前奏--學生黨記錄，大佬勿噴
一、准備tensorflow的安裝環境
1、Anaconda3(筆者裝的minconda，Miniconda，顧名思義，它只包含最基本的內容——python與conda，以及相關的必須依賴項，對於空間要求嚴格的用戶，Miniconda是一種選擇。就只包含最基本的東西，其他的庫得自己裝；Anaconda則是一個打包的集合，里面預裝好了conda、某個版本的python、眾多packages、科學計算工具等等，就是把很多常用的不常用的庫都給你裝好了）
2、tensorflow1.x(需要1.12版本以上，筆者裝的1.13，對應的cuda和cudnn版本一定要注意匹配，如下圖）
3、激活tensorflow環境，按照如下鏈接進行安裝
https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/installation.md
按照步驟一步步來裝，cocoapi也需要裝好
二、准備數據集
參考下面鏈接
https://www.cnblogs.com/zongfa/p/9663649.html
https://blog.csdn.net/qq_17854471/article/details/89786400
三、轉換成tensorflow能識別的格式
1、統計xml文件有多少類別
import xml.dom.minidom as xmldom import os #voc數據集獲取所有標簽的所有類別數"
annotation_path="/home/chenxin/下載/Annotations" annotation_names=[os.path.join(annotation_path,i) for i in os.listdir(annotation_path)] labels = list() for names in annotation_names: xmlfilepath = names domobj = xmldom.parse(xmlfilepath) # 得到元素對象
elementobj = domobj.documentElement #獲得子標簽
subElementObj = elementobj.getElementsByTagName("object") for s in subElementObj: label=s.getElementsByTagName("name")[0].firstChild.data #print(label)
if label not in labels: labels.append(label) print(labels)
2、xml轉csv格式（只需識別一個類別）
# -*- coding:utf-8 -*- #!/usr/bin/env python
""" @Time: 2020/2/10 22：00 @Author: chenxin @File Name: Only_one_label_xml2csv.py @Software: PyCharm """
import os import glob import pandas as pd import xml.etree.ElementTree as ET ## xml文件的路徑
os.chdir('/home/chenxin/models-master/research/object_detection/images/train1') path = '/home/chenxin/models-master/research/object_detection/images/train1' img_path = '/home/chenxin/models-master/research/object_detection/images/train1'
def xml_to_csv(path): xml_list = [] for xml_file in glob.glob(path + '/*.xml'): # 返回所有匹配的文件路徑列表即對於path目錄下的每一個xml文件
tree = ET.parse(xml_file)#獲得xml對應的解析樹
root = tree.getroot()#獲得根標簽annotation
for member in root.findall('object'):#對於每一個object標簽
value = (root.find('filename').text,#在根標簽下查找filename標簽，並獲得其文本信息
#int(root.find('size')[0].text),#在根標簽下查找size標簽，並獲得size的第0個子標簽（width）的文本信息，並轉化為int
#int(root.find('size')[1].text),#在根標簽下查找size標簽，並獲得size的第1個子標簽（width）的文本信息，並轉化為int
member[0].text, # 獲得object的第0個子標簽(name)的文本信息
float(member[4][0].text),#獲得object的第4個子標簽(bndbox),並獲得bndbox的第0個子標簽（xmin)的文本信息，並轉換為int
float(member[4][1].text),#獲得object的第4個個子標簽(bndbox),並獲得bndbox的第1個子標簽（ymin)的文本信息，並轉換為int
float(member[4][2].text),#獲得object的第4個個子標簽(bndbox),並獲得bndbox的第2個子標簽（xmax)的文本信息，並轉換為int
float(member[4][3].text) #獲得object的第4個個子標簽(bndbox),並獲得bndbox的第3個子標簽（ymax)的文本信息，並轉換為int
) # value = (img_path + '/' + root.find('filename').text,
# int(member[1][0].text),
# int(member[1][1].text),
# int(member[1][2].text),
# int(member[1][3].text),
# member[0].text
# )
xml_list.append(value) # column_name = ['filename', 'width', 'height', 'class', 'xmin', 'ymin', 'xmax', 'ymax']
column_name = ['filename', 'xmin', 'ymin', 'xmax', 'ymax', 'class'] xml_df = pd.DataFrame(xml_list, columns=column_name) return xml_df if __name__ == '__main__': #for folder in ['train','test']:
# image_path = os.path.join(os.getcwd(),('images/'+folder))
# xml_path = xml_to_csv(image_path)
# xml_df.to_csv(('images/'+folder+'_'labels.csv'),index = None
image_path = path xml_df = xml_to_csv(image_path) ## 修改文件名稱
xml_df.to_csv(('/home/chenxin/models-master/research/object_detection/images/train1/'+'train1.csv'), index=None) print('Successfully converted xml to csv.')
