随笔- 240 文章- 0 评论- 4 阅读- 23万
08-人脸识别-FaceNet-classify.py代码阅读（说明见注释）

100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
"""An example of how to use your own dataset to train a classifier that recognizes people.
"""
# MIT License
# 
# Copyright (c) 2016 David Sandberg
# 
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
# 
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
# 
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
 
# @ 调用格式：
# @
# @ 训练模型记住人脸（不是训练网络，网络在这之前已经先训练好了）。
# @ ../lfw/ 是lfw数据集经过 mtcnn 截取以后的结果。否则会影响效果（去除数据集中的人脸外部干扰）
# @     python classifier.py TRAIN ../lfw/ 20170511-185253/ train_20180419_2048.pkl 
# @
# @ 测试模型记住人脸的结果。（../data 是测试用的图的路径。）
# @     python classifier.py CLASSIFY ../data/ 20170511-185253/ train_20180419_2048.pkl 
 
 
 
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
 
import tensorflow as tf
import numpy as np
import argparse
import facenet
import os
import sys
import math
import pickle
from sklearn.svm import SVC
 
# @ args内中参数见函数 parse_arguments
def main(args):
    # @ 声明一个计算图，都这么写，没有就是默认一个。
    with tf.Graph().as_default():
        # @ 声明一个 Session
        with tf.Session() as sess:
         
            # @ Part I
            # @ 这部分是计算人脸的 embedding 特征。费时。 
            # @ 
             
            # @ 加随机数seed，调用np.random.random()的结果都会相同。
            np.random.seed(seed=args.seed)
             
            if args.use_split_dataset:
                dataset_tmp = facenet.get_dataset(args.data_dir)
                train_set, test_set = split_dataset(dataset_tmp, args.min_nrof_images_per_class, args.nrof_train_images_per_class)
                if (args.mode=='TRAIN'):
                    dataset = train_set
                elif (args.mode=='CLASSIFY'):
                    dataset = test_set
            else:
                dataset = facenet.get_dataset(args.data_dir)
 
            # Check that there are at least one training image per class
            # @ cls.image_paths 是每张图的路径，包含文件名。
            for cls in dataset:
                assert(len(cls.image_paths)>0, 'There must be at least one image for each class in the dataset')            
 
            # @ 分离出图片路径名paths，和类型labels（人脸所属人名）
            paths, labels = facenet.get_image_paths_and_labels(dataset)
             
            print('Number of classes: %d' % len(dataset))
            print('Number of images: %d' % len(paths))
             
            # Load the model
            # @ 这里加的 model 使用于生成人脸的 embedding 特征的网络。
            # @ 这个网络是事先已经生成好的。
            # @ 网络可以根据运行的平台，设计成不同大小。比如基于GoogleNet/AlexNet等
            print('Loading feature extraction model')
            facenet.load_model(args.model)
             
            # Get input and output tensors
            # @ TensorFlow的参数准备。embeddings 是网络的输出，是后续分类的输入。
            images_placeholder = tf.get_default_graph().get_tensor_by_name("input:0")
            embeddings = tf.get_default_graph().get_tensor_by_name("embeddings:0")
            phase_train_placeholder = tf.get_default_graph().get_tensor_by_name("phase_train:0")
            embedding_size = embeddings.get_shape()[1]
             
            # Run forward pass to calculate embeddings
            print('Calculating features for images')
            nrof_images = len(paths) # @ 图片总数
            nrof_batches_per_epoch = int(math.ceil(1.0*nrof_images / args.batch_size))
            emb_array = np.zeros((nrof_images, embedding_size))
            for i in range(nrof_batches_per_epoch):
                start_index = i*args.batch_size
                end_index = min((i+1)*args.batch_size, nrof_images)
                paths_batch = paths[start_index:end_index]
                images = facenet.load_data(paths_batch, False, False, args.image_size)
                feed_dict = { images_placeholder:images, phase_train_placeholder:False }
                emb_array[start_index:end_index,:] = sess.run(embeddings, feed_dict=feed_dict)
             
            # @ emb_array 是 embedding 结果。一个 embedding 有 18 维。
            # @ 接下来就是用机器学习的方法分类。
            classifier_filename_exp = os.path.expanduser(args.classifier_filename)
 
