实验需要判断人脸是否戴了眼镜，所以参考网上的文章弄了个简单的CNN图片分类器来做戴眼镜与否的判定。
环境如下：
- macOS 10.13.2
- Python 2.7
- TensorFlow 1.2.0
数据集：
要训练我们当然需要训练集，这里我采用的是CelebA的人脸图像数据集，从中筛选出戴了眼镜的人脸和没戴眼镜的人脸分别一千多张也就够了，如何筛选CelebA人脸数据集可以参考我这篇博客：处理筛选CelebA人脸数据集
将两个分别装有戴眼镜与否的人脸图片的文件夹放到我们工程目录下，然后开始写代码。
训练代码：
train_glass.py
# -*- coding: utf-8 -*-
from skimage import io,transform
import glob
import os
import tensorflow as tf
import numpy as np
import time
#数据集地址
path='./'
#模型保存地址
model_path='./model/model.ckpt'
#将所有的图片resize成100*100
w=100
h=100
c=3
#读取图片
def read_img(path):
cate=[path+x for x in os.listdir(path) if os.path.isdir(path+x)]
imgs=[]
labels=[]
for idx,folder in enumerate(cate):
for im in glob.glob(folder+'/*.jpg'):
print('reading the images:%s'%(im))
img=io.imread(im)
img=transform.resize(img,(w,h))
imgs.append(img)
labels.append(idx)
return np.asarray(imgs,np.float32),np.asarray(labels,np.int32)
data,label=read_img(path)
#打乱顺序
num_example=data.shape[0]
arr=np.arange(num_example)
np.random.shuffle(arr)
data=data[arr]
label=label[arr]
#将所有数据分为训练集和验证集
ratio=0.8
s=np.int(num_example*ratio)
x_train=data[:s]
y_train=label[:s]
x_val=data[s:]
y_val=label[s:]
#-----------------构建网络----------------------
#占位符
x=tf.placeholder(tf.float32,shape=[None,w,h,c],name='x')
y_=tf.placeholder(tf.int32,shape=[None,],name='y_')
def inference(input_tensor, train, regularizer):
with tf.variable_scope('layer1-conv1'):
conv1_weights = tf.get_variable("weight",[5,5,3,32],initializer=tf.truncated_normal_initializer(stddev=0.1))
conv1_biases = tf.get_variable("bias", [32], initializer=tf.constant_initializer(0.0))
conv1 = tf.nn.conv2d(input_tensor, conv1_weights, strides=[1, 1, 1, 1], padding='SAME')
relu1 = tf.nn.relu(tf.nn.bias_add(conv1, conv1_biases))
with tf.name_scope("layer2-pool1"):
pool1 = tf.nn.max_pool(relu1, ksize = [1,2,2,1],strides=[1,2,2,1],padding="VALID")
with tf.variable_scope("layer3-conv2"):
conv2_weights = tf.get_variable("weight",[5,5,32,64],initializer=tf.truncated_normal_initializer(stddev=0.1))
conv2_biases = tf.get_variable("bias", [64], initializer=tf.constant_initializer(0.0))
conv2 = tf.nn.conv2d(pool1, conv2_weights, strides=[1, 1, 1, 1], padding='SAME')
relu2 = tf.nn.relu(tf.nn.bias_add(conv2, conv2_biases))
with tf.name_scope("layer4-pool2"):
pool2 = tf.nn.max_pool(relu2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
with tf.variable_scope("layer5-conv3"):
conv3_weights = tf.get_variable("weight",[3,3,64,128],initializer=tf.truncated_normal_initializer(stddev=0.1))
conv3_biases = tf.get_variable("bias", [128], initializer=tf.constant_initializer(0.0))
conv3 = tf.nn.conv2d(pool2, conv3_weights, strides=[1, 1, 1, 1], padding='SAME')
relu3 = tf.nn.relu(tf.nn.bias_add(conv3, conv3_biases))
with tf.name_scope("layer6-pool3"):
pool3 = tf.nn.max_pool(relu3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
with tf.variable_scope("layer7-conv4"):
conv4_weights = tf.get_variable("weight",[3,3,128,128],initializer=tf.truncated_normal_initializer(stddev=0.1))
conv4_biases = tf.get_variable("bias", [128], initializer=tf.constant_initializer(0.0))