3、xml轉csv格式（需識別兩個及以上類別）
# coding: utf-8
import glob import pandas as pd import xml.etree.ElementTree as ET classes = ["person","hat"] def xml_to_csv(path): train_list = [] eval_list = [] for cls in classes: xml_list = [] # 讀取注釋文件
for xml_file in glob.glob(path + '/*.xml'): tree = ET.parse(xml_file) root = tree.getroot() for member in root.findall('object'): if cls == member[0].text: value = (root.find('filename').text, int(root.find('size')[0].text), int(root.find('size')[1].text), member[0].text, float(member[4][0].text), float(member[4][1].text), float(member[4][2].text), float(member[4][3].text) ) xml_list.append(value) for i in range(0, int(len(xml_list) * 0.9)): train_list.append(xml_list[i]) for j in range(int(len(xml_list) * 0.9) + 1, int(len(xml_list))): eval_list.append(xml_list[j]) column_name = ['filename', 'width', 'height', 'class', 'xmin', 'ymin', 'xmax', 'ymax'] # 保存為CSV格式
train_df = pd.DataFrame(xml_list, columns=column_name) eval_df = pd.DataFrame(eval_list, columns=column_name) train_df.to_csv('/home/chenxin/models-master/research/object_detection/images/data/train531.csv', index=None) eval_df.to_csv('/home/chenxin/models-master/research/object_detection/images/data/eval531.csv', index=None) def main(): # path = 'E:\\\data\\\Images'
path = r'/home/chenxin/models-master/research/object_detection/images/train531' # path參數更具自己xml文件所在的文件夾路徑修改
xml_to_csv(path) print('Successfully converted xml to csv.') main()
4、csv轉record格式
（終端運行版：python generate_tfrecord.py --csv_input=/home/chenxin/models-master/research/object_detection/images/data/train.csv --output_path=/home/chenxin/models-master/research/object_detection/images/data/train.record）
# generate_tfrecord.py # -*- coding: utf-8 -*- #!/usr/bin/env python
""" @Time: 2020/2/10 22：00 @Author: chenxin @File Name: csv2tfrecord.py @Software: PyCharm """
""" Usage: # From tensorflow/models/ # Create train data: python generate_tfrecord.py --csv_input=/home/chenxin/models-master/research/object_detection/images/data/train.csv --output_path=/home/chenxin/models-master/research/object_detection/images/data/train.record # Create test data: python generate_tfrecord.py --csv_input=/home/chenxin/models-master/research/object_detection/images/data/test.csv --output_path=/home/chenxin/models-master/research/object_detection/images/data/test.record """
import os import io import pandas as pd import tensorflow as tf from PIL import Image from object_detection.utils import dataset_util from collections import namedtuple,OrderedDict os.chdir('/home/chenxin/models-master/research/object_detection') flags = tf.app.flags flags.DEFINE_string('csv_input','','Path to CSV input') flags.DEFINE_string('out_path','','Path to output TFRecord') FLAGS = flags.FLAGS def class_text_to_int(row_label): if row_label == 'hat': return 1
if row_label == 'person': return 2
else : return 3
def split(df,group): data = namedtuple('data',['filename','object']) gb = df.groupby(group) return [data(filename,gb.get_group(x)) for filename,x in zip(gb.groups.keys(),gb.groups)] def create_tf_example(group,path): with tf.gfile.GFile(os.path.join(path,'{}'.format(group.filename)),'rb') as fid: encoded_jpg = fid.read() encoded_jpg_io = io.BytesIO(encoded_jpg) image = Image.open(encoded_jpg_io) width,height = image.size filename = group.filename.encode('utf8') image_format = b'jpg' xmins = [] xmaxs = [] ymins = [] ymaxs = [] classes_text = [] classes = [] for index,row in group.object.iterrows(): xmins.append(row['xmin']/width) xmaxs.append(row['xmax']/width) ymins.append(row['ymin']/height) ymaxs.append(row['ymax']/height) classes_text.append(row['class'].encode('utf8')) classes.append(class_text_to_int(row['class'])) tf_example = tf.train.Example(features = tf.train.Features(feature = { 'image/height':dataset_util.int64_feature(height), 'image/width':dataset_util.int64_feature(width), 'image/filename':dataset_util.bytes_feature(filename), 'image/source_id':dataset_util.bytes_feature(filename), 'image/encoded':dataset_util.bytes_feature(encoded_jpg), 'image/format':dataset_util.bytes_feature(image_format), 'image/object/bbox/xmin':dataset_util.float_list_feature(xmins), 'image/object/bbox/xmax':dataset_util.float_list_feature(xmaxs), 'image/object/bbox/ymin':dataset_util.float_list_feature(ymins), 'image/object/bbox/ymax':dataset_util.float_list_feature(ymaxs), 'image/object/class/text':dataset_util.bytes_list_feature(classes_text), 'image/object/class/label':dataset_util.int64_list_feature(classes), })) return tf_example def main(self): writer = tf.python_io.TFRecordWriter(FLAGS.output_path) path = os.path.join(os.grtcwd(),FLAGS.image_dir) examples = pd.read_csv(FLAGS.csv_input) grouped = split(examples,'filename') for group in grouped: tf_example = create_tf_example(group,path) writer.write(tf_example.SerialzeToString()) writer.close() output_path = os.path.join(os.getcwd(), FLAGS.output_path) #output_path = os.path.join(os.getcwd(),'images/data')
print('Successfully created the TFRecords:{}'.format(output_path)) if __name__ == '__main__': tf.app.run()
四、配置文件和下載模型
1、 xml位置：/home/chenxin/models-master/research/object_detection/images/xml
2、 xml2csv文件位置：/home/chenxin/models-master/research/object_detection/images/data（自定義）
3、 csv2tfrecord文件位置：/home/chenxin/models-master/research/object_detection/images/data（自定義）
4、 在官方提供的model zoo里下載訓練好的模型。我們使用ssd_mobilenet_v1_coco，先下載它。
在 object_dection文件夾下，解壓ssd_mobilenet_v1_coco.tar.gz
模型存放位置：ssd_mobilenet_v1_coco：/home/chenxin/models-master/research/object_detection/
5、建立training文件夾：將ssd_mobilenet_v1_coco.config 放在training 文件夾下
training文件位置：/home/chenxin/models-master/research/object_detection/training
進行如下更改：
5.1、搜索其中的 PATH_TO_BE_CONFIGURED ，將對應的路徑改為自己的路徑;
注意最后train input reader和evaluation input reader中label_map_path必須保持一致。
5.2、將 num_classes 按照實際情況更改，我的例子中是1；
5.3、batch_size 原本是24，我在運行的時候出現顯存不足的問題，為了保險起見，改為1，如果1還是出現類似問題的話，建議換電腦……
5.4、fine_tune_checkpoint: "ssd_mobilenet_v1_coco/model.ckpt"
from_detection_checkpoint: true
(pipeline.config里面標紅的都是需要更改的）
model { ssd { num_classes: 2 image_resizer { fixed_shape_resizer { height: 300 width: 300 } } feature_extractor { type: "ssd_mobilenet_v1" depth_multiplier: 1.0 min_depth: 16 conv_hyperparams { regularizer { l2_regularizer { weight: 3.9999999e-05 } } initializer { truncated_normal_initializer { mean: 0.0 stddev: 0.029999999 } } activation: RELU_6 batch_norm { decay: 0.99970001 center: true scale: true epsilon: 0.001 train: true } } use_depthwise: true } box_coder { faster_rcnn_box_coder { y_scale: 10.0 x_scale: 10.0 height_scale: 5.0 width_scale: 5.0 } } matcher { argmax_matcher { matched_threshold: 0.5 unmatched_threshold: 0.5 ignore_thresholds: false negatives_lower_than_unmatched: true force_match_for_each_row: true } } similarity_calculator { iou_similarity { } } box_predictor { convolutional_box_predictor { conv_hyperparams { regularizer { l2_regularizer { weight: 3.9999999e-05 } } initializer { truncated_normal_initializer { mean: 0.0 stddev: 0.029999999 } } activation: RELU_6 batch_norm { decay: 0.99970001 center: true scale: true epsilon: 0.001 train: true } } min_depth: 0 max_depth: 0 num_layers_before_predictor: 0 use_dropout: false dropout_keep_probability: 0.80000001 kernel_size: 3 box_code_size: 4 apply_sigmoid_to_scores: false use_depthwise: true } } anchor_generator { ssd_anchor_generator { num_layers: 6 min_scale: 0.2 max_scale: 0.94999999 aspect_ratios: 1.0 aspect_ratios: 2.0 aspect_ratios: 0.5 aspect_ratios: 