目标检测SSD+Tensorflow 转数据为tfrecord

转载：https://blog.csdn.net/weixin_39881922/article/details/80569803

  用tensorflow做深度学习的目标检测真是艰难困苦啊！

  

1.代码地址：https://github.com/balancap/SSD-Tensorflow，下载该代码到本地

2.解压ssd_300_vgg.ckpt.zip 到checkpoint文件夹下

3.测试一下看看，在notebooks中创建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，对Jupyter不熟，就自己改了，感觉这样要方便一些。

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.slim

10. import matplotlib.pyplot as plt

11. import matplotlib.image as mpimg

12. import sys

13. sys.path.append('../')

14. from nets import ssd_vgg_300, ssd_common, np_methods

15. from preprocessing import ssd_vgg_preprocessing

16. from notebooks import visualization

17. # TensorFlow session: grow memory when needed. TF, DO NOT USE ALL MY GPU MEMORY!!!

18. gpu_options = tf.GPUOptions(allow_growth=True)

19. config = tf.ConfigProto(log_device_placement=False, gpu_options=gpu_options)

20. isess = tf.InteractiveSession(config=config)

21. # Input placeholder.

22. net_shape = (300, 300)

23. data_format = 'NHWC'

24. img_input = tf.placeholder(tf.uint8, shape=(None, None, 3))

25. # Evaluation pre-processing: resize to SSD net shape.

26. image_pre, labels_pre, bboxes_pre, bbox_img = ssd_vgg_preprocessing.preprocess_for_eval(

27. img_input, None, None, net_shape, data_format, resize=ssd_vgg_preprocessing.Resize.WARP_RESIZE)

28. image_4d = tf.expand_dims(image_pre, 0)

29.  

30. # Define the SSD model.

31. reuse = True if 'ssd_net' in locals() else None

32. ssd_net = ssd_vgg_300.SSDNet()

33. with slim.arg_scope(ssd_net.arg_scope(data_format=data_format)):

34. predictions, localisations, _, _ = ssd_net.net(image_4d, is_training=False, reuse=reuse)

35.  

36. # Restore SSD model.

37. ckpt_filename = '../checkpoints/ssd_300_vgg.ckpt'

38. # ckpt_filename = '../checkpoints/VGG_VOC0712_SSD_300x300_ft_iter_120000.ckpt'

39. isess.run(tf.global_variables_initializer())

40. saver = tf.train.Saver()

41. saver.restore(isess, ckpt_filename)

42.  

43. # SSD default anchor boxes.

44. ssd_anchors = ssd_net.anchors(net_shape)

45.  
46.  

47. # Main image processing routine.

48. def process_image(img, select_threshold=0.5, nms_threshold=.45, net_shape=(300, 300)):

49. # Run SSD network.

50. rimg, rpredictions, rlocalisations, rbbox_img = isess.run([image_4d, predictions, localisations, bbox_img],

51. feed_dict={img_input: img})

52.  

53. # Get classes and bboxes from the net outputs.

54. rclasses, rscores, rbboxes = np_methods.ssd_bboxes_select(

55. rpredictions, rlocalisations, ssd_anchors,

56. select_threshold=select_threshold, img_shape=net_shape, num_classes=21, decode=True)

57.  

58. rbboxes = np_methods.bboxes_clip(rbbox_img, rbboxes)

59. rclasses, rscores, rbboxes = np_methods.bboxes_sort(rclasses, rscores, rbboxes, top_k=400)

60. rclasses, rscores, rbboxes = np_methods.bboxes_nms(rclasses, rscores, rbboxes, nms_threshold=nms_threshold)

61. # Resize bboxes to original image shape. Note: useless for Resize.WARP!

62. rbboxes = np_methods.bboxes_resize(rbbox_img, rbboxes)

63. return rclasses, rscores, rbboxes

64. # Test on some demo image and visualize output.

65. #测试的文件夹

66. path = '../demo/'

67. image_names = sorted(os.listdir(path))

68. #文件夹中的第几张图，-1代表最后一张

69. img = mpimg.imread(path + image_names[-1])

70. rclasses, rscores, rbboxes = process_image(img)

71.  

72. # visualization.bboxes_draw_on_img(img, rclasses, rscores, rbboxes, visualization.colors_plasma)

73. visualization.plt_bboxes(img, rclasses, rscores, rbboxes)

 

