为一动漫网站做一个图片搜索引擎,需要先定位动漫人物的脸,直接用opencv里的人脸检测API什么都框不出来,仔细想想也对,漫画里的人物没一个是正常的,眼睛都好大,鼻子几乎没有,木法所以只好自己训练咯。
先安装TensorFlow object detection API,可以参考ubuntu16.04 TensorFlow目标检测API安装,等安装完成后,如果没有安装过tensorboard,最好顺便安装一下,以便观察loss等指标的变化
pip install tensorboard tensorboard --logdir /xxxx/model_ckpt/ # model_ckpt检查点的保存路径
一、准备数据集
tensorflow的目标检测有自己的格式,需要转换一下。首先需要一个图像标注软件,将我们数据集中每副图片的动漫脸标注出来,我使用的是labelImg,下载地址:
链接: https://pan.baidu.com/s/1i7oxr1r 密码: nkmz,解压后进入到\windows_v1.4.0\下运行labelImg.exe就ok了(需要python环境),打开图片后框出需要检测的目标,给一个标签,然后“保存”,labelImg保存的是一个个的XML文件,如此下去将训练集中的图片全部标注出来(一件没任何技术含量而且还很费时间的事情),两个小时后标注了100多张图片,暂时做测试足够了。
生成的XML文件如下:
<annotation> <folder>AIR</folder> <filename>吃团子可爱的神尾观铃.jpg</filename> <path>E:\code\image_search\image\人物\AIR\吃团子可爱的神尾观铃.jpg</path> <source> <database>Unknown</database> </source> <size> <width>800</width> <height>564</height> <depth>3</depth> </size> <segmented>0</segmented> <object> <name>person</name> <pose>Unspecified</pose> <truncated>0</truncated> <difficult>0</difficult> <bndbox> <xmin>382</xmin> <ymin>163</ymin> <xmax>572</xmax> <ymax>313</ymax> </bndbox> </object> </annotation>图片中标注了多少目标就会有几个<object>,然后开始转成tf 所需要的数据格式,具体格式说明请参考https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/using_your_own_dataset.md,这里我是先把xml转成了csv然后又转换成tf.record格式的
xml_metadata_csv.py
import os import glob import cv2 import numpy as np import pandas as pd import xml.etree.ElementTree as ET def png_to_jpg(png_path): jpg_path = png_path.replace('.png','.jpg') im = cv2.imdecode(np.fromfile(png_path, dtype=np.uint8), -1) cv2.imencode('.jpg', im)[1].tofile(jpg_path) return jpg_path def xml_to_csv(path): xml_list = [] for dir in os.listdir(path): img_dir = os.path.join(path, dir) for xml_file in glob.glob(img_dir + '\*.xml'): tree = ET.parse(xml_file) root = tree.getroot() file_name = root.find('path').text # if file_name.find('.png') > 0: # file_name = png_to_jpg(file_name) for member in root.findall('object'): value = (file_name, int(root.find('size')[0].text), int(root.find('size')[1].text), member[0].text, int(member[4][0].text), int(member[4][1].text), int(member[4][2].text), int(member[4][3].text) ) xml_list.append(value) column_name = ['filename', 'width', 'height', 'class', 'xmin', 'ymin', 'xmax', 'ymax'] xml_df = pd.DataFrame(xml_list, columns=column_name) return xml_df def main(): image_path = r'E:\code\MechineLeaning\image_search\CNN' xml_df = xml_to_csv(image_path) xml_df.to_csv('CSVMETADATAFILE.csv', index=None) print('successfully converted xml metadata to csv') if __name__ == '__main__': main()
png_to_jpg()需要注意,这个是python3写带中文路径图片的一个处理方法,不能直接用cv2.write()
csv转tf.record,generate_tfrecord.py
from __future__ import division from __future__ import print_function from __future__ import absolute_import import os import io import pandas as pd import tensorflow as tf from PIL import Image from collections import namedtuple import dataset_util import sys # reload(sys) # sys.setdefaultencoding('utf-8') flags = tf.app.flags flags.DEFINE_string('csv_input', 'CSVMETADATAFILE.csv', 'Path to the CSV input') flags.DEFINE_string('output_path', 'person_val.record', 'Path to output TFRecord') FLAGS = flags.FLAGS # 将标签转换成对应的数值,从1开始 def class_text_to_int(row_label): if row_label == 'person': return 1 else: None 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(group): with tf.gfile.GFile(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(_): writer = tf.python_io.TFRecordWriter(FLAGS.output_path) # path = os.path.join(os.getcwd(), path_images) examples = pd.read_csv(FLAGS.csv_input) grouped = split(examples, 'filename') for group in grouped: tf_example = create_tf(group) writer.write(tf_example.SerializeToString()) writer.close() output_path = os.path.join(os.getcwd(), FLAGS.output_path) print('Successfully created the TFRecords: {}'.format(output_path)) if __name__ == '__main__': tf.app.run()二、使用tf object detection API
