tensorflow 目标检测API及jupyter notebook 虚拟环境配置

环境

tensorflow == 2.8.0
win10 or linux

概要

目标检测项目的主要步骤如下：

搭建虚拟环境
采集图像并打标
训练
预测
模型的保存和转换
调优
项目部署

1. 搭建虚拟环境

步骤1： 创建虚拟环境

创建文件夹TFODCourse,并在该目录下安装虚拟环境

cd TFODCourse
python -m venv tfod

步骤2： 激活虚拟环境

LInux 环境

source tfod/bin/activate

Windows环境

.\tfod\Scripts\activate

步骤3： 安装依赖包并添加虚拟环境到Jupyter Notebook

python -m pip install --upgrade pip                       # 升级pip
pip install ipykernel                                               # jupyter notebook 与 虚拟环境关联
python -m ipykernel install --user --name=tfod     # 将虚拟环境安装到hjupyter notebook里

提示：Installed kernelspec tfod in C:\Users\einrj\AppData\Roaming\jupyter\kernels\tfod表示安装成功

步骤4： 打开jupyter notebook

jupyter notebook

错误提示1 ： Jupyter command jupyter-notebook not found.

解决方法，在命令行安装 jupyter

pip install jupyter

然后执行 jupyter notebook 就可以打开 notebook了

可以看到tfod 虚拟环境已经成功安装到Jupyter notebook中了

注意1 如果jupyter notebook没有处在虚拟环境中，如下所示:

需要切换到虚拟环境tfod,通过 Kernel -> Change Kernel - > tfod

2. 图片采集并打标

Tips:

打标尽量的紧凑，紧贴目标(Keep your label as ‘tight’ as possible)
采集的图片应该包含多种角度和光照条件
每种类型的图片可以预先采集10-20张进行训练

2.1 导入依赖包

!pip install opencv-python    # jupyter notebook中运行需要加 !

import cv2
import uuid   # 产生唯一表示，用来进行图片命名
import os
import time

2.2 定义图片种类

labels= ['thumbsup','thumbsdown','thankyou','livelong'] #可自定义标签
number_imgs=5  #理想最少10-20张

2.3 创建文件夹

IMAGES_PATH=os.path.join(`Tensorflow`,'workspace','images','collectedimages')

if not os.path.exists(IMAGES_PATH):if os.name == 'posix':          # linux !mkdir -p {IMAGES_PATH}if os.name == 'nt':             # Windows!mkdir {IMAGES_PATH}
for label in labels:path=os.path.join(IMAGES_PATH,label)if not os.path.exists(path):!mkdir {path}

2.4 拍摄图片

for label in labels:cap=cv2.VideoCapture(0)  print('Collecting images for {}'.format(label))time.sleep(5)   # leave time to chanage posefor imgnum in range (number_imgs):print('Collecting image {}'.format(imgnum))ret,frame = cap.read()imgname=os.path.join(IMAGES_PATH,label,label+'.'+'{}.jpg'.format(str(uuid.uuid1())))cv2.imwrite(imgname,frame)cv2.imshow('frame',frame)time.sleep(2)if cv2.waitKey(1)  &  0xFF == ord('q'):break
cap.release()
cv2.destroyAllWindows()

拍摄好图片，其中如果有部分照片没拍好，可以删掉重新针对缺失的照片进行采集

2.5 图片标注

使用labelimg工具进行标注

github给出的labelimg编译教程，如下

安装labelmg依赖包

!pip install --upgrade pyqt5 lxml

LABELING_PATH=os.path.join('Tensorflow','labelimg')

将源码拷贝到labelimg文件夹下

If not os.path.exists(LABELING_PATH):!mkdir {LABELING_PATH}!git clone https://github.com/tzutalin/labelImg  {LABELING_PATH}

源码编译labelimg

if os.name == 'posix':!make qt5py3
if os.name == 'nt':! cd {LABELING_PATH} && pyrcc5 -o libs/resources.py resources.qrc