            # @ Part II 也较费时。
            # @ 这部分是训练分类人脸的机器学习模型，这里使用的SVC，是SVM的一种。
            # @ 若是 CLASSIFY ，则是加载训练结果，建立 SVC 分类器。
             
            if (args.mode=='TRAIN'):
                # Train classifier
                # @ SVC是SVM的一种Type，是用来的做分类的；同样还有SVR，是SVM的另一种Type，是用来的做回归的。
                print('Training classifier')
                model = SVC(kernel='linear', probability=True)
                model.fit(emb_array, labels) # @ 训练过程
             
                # @ 训练结束，保存数据
                # Create a list of class names
                class_names = [ cls.name.replace('_', ' ') for cls in dataset]
 
                # Saving classifier model
                with open(classifier_filename_exp, 'wb') as outfile:
                    pickle.dump((model, class_names), outfile)
                print('Saved classifier model to file "%s"' % classifier_filename_exp)
                 
            elif (args.mode=='CLASSIFY'):
                # Classify images
                print('Testing classifier')
                # @ 加载数据，建立分类器
                with open(classifier_filename_exp, 'rb') as infile:
                    (model, class_names) = pickle.load(infile)
 
                print('Loaded classifier model from file "%s"' % classifier_filename_exp)
 
                # @ 预测，标签结果应该是 one_hot 的。
                predictions = model.predict_proba(emb_array)
                best_class_indices = np.argmax(predictions, axis=1) # @ 输出每列最大的序号。
                best_class_probabilities = predictions[np.arange(len(best_class_indices)), best_class_indices]
                 
                for i in range(len(best_class_indices)):
                    print('%4d  %s: %.3f' % (i, class_names[best_class_indices[i]], best_class_probabilities[i]))
                     
                # @ 评估结果。labels 是测试集的实际结果，best_class_indices是预测结果。
                accuracy = np.mean(np.equal(best_class_indices, labels))
                print('Accuracy: %.3f' % accuracy)
                 
# @ 将数据集分成训练集和测试集
def split_dataset(dataset, min_nrof_images_per_class, nrof_train_images_per_class):
    train_set = []
    test_set = []
    for cls in dataset:
        paths = cls.image_paths
        # Remove classes with less than min_nrof_images_per_class
        if len(paths)>=min_nrof_images_per_class:
            np.random.shuffle(paths)
            train_set.append(facenet.ImageClass(cls.name, paths[:nrof_train_images_per_class]))
            test_set.append(facenet.ImageClass(cls.name, paths[nrof_train_images_per_class:]))
    return train_set, test_set
     
# @ 命令行参数，使用的系统库 argparse
# @ ** 写法值得记住 **
def parse_arguments(argv):
    parser = argparse.ArgumentParser()
     
    parser.add_argument('mode', type=str, choices=['TRAIN', 'CLASSIFY'],
        help='Indicates if a new classifier should be trained or a classification ' +
        'model should be used for classification', default='CLASSIFY')
    parser.add_argument('data_dir', type=str,
        help='Path to the data directory containing aligned LFW face patches.')
    parser.add_argument('model', type=str, 
        help='Could be either a directory containing the meta_file and ckpt_file or a model protobuf (.pb) file')
    parser.add_argument('classifier_filename', 
        help='Classifier model file name as a pickle (.pkl) file. ' +
        'For training this is the output and for classification this is an input.')
    parser.add_argument('--use_split_dataset', 
        help='Indicates that the dataset specified by data_dir should be split into a training and test set. ' + 
        'Otherwise a separate test set can be specified using the test_data_dir option.', action='store_true')
    parser.add_argument('--test_data_dir', type=str,
        help='Path to the test data directory containing aligned images used for testing.')
    parser.add_argument('--batch_size', type=int,
        help='Number of images to process in a batch.', default=90)
    parser.add_argument('--image_size', type=int,
        help='Image size (height, width) in pixels.', default=160)
    parser.add_argument('--seed', type=int,
        help='Random seed.', default=666)
    parser.add_argument('--min_nrof_images_per_class', type=int,
        help='Only include classes with at least this number of images in the dataset', default=20)
    parser.add_argument('--nrof_train_images_per_class', type=int,
        help='Use this number of images from each class for training and the rest for testing', default=10)
     
    return parser.parse_args(argv)
 
# @ 主函数
# @ sys.argv[1:] 就是命令行输入的 classify.py 后面的所有字符串，以空格分隔。
if __name__ == '__main__':
    main(parse_arguments(sys.argv[1:]))