在安装好API后的object_detection目录下有好几个py文件,使用train.py来进行模型训练,tf给出的调用例子:
Example usage: ./train \ --logtostderr \ --train_dir=path/to/train_dir \ --model_config_path=model_config.pbtxt \ --train_config_path=train_config.pbtxt \ --input_config_path=train_input_config.pbtxt所以我们还需要准备运行train.py需要的三个文件,分别是 model_config.pbtxt、train_config.pbtxt、train_input_config.pbtxt
还有一种调用方式,是配置一个pipeline_config.pbtxt文件,其实就是把上面的三个文件写在一个文件里面了,我是为了清晰点所以写了三个文件。
model_config.pbtxt
ssd { num_classes: 1 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 { } } anchor_generator { ssd_anchor_generator { num_layers: 6 min_scale: 0.2 max_scale: 0.95 aspect_ratios: 1.0 aspect_ratios: 2.0 aspect_ratios: 0.5 aspect_ratios: 3.0 aspect_ratios: 0.3333 } } image_resizer { fixed_shape_resizer { height: 300 width: 300 } } box_predictor { convolutional_box_predictor { min_depth: 0 max_depth: 0 num_layers_before_predictor: 0 use_dropout: false dropout_keep_probability: 0.8 kernel_size: 1 box_code_size: 4 apply_sigmoid_to_scores: false conv_hyperparams { activation: RELU_6, regularizer { l2_regularizer { weight: 0.00004 } } initializer { truncated_normal_initializer { stddev: 0.03 mean: 0.0 } } batch_norm { train: true, scale: true, center: true, decay: 0.9997, epsilon: 0.001, } } } } feature_extractor { type: 'ssd_mobilenet_v1' min_depth: 16 depth_multiplier: 1.0 conv_hyperparams { activation: RELU_6, regularizer { l2_regularizer { weight: 0.00004 } } initializer { truncated_normal_initializer { stddev: 0.03 mean: 0.0 } } batch_norm { train: true, scale: true, center: true, decay: 0.9997, epsilon: 0.001, } } } loss { classification_loss { weighted_sigmoid { anchorwise_output: true } } localization_loss { weighted_smooth_l1 { anchorwise_output: true } } hard_example_miner { num_hard_examples: 3000 iou_threshold: 0.99 loss_type: CLASSIFICATION max_negatives_per_positive: 3 min_negatives_per_image: 0 } classification_weight: 1.0 localization_weight: 1.0 } normalize_loss_by_num_matches: true post_processing { batch_non_max_suppression { score_threshold: 1e-8 iou_threshold: 0.6 max_detections_per_class: 100 max_total_detections: 100 } score_converter: SIGMOID } }train_config.pbtxt
batch_size: 24 optimizer { rms_prop_optimizer: { learning_rate: { exponential_decay_learning_rate { initial_learning_rate: 0.004 decay_steps: 800720 decay_factor: 0.95 } } momentum_optimizer_value: 0.9 decay: 0.9 epsilon: 1.0 } } fine_tune_checkpoint: "/home/bqh/Code/image_search/model_ckpt/model.ckpt" from_detection_checkpoint: false # Note: The below line limits the training process to 200K steps, which we # empirically found to be sufficient enough to train the pets dataset. This # effectively bypasses the learning rate schedule (the learning rate will # never decay). Remove the below line to train indefinitely. num_steps: 200000 data_augmentation_options { random_horizontal_flip { } } data_augmentation_options { ssd_random_crop { } }train_input_config.pbtxt
tf_record_input_reader { input_path: "$PATH/train.record" } label_map_path: "$PATH/label_map.pbtxt"说明:在model_config文件中,有一项 num_classes: 1,因为这里训练的只是找出图片中的漫画人物的脸所以就一个类别,如果检测的是多目标的话,有几项这里就改成几;这里使用的是SSD检测模型也可以使用别的检测模型,比如faster rcnn(理论请参考 http://blog.csdn.net/luoyang224/article/details/77529536);train_input_config文件中又多出来两个文件的路径,分别是第一步生成的训练集路径和训练集中目标的标签
label_map.pbtxt
item{ id:1 name:'person' }三、开始训练
我手头没有可供规模大点的数据集使用的GPU,所以就直接在CPU上训练的,不过速度真的太慢了,按照上面的配置文件每个batch 24长图片,大概需要16G的内存,我这里训练了四天多才迭代了进8W次。如果有条件的话还是使用GPU吧!运行命令:python train.py --logtostderr --train_dir=$PATH/model_ckpt --model_config_path=$PATH/config/model_config.pbtxt --train_config_path=$PATH/config/train_config.pbtxt --input_config_path=$PATH/config/train_input_config.pbtxt
四、检测
在API里有个文件export_inference_graph.py可以导出模型,但是运行的时候报错,在网上查了下说是因为TF的版本问题,反正不导出也OK,就直接测试吧。使用eval.py,运行的格式和上面train.py差不多:
python eval.py --logtostderr --checkpoint_dir=$PATH/model_ckpt --eval_dir=$PATH/eval_dir --eval_config_path=$PATH/config/eval_config.pbtxt --model_config_path=$PATH/config/model_config.pbtxt --input_config_path=$PATH/config/eval_input_config.pbtxt