启动labelmg

!cd {LABELING_PATH} && python labelImg.py

标注

具体labelimg的使用参考labelimg使用教程

2.6 划分测试集和训练集

创建train 和test文件夹

TRAIN_PATH=os.path.join('Tensorflow','workspace','images','train')
TEST_PATH=os.path.join('Tensorflow','workspace','images','test')

if not os.path.exists(TRAIN_PATH):if os.name == 'posix':          # linux !mkdir -p {TRAIN_PATH}if os.name == 'nt':             # Windows!mkdir {TRAIN_PATH}if not os.path.exists(TEST_PATH):if os.name == 'posix':          # linux !mkdir -p {TEST_PATH}if os.name == 'nt':             # Windows!mkdir {TEST_PATH}

由于文件比较少，直接将collectedimages下各个类别的图片前80%的图片及标注xml文件迁移到train文件夹，剩余的图片及标注xml文件迁移到test文件夹

train 文件
test文件

3. 训练

Tensorflow Model Zoo

使用迁移学习，对模型进行训练。Tensorflow Model Zoo有大量预训练模型可供选择。
可以不知道模型结构的细节，作为一个开发应用人员可以直接使用这些模型，不同模型对于不同的精度和运行速度。本文使用的模型是SSD MobileNet V2 FPNLite 320x320,用过鼠标右键选择下载链接或者直接下载模型。

其中320 x320 我们自己不需要把图片压缩到320 x320，传入图片后模型自动压缩到320 x 320，然后转换为原始图片的分辨率。预训练模型我们不需要做大量的前处理和后处理工作，模型已经帮忙做好。

3.1 设置路径

import os

CUSTOM_MODEL_NAME='my_ssd_mobnet'
PRETRAINED_MODEL_NAME = 'ssd_mobilenet_v2_fpnlite_320x320_coco17_tpu-8'
PRETRAIN_MODEL_URL = 'http://download.tensorflow.org/models/object_detection/tf2/20200711/ssd_mobilenet_v2_fpnlite_320x320_coco17_tpu-8.tar.gz'
TF_RECORD_SCRIPT_NAME = 'generate_tfrecord.py'
LABEL_MAP_NAME='label_map.pbtxt'

paths={'WORKWAPCE_PATH':os.path.join('Tensorflow','workspace'),
'SCRIPTS_PATH':os.path.join('Tensorflow','scripts'),
'APIMODEL_PATH':os.path.join('Tensorflow','models'),
'ANNOTATION_PATH':os.path.join('Tensorflow','workspace','annotations'),
'IMAGE_PATH':os.path.join('Tensorflow','workspace','images'),
'MODEL_PATH':os.path.join('Tensorflow','workspace','models'),
'PRETRAINED_MODEL_PATH':os.path.join('Tensorflow','workspace','pre-trained-models'),
'CHECKPOINT_PATH':os.path.join('Tensorflow','workspace','models',CUSTOM_MODEL_NAME),
'OUTPUT_PATH':os.path.join('Tensorflow','workspace','models',CUSTOM_MODEL_NAME,'export'),
'TFJS_PATH':os.path.join('Tensorflow','workspace','models',CUSTOM_MODEL_NAME,'tfjsexport'),
'TFLITE_PATH':os.path.join('Tensorflow','workspace','models',CUSTOM_MODEL_NAME,'tfliteexport'),
'PROTOC_PATH':os.path.join('Tensorflow','protoc')}

files = {'PIPELINE_CONFIG':os.path.join('Tensorflow','workspace','models',CUSTOM_MODEL_NAME,'pipeline.config'),'TF_RECORD_SCRIPT':os.path.join(paths['SCRIPTS_PATH'],TF_RECORD_SCRIPT_NAME),'LABELMAP':os.path.join(paths['ANNOTATION_PATH'],LABEL_MAP_NAME)
}

for path in paths.values():if not os.path.exists(path):if os.name == 'posix':!mkdir -p {path}if os.name == 'nt':!mkdir {path}

3.2 下载Tensorflow Model Zoo 并安装TFOD（tensorflow object detection）

if os.name == 'nt':!pip install wgetimport wget

下载Tensorflow/models源码到 APIMODEL_PATH

if not os.path.exists(os.path.join(paths['APIMODEL_PATH'],'research','object_detection')):!git clone https://github.com/tensorflow/models  {paths['APIMODEL_PATH']}