3.0 aspect_ratios: 0.33329999 } } post_processing { batch_non_max_suppression { score_threshold: 9.9999999e-09 iou_threshold: 0.60000002 max_detections_per_class: 100 max_total_detections: 100 } score_converter: SIGMOID } normalize_loss_by_num_matches: true loss { localization_loss { weighted_smooth_l1 { } } classification_loss { weighted_sigmoid { } } hard_example_miner { num_hard_examples: 3000 iou_threshold: 0.99000001 loss_type: CLASSIFICATION max_negatives_per_positive: 3 min_negatives_per_image: 0 } classification_weight: 1.0 localization_weight: 1.0 } } } train_config { batch_size: 1 data_augmentation_options { random_horizontal_flip { } } data_augmentation_options { ssd_random_crop { } } optimizer { rms_prop_optimizer { learning_rate { exponential_decay_learning_rate { initial_learning_rate: 0.0040000002 decay_steps: 800720 decay_factor: 0.94999999 } } momentum_optimizer_value: 0.89999998 decay: 0.89999998 epsilon: 1.0 } } fine_tune_checkpoint: "/home/chenxin/models-master/research/object_detection/mask_rcnn_inception_v2_coco/model.ckpt" from_detection_checkpoint: true num_steps: 20000 } train_input_reader { label_map_path: "/home/chenxin/models-master/research/object_detection/images/data/chenxin.pbtxt" tf_record_input_reader { input_path: "/home/chenxin/models-master/research/object_detection/images/data/train.record" } } eval_config { num_examples: 20 max_evals: 10 retain_original_images: true } eval_input_reader { label_map_path: "/home/chenxin/models-master/research/object_detection/images/data/chenxin.pbtxt" shuffle: false num_readers: 1 tf_record_input_reader { input_path: "/home/chenxin/models-master/research/object_detection/images/data/eval.record" } }
6、在前面xml2csv文件夾（/data）下，創建一個 xxx.pbtxt的文本文件(這個id順序要與前面生成tfrecod的py文件代碼順序一樣）
item{ id:1 name:'hat' } item{ id:2 name:'person' }
五、訓練模型
1、本地GPU訓練（本機環境：Ubuntu 18.04），終端進入 object_detection目錄下，最新版用model_main.py
2、如果是python3訓練，添加list()到 model_lib.py的大概390行category_index.values()變成：list(category_index.values())，否則會有 can't pickle dict_values ERROR出現
3、准備好所有后，在終端輸入如下命令：
python model_main.py --logtostderr --model_dir=/home/chenxin/models-master/research/object_detection/training/ --pipeline_config_path=/home/chenxin/models-master/research/object_detection/training/pipeline.config
4、另開一個終端，同樣進入到object_detection目錄下，輸入：
tensorboard --logdir=training
5、運行一段時間后，我們可以看到我們的training文件夾下已經有模型數據保存了，接下來就可以生成我們的需要的模型文件了，終端在object_detection目錄下，輸入：
python3 export_inference_graph.py --input_type image_tensor --pipeline_config_path /home/chenxin/models-master/research/object_detection/training/ssd_mobilenet_v1_coco.config --trained_checkpoint_prefix /home/chenxin/models-master/research/object_detection/training/model.ckpt-11945 --output_directory ssd_mobilenet_v1_coco
其中，trained checkpoint 后面接着的數字改為如上圖里面的任意一個數字， output為想要將模型存放在何處，新建了一個文件夾xxx_detction 。運行結束后，就可以在xxx_detction文件夾下看到若干文件，有saved_model、checkpoint、frozen_inference_graph.pb等。 .pb結尾的就是最重要的frozen model.
六、測試模型
1、將object_detection目錄下的object_detection_tutorial.ipynb打開
2、不用下載模型，下載相關代碼可以刪除，model name, path to labels , num classes 更改成自己的，download model部分都刪去。
3、將訓練好的模型放到自定義位置，我的文件位置：/home/chenxin/文檔/ssd_mobilenet_v1_coco，測試圖片，准備幾張放入/home/chenxin/文檔/ssd_mobilenet_v1_coco/test images文件夾中，命名images+數字.jpg的格式，就不用改代碼,再在ssd_mobilenet_v1_coco里面新建data/文件夾，將xxx.pbtxt
4、一行更改自己圖片的數字序列就好了，range(1,10)，我的圖片命名從1至9.
# For the sake of simplicity we will use only 2 images: # image1.jpg # image2.jpg # If you want to test the code with your images, just add path to the images to the TEST_IMAGE_PATHS. PATH_TO_TEST_IMAGES_DIR = '/home/chenxin/文檔/ssd_mobilenet_v1_coco/test_images' TEST_IMAGE_PATHS = [ os.path.join(PATH_TO_TEST_IMAGES_DIR, 'image{}.jpg'.format(i)) for i in range(1, 10) ]