4.将自己的数据集做成VOC2007格式放在该工程下面

5. 修改datasets文件夹中pascalvoc_common.py文件，将训练类修改别成自己的

1. #原始的

2. # VOC_LABELS = {

3. # 'none': (0, 'Background'),

4. # 'aeroplane': (1, 'Vehicle'),

5. # 'bicycle': (2, 'Vehicle'),

6. # 'bird': (3, 'Animal'),

7. # 'boat': (4, 'Vehicle'),

8. # 'bottle': (5, 'Indoor'),

9. # 'bus': (6, 'Vehicle'),

10. # 'car': (7, 'Vehicle'),

11. # 'cat': (8, 'Animal'),

12. # 'chair': (9, 'Indoor'),

13. # 'cow': (10, 'Animal'),

14. # 'diningtable': (11, 'Indoor'),

15. # 'dog': (12, 'Animal'),

16. # 'horse': (13, 'Animal'),

17. # 'motorbike': (14, 'Vehicle'),

18. # 'person': (15, 'Person'),

19. # 'pottedplant': (16, 'Indoor'),

20. # 'sheep': (17, 'Animal'),

21. # 'sofa': (18, 'Indoor'),

22. # 'train': (19, 'Vehicle'),

23. # 'tvmonitor': (20, 'Indoor'),

24. # }

25. #修改后的

26. VOC_LABELS = {

27. 'none': (0, 'Background'),

28. 'pantograph':(1,'Vehicle'),

29. }

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。

此外， 修改67行，可以修改几张图片转为一个tfrecords

7.运行tf_convert_data.py文件，但是需要传给它一些参数：

linux	在SSD-Tensorflow-master文件夹下创建tf_conver_data.sh,文件写入内容如下： DATASET_DIR=./VOC2007/     #VOC数据保存的文件夹（VOC的目录格式未改变）   OUTPUT_DIR=./tfrecords_  #自己建立的保存tfrecords数据的文件夹        python ./tf_convert_data.py \        --dataset_name=pascalvoc \            --dataset_dir=${DATASET_DIR} \      --output_name=voc_2007_train \    --output_dir=${OUTPUT_DIR}
windows     +pychram	配置pycharm-->run-->Edit Configuration Script parameters中写入：--dataset_name=pascalvoc --dataset_dir=./VOC2007/ --output_name=voc_2007_train --output_dir=./tfrecords_ 运行tf_convert_data.py文件

生成tfrecords文件过程中你会看到	生成tfrecords文件完毕后你会看到

8.训练模型train_ssd_network.py文件中修改

 

train_ssd_network.py文件中网络参数配置，若需要改，在此文件中进行修改，如：

其他需要修改的地方

a.   nets/ssd_vgg_300.py  (因为使用此网络结构) ，修改87 和88行的类别
b. train_ssd_network.py,修改类别120行，GPU占用量，学习率，batch_size等
c eval_ssd_network.py 修改类别，66行
d. datasets/pascalvoc_2007.py 根据自己的训练数据修改整个文件	# (Images, Objects) statistics on every class. # TRAIN_STATISTICS = { # 'none': (0, 0), # 'aeroplane': (238, 306), # 'bicycle': (243, 353), # 'bird': (330, 486), # 'boat': (181, 290), # 'bottle': (244, 505), # 'bus': (186, 229), # 'car': (713, 1250), # 'cat': (337, 376), # 'chair': (445, 798), # 'cow': (141, 259), # 'diningtable': (200, 215), # 'dog': (421, 510), # 'horse': (287, 362), # 'motorbike': (245, 339), # 'person': (2008, 4690), # 'pottedplant': (245, 514), # 'sheep': (96, 257), # 'sofa': (229, 248), # 'train': (261, 297), # 'tvmonitor': (256, 324), # 'total': (5011, 12608), # } # TEST_STATISTICS = { # 'none': (0, 0), # 'aeroplane': (1, 1), # 'bicycle': (1, 1), # 'bird': (1, 1), # 'boat': (1, 1), # 'bottle': (1, 1), # 'bus': (1, 1), # 'car': (1, 1), # 'cat': (1, 1), # 'chair': (1, 1), # 'cow': (1, 1), # 'diningtable': (1, 1), # 'dog': (1, 1), # 'horse': (1, 1), # 'motorbike': (1, 1), # 'person': (1, 1), # 'pottedplant': (1, 1), # 'sheep': (1, 1), # 'sofa': (1, 1), # 'train': (1, 1), # 'tvmonitor': (1, 1), # 'total': (20, 20), # } # SPLITS_TO_SIZES = { # 'train': 5011, # 'test': 4952, # } # (Images, Objects) statistics on every class.   TRAIN_STATISTICS = { 'none': (0, 0), 'pantograph': (1000, 1000), } TEST_STATISTICS = { 'none': (0, 0), 'pantograph': (1000, 1000), } SPLITS_TO_SIZES = { 'train': 500, 'test': 500, }

9.通过加载预训练好的vgg16模型,训练网络

下载预训练好的vgg16模型，解压放入checkpoint文件中，如果找不到vgg_16.ckpt文件，可以在下面的链接中点击下载。

链接：https://pan.baidu.com/s/1diWbdJdjVbB3AWN99406nA 密码：ge3x

按照之前的方式，同样，如果你是linux用户，你可以新建一个.sh文件，文件里写入

1. DATASET_DIR=./tfrecords_/

2. TRAIN_DIR=./train_model/

3. CHECKPOINT_PATH=./checkpoints/vgg_16.ckpt

4.  

5. python3 ./train_ssd_network.py \

6. --train_dir=./train_model/ \ #训练生成模型的存放路径

7. --dataset_dir=./tfrecords_/ \ #数据存放路径

8. --dataset_name=pascalvoc_2007 \ #数据名的前缀

9. --dataset_split_name=train \

10. --model_name=ssd_300_vgg \ #加载的模型的名字

11. --checkpoint_path=./checkpoints/vgg_16.ckpt \ #所加载模型的路径

12. --checkpoint_model_scope=vgg_16 \ #所加载模型里面的作用域名

13. --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 \

14. --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 \

15. --save_summaries_secs=60 \ #每60s保存一下日志

16. --save_interval_secs=600 \ #每600s保存一下模型

17. --weight_decay=0.0005 \ #正则化的权值衰减的系数

18. --optimizer=adam \ #选取的最优化函数

19. --learning_rate=0.001 \ #学习率

20. --learning_rate_decay_factor=0.94 \ #学习率的衰减因子

21. --batch_size=24 \

22. --gpu_memory_fraction=0.9 #指定占用gpu内存的百分比

如果你是windows+pycharm中运行，除了在上述的run中Edit Configuration配置，你还可以打开Terminal，在这里运行代码，输入即可

python ./train_ssd_network.py  --train_dir=./train_model/  --dataset_dir=./tfrecords_/  --dataset_name=pascalvoc_2007 --dataset_split_name=train --model_name=ssd_300_vgg   --checkpoint_path=./checkpoints/vgg16.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