安装Tensorflow object Dectection API

if os.name == 'posix':!apt-get install protobuf-compiler!cd Tensorflow/models/research && protoc object_detection/protos/*.proto --python_out=. && cp object_detection/packages/tf2/setup.py . && python -m pip install . if os.name=='nt':url="https://github.com/protocolbuffers/protobuf/releases/download/v3.15.6/protoc-3.15.6-win64.zip"wget.download(url)!move protoc-3.15.6-win64.zip {paths['PROTOC_PATH']}!cd {paths['PROTOC_PATH']} && tar -xf protoc-3.15.6-win64.zipos.environ['PATH'] += os.pathsep + os.path.abspath(os.path.join(paths['PROTOC_PATH'], 'bin'))   !cd Tensorflow/models/research && protoc object_detection/protos/*.proto --python_out=. && copy object_detection\\packages\\tf2\\setup.py setup.py && python setup.py build && python setup.py install!cd Tensorflow/models/research/slim && pip install -e .

注: 提示git 或者tar 不是内部或者外部命令，需要安装git、tar并加入环境变量，如果不想安装可以手动下载到指定的位置

检测是否安装成功

VERIFICATION_SCRIPT = os.path.join(paths['APIMODEL_PATH'], 'research', 'object_detection', 'builders', 'model_builder_tf2_test.py')
# Verify Installation
!python {VERIFICATION_SCRIPT}

运行验证脚本，如果验证没有问题，说明Tensorflow 目标检测API已安装成功

出现的问题1： ModuleNotFoundError: No module named ‘tensorflow’
解决：运行如下命令

!pip install tensorflow --upgrade

重新运行

!python {VERIFICATION_SCRIPT}

出现的问题2： ModuleNotFoundError: No module named ‘cycler’

解决2：

!pip install cycler

然后重新运行验证脚本
问题3：ModuleNotFoundError: No module named ‘yaml’
解决3：

!pip install pyyaml

出现如下所示，说明tensorflow 目标检测API已经安装成功

导入object_detection

import object_detection

解决方法：重启jupyter notebook内核

如果需要执行GPU加速，需要下载tensorflow对应版本的cudnn和cuda
参见：
- TensorFlow与cudnn和cuda的对应关系 ,同时需要下载对应Visual Studio 编译器
- Tensorflow2.0 安装cuda和cudnn教程
从Tensorflow Model Zoo 下载预训练模型

if os.name =='posix':!wget {PRETRAINED_MODEL_URL}!mv {PRETRAINED_MODEL_NAME+'.tar.gz'} {paths['PRETRAINED_MODEL_PATH']}!cd {paths['PRETRAINED_MODEL_PATH']} && tar -zxvf {PRETRAINED_MODEL_NAME+'.tar.gz'}
if os.name == 'nt':wget.download(PRETRAINED_MODEL_URL)!move {PRETRAINED_MODEL_NAME+'.tar.gz'} {paths['PRETRAINED_MODEL_PATH']}!cd {paths['PRETRAINED_MODEL_PATH']} && tar -zxvf {PRETRAINED_MODEL_NAME+'.tar.gz'}

3.3 创建标签映射

labels = [{'name':'ThumbsUp', 'id':1}, {'name':'ThumbsDown', 'id':2}, {'name':'ThankYou', 'id':3}, {'name':'LiveLong', 'id':4}]with open(files['LABELMAP'], 'w') as f:for label in labels:f.write('item { \n')f.write('\tname:\'{}\'\n'.format(label['name']))f.write('\tid:{}\n'.format(label['id']))f.write('}\n')

3.3 创建TF Records

下载 generate_tfrecord.py

if not os.path.exists(files['TF_RECORD_SCRIPT']):!git clone https://github.com/nicknochnack/GenerateTFRecord {paths['SCRIPTS_PATH']}

执行脚本生成TF Records

!python {files['TF_RECORD_SCRIPT']} -x {os.path.join(paths['IMAGE_PATH'], 'train')} -l {files['LABELMAP']} -o {os.path.join(paths['ANNOTATION_PATH'], 'train.record')}
!python {files['TF_RECORD_SCRIPT']} -x {os.path.join(paths['IMAGE_PATH'], 'test')} -l {files['LABELMAP']} -o {os.path.join(paths['ANNOTATION_PATH'], 'test.record')}

