用tensorflow做深度學習的目標檢測真是艱難困苦啊！

1.程式碼地址：https://github.com/balancap/SSD-Tensorflow，下載該程式碼到本地

2.解壓ssd_300_vgg.ckpt.zip 到checkpoint資料夾下

3.測試一下看看，在notebooks中建立demo_test.py,其實就是複製ssd_notebook.ipynb中的程式碼，該py檔案是完成對於單張圖片的測試，對Jupyter不熟，就自己改了，感覺這樣要方便一些。

1. import os

2. import math

3. import random

4. import numpy as np

5. import tensorflow as tf

6. import cv2

7. slim = tf.contrib.slim

8. import matplotlib.pyplot as plt

9. import matplotlib.image as mpimg

10. import sys

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

12. from nets import ssd_vgg_300, ssd_common, np_methods

13. from preprocessing import ssd_vgg_preprocessing

14. from notebooks import visualization

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

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

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

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

19. # Input placeholder.

20. net_shape = (300, 300)

21. data_format = 'NHWC'

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

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

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

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

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

27. # Define the SSD model.

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

29. ssd_net = ssd_vgg_300.SSDNet()

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

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

32. # Restore SSD model.

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

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

35. isess.run(tf.global_variables_initializer())

36. saver = tf.train.Saver()

37. saver.restore(isess, ckpt_filename)

38. # SSD default anchor boxes.

39. ssd_anchors = ssd_net.anchors(net_shape)

40. # Main image processing routine.

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

42. # Run SSD network.

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

44. feed_dict={img_input: img})

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

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

47. rpredictions, rlocalisations, ssd_anchors,

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

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

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

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

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

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

54. return rclasses, rscores, rbboxes

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

56. #測試的資料夾

57. path = '../demo/'

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

59. #資料夾中的第幾張圖，-1代表最後一張

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

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

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

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

4.將自己的資料集做成VOC2007格式放在該工程下面

5. 修改datasets資料夾中pascalvoc_common.py檔案，將訓練類修改別成自己的

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. #修改後的

26. VOC_LABELS = {

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

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

29. }

6.  將影象資料轉換為tfrecods格式，修改datasets資料夾中的pascalvoc_to_tfrecords.py檔案，然後更改檔案的83行讀取方式為’rb‘，如果你的檔案不是.jpg格式，也可以修改圖片的型別。

此外， 修改67行，可以修改幾張圖片轉為一個tfrecords

7.執行tf_convert_data.py檔案，但是需要傳給它一些引數：

生成tfrecords檔案過程中你會看到	生成tfrecords檔案完畢後你會看到

8.訓練模型train_ssd_network.py檔案中修改

train_ssd_network.py檔案中網路引數配置，若需要改，在此檔案中進行修改，如：

其他需要修改的地方

a.   nets/ssd_vgg_300.py  (因為使用此網路結構) ，修改87 和88行的類別
b. train_ssd_network.py,修改類別120行，GPU佔用量，學習率，batch_size等
c eval_ssd_network.py 修改類別，66行
d. datasets/pascalvoc_2007.py 根據自己的訓練資料修改整個檔案	# (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.通過載入預訓練好的vgg16模型,訓練網路

下載預訓練好的vgg16模型，解壓放入checkpoint檔案中，如果找不到vgg_16.ckpt檔案，可以在下面的連結中點選下載。

連結：https://pan.baidu.com/s/1diWbdJdjVbB3AWN99406nA 密碼：ge3x

按照之前的方式，同樣，如果你是linux使用者，你可以新建一個.sh檔案，檔案裡寫入

1. DATASET_DIR=./tfrecords_/

2. TRAIN_DIR=./train_model/

3. CHECKPOINT_PATH=./checkpoints/vgg_16.ckpt

4. python3 ./train_ssd_network.py \

5. --train_dir=./train_model/ \ #訓練生成模型的存放路徑

6. --dataset_dir=./tfrecords_/ \ #資料存放路徑

7. --dataset_name=pascalvoc_2007 \ #資料名的字首

8. --dataset_split_name=train \

9. --model_name=ssd_300_vgg \ #載入的模型的名字

10. --checkpoint_path=./checkpoints/vgg_16.ckpt \ #所載入模型的路徑

11. --checkpoint_model_scope=vgg_16 \ #所載入模型裡面的作用域名

12. --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 \

13. --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 \

14. --save_summaries_secs=60 \ #每60s儲存一下日誌

15. --save_interval_secs=600 \ #每600s儲存一下模型

16. --weight_decay=0.0005 \ #正則化的權值衰減的係數

17. --optimizer=adam \ #選取的最優化函式

18. --learning_rate=0.001 \ #學習率

19. --learning_rate_decay_factor=0.94 \ #學習率的衰減因子

20. --batch_size=24 \

21. --gpu_memory_fraction=0.9 #指定佔用gpu記憶體的百分比

如果你是windows+pycharm中執行，除了在上述的run中Edit Configuration配置，你還可以開啟Terminal，在這裡執行程式碼，輸入即可

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