conv4 = tf.nn.conv2d(pool3, conv4_weights, strides=[1, 1, 1, 1], padding='SAME')
relu4 = tf.nn.relu(tf.nn.bias_add(conv4, conv4_biases))
with tf.name_scope("layer8-pool4"):
pool4 = tf.nn.max_pool(relu4, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
nodes = 6*6*128
reshaped = tf.reshape(pool4,[-1,nodes])
with tf.variable_scope('layer9-fc1'):
fc1_weights = tf.get_variable("weight", [nodes, 1024],
initializer=tf.truncated_normal_initializer(stddev=0.1))
if regularizer != None: tf.add_to_collection('losses', regularizer(fc1_weights))
fc1_biases = tf.get_variable("bias", [1024], initializer=tf.constant_initializer(0.1))
fc1 = tf.nn.relu(tf.matmul(reshaped, fc1_weights) + fc1_biases)
if train: fc1 = tf.nn.dropout(fc1, 0.5)
with tf.variable_scope('layer10-fc2'):
fc2_weights = tf.get_variable("weight", [1024, 512],
initializer=tf.truncated_normal_initializer(stddev=0.1))
if regularizer != None: tf.add_to_collection('losses', regularizer(fc2_weights))
fc2_biases = tf.get_variable("bias", [512], initializer=tf.constant_initializer(0.1))
fc2 = tf.nn.relu(tf.matmul(fc1, fc2_weights) + fc2_biases)
if train: fc2 = tf.nn.dropout(fc2, 0.5)
with tf.variable_scope('layer11-fc3'):
fc3_weights = tf.get_variable("weight", [512, 2],
initializer=tf.truncated_normal_initializer(stddev=0.1))
if regularizer != None: tf.add_to_collection('losses', regularizer(fc3_weights))
fc3_biases = tf.get_variable("bias", [2], initializer=tf.constant_initializer(0.1))
logit = tf.matmul(fc2, fc3_weights) + fc3_biases
return logit
#---------------------------网络结束---------------------------
regularizer = tf.contrib.layers.l2_regularizer(0.0001)
logits = inference(x,False,regularizer)
#(小处理)将logits乘以1赋值给logits_eval，定义name，方便在后续调用模型时通过tensor名字调用输出tensor
b = tf.constant(value=1,dtype=tf.float32)
logits_eval = tf.multiply(logits,b,name='logits_eval')
loss=tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=y_)
train_op=tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)
correct_prediction = tf.equal(tf.cast(tf.argmax(logits,1),tf.int32), y_)
acc= tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
#定义一个函数，按批次取数据
def minibatches(inputs=None, targets=None, batch_size=None, shuffle=False):
assert len(inputs) == len(targets)
if shuffle:
indices = np.arange(len(inputs))
np.random.shuffle(indices)
for start_idx in range(0, len(inputs) - batch_size + 1, batch_size):
if shuffle:
excerpt = indices[start_idx:start_idx + batch_size]
else:
excerpt = slice(start_idx, start_idx + batch_size)
yield inputs[excerpt], targets[excerpt]
#训练和测试数据，可将n_epoch设置更大一些
n_epoch=10
batch_size=64
saver=tf.train.Saver()
sess=tf.Session()
sess.run(tf.global_variables_initializer())
for epoch in range(n_epoch):
start_time = time.time()
print("====epoch %d====="%epoch)
#training
train_loss, train_acc, n_batch = 0, 0, 0
for x_train_a, y_train_a in minibatches(x_train, y_train, batch_size, shuffle=True):
_,err,ac=sess.run([train_op,loss,acc], feed_dict={x: x_train_a, y_: y_train_a})
train_loss += err; train_acc += ac; n_batch += 1
print(" train loss: %f" % (np.sum(train_loss)/ n_batch))