运行提示如下报错信息

执行

!pip install pytz

3.4 复制Model Config 到训练目录下

if os.name =='posix':!cp {os.path.join(paths['PRETRAINED_MODEL_PATH'], PRETRAINED_MODEL_NAME, 'pipeline.config')} {os.path.join(paths['CHECKPOINT_PATH'])}
if os.name == 'nt':!copy {os.path.join(paths['PRETRAINED_MODEL_PATH'], PRETRAINED_MODEL_NAME, 'pipeline.config')} {os.path.join(paths['CHECKPOINT_PATH'])}

3.4 更新Config文件用来做迁移学习

import tensorflow as tf
from object_detection.utils import config_util
from object_detection.protos import pipeline_pb2
from google.protobuf import text_format

config = config_util.get_configs_from_pipeline_file(files['PIPELINE_CONFIG'])

根据config文件，可以看出：label_map_path: "PATH_TO_BE_CONFIGURED"， input_path: "PATH_TO_BE_CONFIGURED"，fine_tune_checkpoint: "PATH_TO_BE_CONFIGURED",tf_record_input_reader:"PATH_TO_BE_CONFIGURED"，这些路径需要我们重新去定义

执行代码，更新config文件。

pipeline_config = pipeline_pb2.TrainEvalPipelineConfig()
with tf.io.gfile.GFile(files['PIPELINE_CONFIG'], "r") as f:                                                                                                                                                                                                                     proto_str = f.read()                                                                                                                                                                                                                                          text_format.Merge(proto_str, pipeline_config)

pipeline_config.model.ssd.num_classes = len(labels)
pipeline_config.train_config.batch_size = 4
pipeline_config.train_config.fine_tune_checkpoint = os.path.join(paths['PRETRAINED_MODEL_PATH'], PRETRAINED_MODEL_NAME, 'checkpoint', 'ckpt-0')
pipeline_config.train_config.fine_tune_checkpoint_type = "detection"
pipeline_config.train_input_reader.label_map_path= files['LABELMAP']
pipeline_config.train_input_reader.tf_record_input_reader.input_path[:] = [os.path.join(paths['ANNOTATION_PATH'], 'train.record')]
pipeline_config.eval_input_reader[0].label_map_path = files['LABELMAP']
pipeline_config.eval_input_reader[0].tf_record_input_reader.input_path[:] = [os.path.join(paths['ANNOTATION_PATH'], 'test.record')]

注：``

config_text = text_format.MessageToString(pipeline_config)
with tf.io.gfile.GFile(files['PIPELINE_CONFIG'], "wb") as f:                                                                                                                                                                                                                     f.write(config_text)

4 训练模型

4.1 训练模型

TRAINING_SCRIPT = os.path.join(paths['APIMODEL_PATH'], 'research', 'object_detection', 'model_main_tf2.py')

command = "python {} --model_dir={} --pipeline_config_path={} --num_train_steps=3000".format(TRAINING_SCRIPT, paths['CHECKPOINT_PATH'],files['PIPELINE_CONFIG'])

print(command)

python Tensorflow\models\research\object_detection\model_main_tf2.py --model_dir=Tensorflow\workspace\models\my_ssd_mobnet_tuned --pipeline_config_path=Tensorflow\workspace\models\my_ssd_mobnet_tuned\pipeline.config --num_train_steps=2000

为了看到模型训练的进度，复制训练脚本到windows 命令窗口中进行训练，因为jupyter notebook观察不到训练进度

提示No module named 'gin'

pip install gin-config == 0.1.1

提示：No module named 'tensorflow_addons'

pip install tensorflow_addons

![在这里插入图片描述](https://img-blog.csdnimg.cn/461df7aeeb524a4b99639fff77b13e4f.png

训练完成

4.2 模型评估

command = "python {} --model_dir={} --pipeline_config_path={} --checkpoint_dir={}".format(TRAINING_SCRIPT, paths['CHECKPOINT_PATH'],files['PIPELINE_CONFIG'], paths['CHECKPOINT_PATH'])

print(command)

python Tensorflow\models\research\object_detection\model_main_tf2.py --model_dir=Tensorflow\workspace\models\my_ssd_mobnet_tuned --pipeline_config_path=Tensorflow\workspace\models\my_ssd_mobnet_tuned\pipeline.config --checkpoint_dir=Tensorflow\workspace\models\my_ssd_mobnet_tuned

4.3 tensorboard 展示

模型训练和验证后，生成的结果保存如下：

cd 到train 或者 eval 执行如下命令

tensorboard --logdir=.