同样需要准备eval_config.pbtxt、eval_input_config.pbtxt文件
eval_config.pbtxt
num_examples: 2000 # Note: The below line limits the evaluation process to 10 evaluations. # Remove the below line to evaluate indefinitely. # max_evals: 10eval_input_config.pbtxt
tf_record_input_reader { input_path: "$PATH/person_val.record" } label_map_path: "$PATH/eval_label_map.pbtxt" shuffle: false num_readers: 1这里的person_val.record是测试集的tf.record文件,生成方法同第一步;eval_label_map.pbtxt文件同第一步的label_map.pbtxt文件。这里遇见一个问题,源码的utils/object_detection_evaluation.py文件中的第433行有个判断
if groundtruth_is_difficult_list is None:但是没有判断groundtruth_is_difficult_list 是空的情况,导致groundtruth_is_difficult_list变量没有赋值,后续在utils/per_image_evaluation.py的202行调用的时候会报错,没继续研究下去到底是我配置的问题还是源码这里应该判断一次,我把utils/object_detection_evaluation.py的433行改成了:
if groundtruth_is_difficult_list is None or len(groundtruth_is_difficult_list) == 0:再运行就没问题了。运行后在$PATH/eval_dir目录下会生成out文件,然后打开tensorboard
tensorboard -log $PATH/eval_dir/在浏览器中查看
可以看到已经把结果显示出来的,之所以有偏差是因为还在训练当中,train.py还没收敛到一个满意的loss,还有就是训练集数据量太少,需要加大样本量,不过已经看到结果了。对于我的图片搜索来说,已经可以区分出图片中是否有人脸啦!
五、模型导出
如同上面说的,直接使用
python export_inference_graph.py --input_type image_tensor --pipeline_config_path $PATH/config/ttadm_ssd_inception.config --trained_checkpoint_prefix $PATH/model_ckpt/model.ckpt-* --output_directory $PATH/exported_model_directory
exporter.py 71行会报错,
rewrite_options = rewriter_config_pb2.RewriterConfig(optimize_tensor_layout=True)
修改为:
rewrite_options = rewriter_config_pb2.RewriterConfig()
六、执行预测
import cv2 import numpy as np import tensorflow as tf from object_detection.utils import label_map_util from object_detection.utils import visualization_utils as vis_util class TOD(object): def __init__(self): self.PATH_TO_CKPT = r'$PATH/exported_model_directory/frozen_inference_graph.pb' self.PATH_TO_LABELS = r'$PATH/eval_label_map.pbtxt' self.NUM_CLASSES = 1 self.detection_graph = self._load_model() self.category_index = self._load_label_map() def _load_model(self): detection_graph = tf.Graph() with detection_graph.as_default(): od_graph_def = tf.GraphDef() with tf.gfile.GFile(self.PATH_TO_CKPT, 'rb') as fid: serialized_graph = fid.read() od_graph_def.ParseFromString(serialized_graph) tf.import_graph_def(od_graph_def, name='') return detection_graph def _load_label_map(self): label_map = label_map_util.load_labelmap(self.PATH_TO_LABELS) categories = label_map_util.convert_label_map_to_categories(label_map, max_num_classes=self.NUM_CLASSES, use_display_name=True) category_index = label_map_util.create_category_index(categories) return category_index def detect(self, image): with self.detection_graph.as_default(): with tf.Session(graph=self.detection_graph) as sess: # Expand dimensions since the model expects images to have shape: [1, None, None, 3] image_np_expanded = np.expand_dims(image, axis=0) image_tensor = self.detection_graph.get_tensor_by_name('image_tensor:0') boxes = self.detection_graph.get_tensor_by_name('detection_boxes:0') scores = self.detection_graph.get_tensor_by_name('detection_scores:0') classes = self.detection_graph.get_tensor_by_name('detection_classes:0') num_detections = self.detection_graph.get_tensor_by_name('num_detections:0') # Actual detection. (boxes, scores, classes, num_detections) = sess.run( [boxes, scores, classes, num_detections], feed_dict={image_tensor: image_np_expanded}) # Visualization of the results of a detection. vis_util.visualize_boxes_and_labels_on_image_array( image, np.squeeze(boxes), np.squeeze(classes).astype(np.int32), np.squeeze(scores), self.category_index, use_normalized_coordinates=True, line_thickness=8) cv2.namedWindow("detection", cv2.WINDOW_NORMAL) cv2.imshow("detection", image) cv2.waitKey(0) def main(): image = cv2.imread('1.jpg') detecotr = TOD() detecotr.detect(image) pass if __name__ == '__main__': main()