深度学习实战 | 使用Kera预测人物年龄

正文共：11452个字，8张图，预计阅读时间：29分钟。

01

问题描述

我们的任务是从一个人的面部特征来预测他的年龄(用“Young”“Middle ”“Old”表示)，我们训练的数据集大约有19906多张照片及其每张图片对应的年龄（全是阿三的头像。。。），测试集有6636张图片，首先我们加载数据集，然后我们通过深度学习框架Keras建立、编译、训练模型，预测出6636张人物头像对应的年龄。

02

引入所需要的模块

import os

import random

import pandas as pd

import numpy as np

from PIL import Image

03

加载数据集

root_dir=os.path.abspath('E:/data/age') train=pd.read_csv(os.path.join(root_dir,'train.csv')) test=pd.read_csv(os.path.join(root_dir,'test.csv')) print(train.head()) print(test.head())  

ID   Class

0    377.jpg  MIDDLE

1  17814.jpg   YOUNG

2  21283.jpg  MIDDLE

3  16496.jpg   YOUNG

4   4487.jpg  MIDDLE

         ID

0  25321.jpg

1    989.jpg

2  19277.jpg

3  13093.jpg

4   5367.jpg

04

随机读取一张图片试下

i=random.choice(train.index) img_name=train.ID[i] print(img_name) img=Image.open(os.path.join(root_dir,'Train',img_name)) img.show() print(train.Class[i])

20188.jpg

MIDDLE

05

难点

我们随机打开几张图片之后，可以发现图片之间的差别比较大。大家感受下：

质量好的图片：

Middle:

*Middle**

Young:

**Young**

Old:

*Old**

质量差的：

Middle:

**Middle**

下面是我们需要面临的问题：

1、图片的尺寸差别：有的图片的尺寸是66x46,而另一张图片尺寸为102x87

2、人物面貌角度不同：

侧脸：

正脸：

3、图片质量不一（直接上图）:

插图

4、亮度和对比度的差异

亮度

对比度

现在，我们只专注下图片尺寸处理，将每一张图片尺寸重置为32x32；

06

格式化图片尺寸和将图片转换成numpy数组

temp=[]for img_name in train.ID:    img_path=os.path.join(root_dir,'Train',img_name)    img=Image.open(img_path)    img=img.resize((32,32))    array=np.array(img)    temp.append(array.astype('float32')) train_x=np.stack(temp) print(train_x.shape) print(train_x.ndim)(19906, 32, 32, 3) 4temp=[]for img_name in test.ID:    img_path=os.path.join(root_dir,'Test',img_name)    img=Image.open(img_path)    img=img.resize((32,32))    array=np.array(img)    temp.append(array.astype('float32')) test_x=np.stack(temp) print(test_x.shape)(6636, 32, 32, 3)

另外我们再归一化图像，这样会使模型训练的更快

train_x = train_x / 255.test_x = test_x / 255.

我们看下图片年龄大致分布：

train.Class.value_counts(normalize=True)

MIDDLE    0.542751

YOUNG     0.336883

OLD       0.120366

Name: Class, dtype: float64

test['Class'] = 'MIDDLE

'test.to_csv('sub01.csv', index=False)

将目标变量处理虚拟列，能够使模型更容易接受识别它

import keras

from sklearn.preprocessing import LabelEncoder

lb=LabelEncoder() train_y=lb.fit_transform(train.Class) print(train_y) train_y=keras.utils.np_utils.to_categorical(train_y) print(train_y) print(train_y.shape)

[0 2 0 ..., 0 0 0]

[[ 1.  0.  0.] [ 0.  0.  1.] [ 1.  0.  0.] ..., [ 1.  0.  0.] [ 1.  0.  0.] [ 1.  0.  0.]] (19906, 3)

