0x00 环境
OS: Ubuntu 18.10 x64
Anaconda: 4.6.12
Python: 3.6.8
TensorFlow: 1.13.1
OpenCV: 3.4.1
0x01 基础环境配置
Anaconda 下载地址: Anaconda-4.6.12-Linux
conda install python=3.6安装 OpenCV 3.4.1
conda install opencv=3.4.1安装 TensorFlow 1.13.1
conda install tensorflow=1.13.10x02 TensorFlow Models
下载地址: Github – TensorFlow Models
下载后得到一个 models-master.zip 文件，解压后移动到 /usr/local/anaconda3/lib/python3.6/site-packages/tensorflow 文件夹下，并重命名为 models
unzip models-master.zip mv models /usr/local/anaconda3/lib/python3.6/site-packages/tensorflow进入 models/research/ 目录，并编译 protobuf
cd /usr/local/anaconda3/lib/python3.6/site-packages/tensorflow/models/research protoc object_detection/protos/*.proto --python_out=.安装 object_detection 库
python setup.py build python setup.py install设置 PYTHONPATH
export PYTHONPATH=$PYTHONPATH:/usr/local/anaconda3/lib/python3.6/site-packages/tensorflow/models/research export PYTHONPATH=$PYTHONPATH:/usr/local/anaconda3/lib/python3.6/site-packages/tensorflow/models/research/slim直接执行以上命令只会在当前终端生效，将以上命令写入 ~/.bashrc 并执行如下命令可以永久保存
source ~/.bashrc测试 object_detection 库是否安装成功
python object_detection/builders/model_builder_test.py进入 object_detection/ 目录并启动 jupyter-notebook，测试目标检测
cd object_detection/ jupyter-notebook在浏览器中打开 http://localhost:8888/，进入 jupyter-notebook 控制台，打开 object_detection_tutorial.ipynb 文件并执行，待模型下载完成并检测完成后会在页面底部出现两张标注后的图片
0x03 训练
下载 VOC 2012 数据集: VOCtrainval_11-May-2012.tar
在 object_detection/ 目录下创建目录 ssd_model，并解压数据集至 object_detection/ssd_model
mkdir ssd_model/ cd ssd_model tar xvf VOCtrainval_11-May-2012.tar返回 research/ 目录，执行 train 和 val 脚本
cd ../.. python ./object_detection/dataset_tools/create_pascal_tf_record.py --label_map_path=./object_detection/data/pascal_label_map.pbtxt --data_dir=object_detection/ssd_model/VOCdevkit/ --year=VOC2012 --set=train --output_path=./object_detection/ssd_model/pascal_train.record python ./object_detection/dataset_tools/create_pascal_tf_record.py --label_map_path=./object_detection/data/pascal_label_map.pbtxt --data_dir=./object_detection/ssd_model/VOCdevkit/ --year=VOC2012 --set=val --output_path=./object_detection/ssd_model/pascal_val.record这两个脚本会在 ssd_model/ 目录下生成 pascal_train.record 和 pascal_val.record 两个文件，各 600M 左右
复制配置文件，在此基础上修改，并训练数据
cp object_detection/data/pascal_label_map.pbtxt object_detection/ssd_model/ cp object_detection/samples/configs/ssd_mobilenet_v1_pets.config object_detection/ssd_model/pascal_label_map.pbtxt 文件中保存了数据集中有哪些 label
将 ssd_mobilenet_v1_pets.config 中的 num_classes 改为 pascal_label_map.pbtxt 中列出的文件数量，这里是 20，并修改迭代次数 num_steps，并将配置文件末尾的路径按照如下格式修改
train_input_reader: { tf_record_input_reader { input_path: "/usr/local/anaconda3/lib/python3.6/site-packages/tensorflow/models/research/object_detection/ssd_model/pascal_train.record" } label_map_path: "/usr/local/anaconda3/lib/python3.6/site-packages/tensorflow/models/research/object_detection/ssd_model/pascal_label_map.pbtxt" } eval_input_reader: { tf_record_input_reader { input_path: "/usr/local/anaconda3/lib/python3.6/site-packages/tensorflow/models/research/object_detection/ssd_model/pascal_val.record" } label_map_path: "/usr/local/anaconda3/lib/python3.6/site-packages/tensorflow/models/research/object_detection/ssd_model/pascal_label_map.pbtxt" shuffle: false num_readers: 1 }下载 ssd_mobilenet 至 ssd_model/ 目录下，解压并重命名为 ssd_mobilenet
ssd_mobilenet: ssd_mobilenet_v1_coco_11_06_2017.tar.gz
tar zxvf ssd_mobilenet_v1_coco_11_06_2017.tar.gz mv ssd_mobilenet_v1_coco_11_06_2017 ssd_mobilenet将 ssd_mobilenet_v1_pets.config 中 fine_tune_checkpoint 修改为如下格式的路径
fine_tune_checkpoint: "/usr/local/anaconda3/lib/python3.6/site-packages/tensorflow/models/research/object_detection/ssd_model/ssd_mobilenet/model.ckpt"使用 train.py 脚本训练模型
注意：脚本可能位于 object_detection/ 或 object_detection/legacy/ 目录下
这里位于 object_detection/legacy/ 目录
python ./object_detection/legacy/train.py --train_dir ./object_detection/legacy/train/ --pipeline_config_path ./object_detection/ssd_model/ssd_mobilenet_v1_pets.config运行 export_inference_graph.py 脚本将训练出的模型固化成 TensorFlow 的 .pb 模型，其中 trained_checkpoint_prefix 要设置成 model.ckpt-[step]，其中 step 要与训练迭代次数相同
python ./object_detection/export_inference_graph.py --input_type image_tensor --pipeline_config_path ./object_detection/ssd_model/ssd_mobilenet_v1_pets.config --trained_checkpoint_prefix ./object_detection/legacy/train/model.ckpt-9000 --output_directory ./object_detection/ssd_model/model/转换后生成的 .pb 模型位于 object_detection/ssd_model/model/ 目录下
将 pascal_label_map.pbtxt 作为 label 文件
