Tensorflow Object Detection API使用

安装

安装tensorflow-gpu

最新版的tensorflow不支持cuda8.0， 因此，这里安装了tensorflow1.4，用清华的镜像：

pip install \
-i https://pypi.tuna.tsinghua.edu.cn/simple/ \
https://mirrors.tuna.tsinghua.edu.cn/tensorflow/linux/gpu/tensorflow_gpu-1.4.0-cp27-none-linux_x86_64.whl

获取 TensorFlow Models 源码：

git clone https://github.com/tensorflow/models –recursive
cd models/

安装必要的python包，如pillow，lxml，jupyter，matplotlib
手动编译安装Protobuf 3.3：

wget https://github.com/google/protobuf/archive/v3.3.0.tar.gz
tar zxvf v3.3.0.tar.gz
cd protobuf-3.3.0/
./autogen.sh
./configure
make -j8
make install

如果出错，configure.ac:17: error: possibly undefined macro: AC_PROG_LIBTOOL，则安装sudo apt-get install libtool

添加环境变量export LD_LIBRARY_PATH=/usr/local/lib

编译proto:

(Folder tensorflow/models/research)
protoc object_detection/protos/*.proto –python_out=.
export PYTHONPATH=$PYTHONPATH:pwd:pwd/slim

测试是否成功

python object_detection/builders/model_builder_test.py
Ran 13 tests in 0.045s
OK

http://mp.weixin.qq.com/s/eN41M1sMLxvQMEl0IhyC9w

准备数据集，并转成tfrecords, 利用脚本create_pet_tf_record.py即可，需要调整数据目录

• images
• annotations
  
  • xmls
  • trainval.txt

python create_pet_tf_record.py –data_dir=../../Deep-Learning-master/tensorflow_toy_detector –output_dir=../../Deep-Learning-master/tensorflow_toy_detector –label_map_path=../../Deep-Learning-master/tensorflow_toy_detector/toy_label_map.pbtxt

其中，label_map是item { id: 1 name: ‘toy’}

修改配置文件，object_detection/samples/configs,修改num以及路径

训练模型，我一开始用虚拟机训练，可能内存不够，程序直接被kill了，换到gpu上就可以了。另外，注意要先删掉logs，否则总是从上一次恢复运行。

错误：local variable ‘total_loss’ referenced before assignment

似乎是一个bug，可以直接在/usr/lib/python2.7/site-packages/tensorflow/contrib/slim/python/slim/learning.py那里把total_loss=None

错误：:Caught OutOfRangeError. Stopping Training.
检查下dataset的路径是否错误

python train.py --train_dir=../../../toy/Deep-Learning-master/tensorflow_toy_detector/logs/  --pipeline_config_path=../../../toy/Deep-Learning-master/tensorflow_toy_detector/faster_rcnn_resnet101_coco.config

测试模型eval.py

python eval.py --pipeline_config_path=../../../toy/Deep-Learning-master/tensorflow_toy_detector/faster_rcnn_resnet101_coco.config --checkpoint_dir=../../../toy/Deep-Learning-master/tensorflow_toy_detector/logs/ --eval_dir=../../../toy/Deep-Learning-master/tensorflow_toy_detector/eval_logs/

可以用tensorboard –logdir=eval_logs查看训练结果

将训练好的模型导出，产生.pb的模型文件，保存在output文件夹中

python export_inference_graph.py --input_type=image_tensor --pipeline_config_path=../../../toy/Deep-Learning-master/tensorflow_toy_detector/faster_rcnn_resnet101_coco.config --trained_checkpoint_prefix=../../../toy/Deep-Learning-master/tensorflow_toy_detector/logs/model.ckpt-1058 --output_directory ../../../toy/Deep-Learning-master/tensorflow_toy_detector/output

使用保存好的模型对图片进行物体检测
我写了一个脚本detection_pic.py：

import tensorflow as tf
from PIL import Image
import numpy as np

from utils import label_map_util
from utils import visualization_utils as vis_util


PATH_TO_CKPT = '.pb'
PATH_TO_LABELS = ''
NUM_CLASS = 1

flags = tf.app.flags
flags.DEFINE_string('ckpt_path', '', 'ckpt path')
flags.DEFINE_integer('num_class', 0, 'number of classes')
flags.DEFINE_string('label_path', '', 'label path')
flags.DEFINE_string('image_path', '', 'image path')

FLAGS = flags.FLAGS

def load_image_into_numpy_array(image):
    (im_width, im_height) = image.size
    return np.array(image.getdata()).reshape(
        (im_height, im_width, 3)).astype(np.uint8)

def main(_):
    ckpt_path = FLAGS.ckpt_path
    num_class = FLAGS.num_class
    label_path = FLAGS.label_path
    image_path = FLAGS.image_path

    # load tensorflow model into memory
    detection_graph = tf.Graph()
    with detection_graph.as_default():
        od_graph_def = tf.GraphDef()
        with tf.gfile.GFile(ckpt_path, 'rb') as fid:
            serialized_graph = fid.read()
            od_graph_def.ParseFromString(serialized_graph)
            tf.import_graph_def(od_graph_def, name='')

    # load label map
    label_map = label_map_util.load_labelmap(label_path)
    categories = label_map_util.convert_label_map_to_categories(label_map,
                                                                max_num_classes=num_class,
                                                                use_display_name=True)
    category_index = label_map_util.create_category_index(categories)

    # detection
    with detection_graph.as_default():
        with tf.Session(graph=detection_graph) as sess:
            image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')
            detection_boxes = detection_graph.get_tensor_by_name('detection_boxes:0')

            detection_scores = detection_graph.get_tensor_by_name('detection_scores:0')
            detection_classes = detection_graph.get_tensor_by_name('detection_classes:0')
            num_detections = detection_graph.get_tensor_by_name('num_detections:0')

            image = Image.open(image_path)
            image_np = load_image_into_numpy_array(image)
            image_np_expanded = np.expand_dims(image_np, axis=0)

            # Actial detection
            (boxes, scores, classes, num) = sess.run(
                [detection_boxes, detection_scores, detection_classes, num_detections],
                feed_dict={image_tensor: image_np_expanded}
            )
            vis_util.visualize_boxes_and_labels_on_image_array(
                image_np,
                np.squeeze(boxes),
                np.squeeze(classes).astype(np.int32),
                np.squeeze(scores),
                category_index,
                use_normalized_coordinates=True,
                line_thickness=8
            )
            output_image = Image.fromarray(image_np)
            output_image.save('out.jpeg')

if __name__ == '__main__':
  tf.app.run()

运行以下命令：

python detection_pic.py --ckpt_path=../../../toy/Deep-Learning-master/tensorflow_toy_detector/output/frozen_inference_graph.pb --num_class=1 --label_path=../../../toy/Deep-Learning-master/tensorflow_toy_detector/toy_label_map.pbtxt --image_path=../../../toy/Deep-Learning-master/tensorflow_toy_detector/images/toy11.jpg