                                               feed_dict={img_input: img}) 55  56     # Get classes and bboxes from the net outputs. 57     rclasses, rscores, rbboxes = np_methods.ssd_bboxes_select( 58         rpredictions, rlocalisations, ssd_anchors, 59         select_threshold=select_threshold, img_shape=net_shape, num_classes=21, decode=True) 60  61     rbboxes = np_methods.bboxes_clip(rbbox_img, rbboxes) 62     rclasses, rscores, rbboxes = np_methods.bboxes_sort(rclasses, rscores, rbboxes, top_k=400) 63     rclasses, rscores, rbboxes = np_methods.bboxes_nms(rclasses, rscores, rbboxes, nms_threshold=nms_threshold) 64     # Resize bboxes to original image shape. Note: useless for Resize.WARP! 65     rbboxes = np_methods.bboxes_resize(rbbox_img, rbboxes) 66     return rclasses, rscores, rbboxes 67  68  69 #===========================================测试部分=========================================== 70 #----------------------------------单张图片测试--------------------------- 71 # Test on some demo image and visualize output. 72 # 测试的文件夹 73 def demo(): 74     path = '../demo/' 75     image_names = sorted(os.listdir(path)) 76     # 文件夹中的第几张图，-1代表最后一张 77     img = mpimg.imread(path + image_names[-1]) 78     print(img.shape) 79     rclasses, rscores, rbboxes = process_image(img) 80  81     # visualization.bboxes_draw_on_img(img, rclasses, rscores, rbboxes, visualization.colors_plasma) 82     visualization.plt_bboxes(img, rclasses, rscores, rbboxes) 83  84  85 #======================================做成实时显示的代码=================================================== 86 L = ["None","aeroplane","bicycle","bird ","boat ","bottle ","bus ","car ","cat ","chair","cow ","diningtable","dog","horse","motorbike","person", 87      "pottedplant","sheep","sofa","train","tvmonitor"] 88 colors_tableau = [(255, 255, 255), (31, 119, 180), (174, 199, 232), (255, 127, 14), (255, 187, 120), 89                   (44, 160, 44), (152, 223, 138), (214, 39, 40), (255, 152, 150), 90                   (148, 103, 189), (197, 176, 213), (140, 86, 75), (196, 156, 148), 91                   (227, 119, 194), (247, 182, 210), (127, 127, 127), (199, 199, 199), 92                   (188, 189, 34), (219, 219, 141), (23, 190, 207), (158, 218, 229)] 93  94 def Load_video_show():  #获取视频 95     video = cv2.VideoCapture("1.mp4")  # 0 表示摄像头 ，   如果为文件路径则 为加载视频 96     while (True): 97         ret, frame = video.read()  #frame为 帧  这里当做一张图 98         frame = cv2.flip( frame ,1)  #镜像变换，图像正与不正 99         cv2.resizeWindow("video", 640, 360)                #设置窗口大小100         frame = cv2.resize(frame, (640, 360))               #设置图大小101         #cv2.imshow("video" ,frame)102         rclasses, rscores, rbboxes = process_image(frame)103         height = frame.shape[0]104         width = frame.shape[1]105         colors = dict()106         for i in range(rclasses.shape[0]):107             cls_id = int(rclasses[i])108             if cls_id >= 0:109                 score = rscores[i]110                 if cls_id not in colors:111                     colors[cls_id] = (random.random(), random.random(), random.random())112                 ymin = int(rbboxes[i, 0] * height)113                 xmin = int(rbboxes[i, 1] * width)114                 ymax = int(rbboxes[i, 2] * height)115                 xmax = int(rbboxes[i, 3] * width)116                 cv2.rectangle(frame, (xmin, ymin), (xmax,ymax), colors[cls_id], 2)  #画矩形框117                 class_name = L[cls_id]118                 cv2.putText(frame, '{:s} | {:.3f}'.format(class_name, score), (xmin, ymin - 2,), cv2.FONT_HERSHEY_COMPLEX, 0.5, (0, 255, 0), 1)  #写文字119         cv2.imshow("video", frame)120         c = cv2.waitKey(50)121         if c == 27: #esc退出122             break123 124 125 #Load_video_show()126 demo()