4.4 加载模型

import os
import tensorflow as tf
from object_detection.utils import label_map_util
from object_detection.utils import visualization_utils as viz_utils
from object_detection.builders import model_builder
from object_detection.utils import config_util

# Load pipeline config and build a detection model
configs = config_util.get_configs_from_pipeline_file(files['PIPELINE_CONFIG'])
detection_model = model_builder.build(model_config=configs['model'], is_training=False)# Restore checkpoint
ckpt = tf.compat.v2.train.Checkpoint(model=detection_model)
ckpt.restore(os.path.join(paths['CHECKPOINT_PATH'], 'ckpt-3')).expect_partial()@tf.function
def detect_fn(image):image, shapes = detection_model.preprocess(image)prediction_dict = detection_model.predict(image, shapes)detections = detection_model.postprocess(prediction_dict, shapes)return detections

其中ckpt.restore(os.path.join(paths['CHECKPOINT_PATH'], 'ckpt-3')).expect_partial() 其中参数ckpt-3根据保存的模型文件最新的ckpt修改。

5 预测

5.1 单张图片检测

import cv2
import numpy as np
from matplotlib import pyplot as plt
%matplotlib inline

category_index = label_map_util.create_category_index_from_labelmap(files['LABELMAP'])

IMAGE_PATH = os.path.join(paths['IMAGE_PATH'], 'test', 'thumbsup.6223230a-947f-11ec-a250-48ba4e5e8c8f.jpg')

img = cv2.imread(IMAGE_PATH)
image_np = np.array(img)input_tensor = tf.convert_to_tensor(np.expand_dims(image_np, 0), dtype=tf.float32)
detections = detect_fn(input_tensor)num_detections = int(detections.pop('num_detections'))
detections = {key: value[0, :num_detections].numpy()for key, value in detections.items()}
detections['num_detections'] = num_detections# detection_classes should be ints.
detections['detection_classes'] = detections['detection_classes'].astype(np.int64)label_id_offset = 1
image_np_with_detections = image_np.copy()viz_utils.visualize_boxes_and_labels_on_image_array(image_np_with_detections,detections['detection_boxes'],detections['detection_classes']+label_id_offset,detections['detection_scores'],category_index,use_normalized_coordinates=True,max_boxes_to_draw=5,min_score_thresh=.8,agnostic_mode=False)plt.imshow(cv2.cvtColor(image_np_with_detections, cv2.COLOR_BGR2RGB))
plt.show()

5.2 实时视频检测

cap = cv2.VideoCapture(0)
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))while cap.isOpened(): ret, frame = cap.read()image_np = np.array(frame)input_tensor = tf.convert_to_tensor(np.expand_dims(image_np, 0), dtype=tf.float32)detections = detect_fn(input_tensor)num_detections = int(detections.pop('num_detections'))detections = {key: value[0, :num_detections].numpy()for key, value in detections.items()}detections['num_detections'] = num_detections# detection_classes should be ints.detections['detection_classes'] = detections['detection_classes'].astype(np.int64)label_id_offset = 1image_np_with_detections = image_np.copy()viz_utils.visualize_boxes_and_labels_on_image_array(image_np_with_detections,detections['detection_boxes'],detections['detection_classes']+label_id_offset,detections['detection_scores'],category_index,use_normalized_coordinates=True,max_boxes_to_draw=5,min_score_thresh=.8,agnostic_mode=False)cv2.imshow('object detection',  cv2.resize(image_np_with_detections, (800, 600)))if cv2.waitKey(10) & 0xFF == ord('q'):cap.release()cv2.destroyAllWindows()break

完整代码：