print(" train acc: %f" % (np.sum(train_acc)/ n_batch))
#validation
val_loss, val_acc, n_batch = 0, 0, 0
for x_val_a, y_val_a in minibatches(x_val, y_val, batch_size, shuffle=False):
err, ac = sess.run([loss,acc], feed_dict={x: x_val_a, y_: y_val_a})
val_loss += err; val_acc += ac; n_batch += 1
print(" validation loss: %f" % (np.sum(val_loss)/ n_batch))
print(" validation acc: %f" % (np.sum(val_acc)/ n_batch))
saver.save(sess,model_path)
sess.close()
代码一开始获取了数据集，因为有两个戴眼镜与否的目录，所以这份代码不需要具体写所有目录，代码会识别有两个目录，也就表示图片有两个类别——戴眼镜与不戴眼镜。
生成的模型文件我们保存在model文件夹下。
代码将80%的图片作为训练集，剩下20%的图片作为测试集，来查看训练效果。
其余部分代码中的注释讲的很清楚了，现在可以直接在终端运行这个python文件开始训练了。
一开始会读取所有图片，然后进行训练，训练有十轮，轮数可以通过修改“n_epoch”变量来改变，但是十轮下来效果已经很好了。用mac跑半小时也就训练完了。
第一轮的损失及准确率情况：
====epoch 0=====
train loss: 1327.149698
train acc: 0.575101
validation loss: 25.071056
validation acc: 0.845982
很正常，毕竟是二分类，瞎蒙也有50%的准确率嘛
第十轮的损失及准确率情况：
====epoch 10=====
train loss: 0.066756
train acc: 1.000000
validation loss: 1.637226
validation acc: 0.991071
可以看到，第十轮的训练损失已经非常小了，只有0.066756，而准确率居然到了100%。。。测试集的准确率也到了99.1%。要说明的是这些值每次都可能不一样，只是参考。
现在可以开始测试了。
测试代码
inference_glass.py
# -*- coding: utf-8 -*-
from skimage import io,transform
import tensorflow as tf
import numpy as np
path1 = "./face范冰冰眼镜.jpg"
path2 = "./face黄晓明眼镜.jpg"
path3 = "./face林志玲眼镜.jpg"
path4 = "./face徐峥无眼镜.jpg"
path5 = "./face赵薇无眼镜.jpg"
face_dict = {1:'Has Glass',0:'No Glass'}
w=100
h=100
c=3
def read_one_image(path):
img = io.imread(path)
img = transform.resize(img,(w,h))
return np.asarray(img)
with tf.Session() as sess:
data = []
data1 = read_one_image(path1)
data2 = read_one_image(path2)
data3 = read_one_image(path3)
data4 = read_one_image(path4)
data5 = read_one_image(path5)
data.append(data1)
data.append(data2)
data.append(data3)
data.append(data4)
data.append(data5)
saver = tf.train.import_meta_graph('./model/model.ckpt.meta')
saver.restore(sess,tf.train.latest_checkpoint('./model/'))
graph = tf.get_default_graph()
x = graph.get_tensor_by_name("x:0")
feed_dict = {x:data}
logits = graph.get_tensor_by_name("logits_eval:0")
classification_result = sess.run(logits,feed_dict)
#打印出预测矩阵
print(classification_result)
#打印出预测矩阵每一行最大值的索引
print(tf.argmax(classification_result,1).eval())
#根据索引通过字典对应人脸的分类
output = []
output = tf.argmax(classification_result,1).eval()
for i in range(len(output)):
print("No.",i+1,"face is belong to:"+face_dict[output[i]])
这里我们放入五张人脸图片作为测试数据，注意，这里的测试与上文训练代码中的测试不是一个意思，这里是真正用来做分类，上文的测试只是用来检验训练结果。
五张测试用的明星脸
“face_dict”数组保存了分类结果的说明，训练后如果分类为1表示戴了眼镜，如果分类为0表示没戴眼镜，最后的代码也可以看到结果输出是用这个数组来转换结果的。
测试结果如下：
[[ -9.42931938 11.11681461]
[ -3.12095881 3.74463916]
[ -4.20803499 8.95589638]
[ 4.09471083 -0.16824517]
[ 6.93871641 -2.67873168]]
[1 1 1 0 0]
('No.', 1, 'face is belong to:Has Glass')
('No.', 2, 'face is belong to:Has Glass')
('No.', 3, 'face is belong to:Has Glass')
('No.', 4, 'face is belong to:No Glass')
('No.', 5, 'face is belong to:No Glass')
可以对比我们的输入图片名来看，发现结果全部正确！
我的工程：https://github.com/Cloudox/CNN_Face_Glass_Classfy
参考文章：http://blog.csdn.net/Enchanted_ZhouH/article/details/74116823