07

创建模型

#构建神经网络

input_num_units=(32,32,3) hidden_num_units=500

output_num_units=3

epochs=5

batch_size=128

from keras.models import Sequential

from keras.layers import Dense,Flatten,InputLayer model=Sequential({    InputLayer(input_shape=input_num_units),    Flatten(),    Dense(units=hidden_num_units,activation='relu'),    Dense(input_shape=(32,32,3),units=output_num_units,activation='softmax') }) model.summary()

_________________________________________________________________

 Layer (type)                 Output Shape              Param #  

========================================

 input_23 (InputLayer)        (None, 32, 32, 3)         0        

_________________________________________________________________

 flatten_23 (Flatten)         (None, 3072)              0        

_________________________________________________________________

dense_45 (Dense)             (None, 500)               1536500  

_________________________________________________________________ 

dense_46 (Dense)             (None, 3)                 1503      

======================================== 

Total params: 1,538,003

Trainable params: 1,538,003

Non-trainable params: 0

_________________________________________________________________

08

编译模型

# model.compile(optimizer='sgd',loss='categorical_crossentropy',metrics['accuracy'])

model.compile(optimizer='sgd',loss='categorical_crossentropy', metrics=['accuracy']) model.fit(train_x,train_y,batch_size=batch_size,epochs=epochs,verbose=1)

Epoch 1/5

19906/19906 [==============================]

- 4s - loss: 0.8878 - acc: 0.5809     Epoch 2/5 19906/19906 [==============================]

- 4s - loss: 0.8420 - acc: 0.6077     Epoch 3/5 19906/19906 [==============================]

- 4s - loss: 0.8210 - acc: 0.6214     Epoch 4/5 19906/19906 [==============================]

- 4s - loss: 0.8149 - acc: 0.6194     Epoch 5/5 19906/19906 [==============================]

- 4s - loss: 0.8042 - acc: 0.6305    

<keras.callbacks.History at 0x1d3803e6278>

model.fit(train_x, train_y, batch_size=batch_size,epochs=epochs,verbose=1, validation_split=0.2)

Train on 15924 samples, validate on 3982 samples Epoch 1/5 15924/15924 [==============================]

- 3s - loss: 0.7970 - acc: 0.6375 - val_loss: 0.7854 - val_acc: 0.6396 Epoch 2/5 15924/15924 [==============================]

- 3s - loss: 0.7919 - acc: 0.6378 - val_loss: 0.7767 - val_acc: 0.6519 Epoch 3/5 15924/15924 [==============================]

- 3s - loss: 0.7870 - acc: 0.6404 - val_loss: 0.7754 - val_acc: 0.6534 Epoch 4/5 15924/15924 [==============================]

- 3s - loss: 0.7806 - acc: 0.6439 - val_loss: 0.7715 - val_acc: 0.6524 Epoch 5/5 15924/15924 [==============================]

- 3s - loss: 0.7755 - acc: 0.6519 - val_loss: 0.7970 - val_acc: 0.6346

<keras.callbacks.History at 0x1d3800a4eb8>

09

优化

我们使用最基本的模型来处理这个年龄预测结果，并且最终的预测结果为0.6375。接下来，从以下角度尝试优化：

1、使用更好的神经网络模型

2、增加训练次数

3、将图片进行灰度处理（因为对于本问题而言，图片颜色不是一个特别重要的特征。）

10

optimize1 使用卷积神经网络

添加卷积层之后，预测准确率有所上涨，从6.3到6.7；最开始epochs轮数是5，训练轮数增加到10，此时准确率为6.87；然后将训练轮数增加到20，结果没有发生变化。

11

Conv2D层

keras.layers.convolutional.Conv2D(filters, kernel_size, strides=(1, 1), padding='valid', data_format=None, dilation_rate=(1, 1), activation=None, use_bias=True, kernel_initializer='glorot_uniform', bias_initializer='zeros', kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None)

• filters:输出的维度

• strides:卷积的步长

更多关于Conv2D的介绍请看Keras文档Conv2D层（http://keras-cn.readthedocs.io/en/latest/layers/convolutional_layer/#conv2d）

#参数初始化

filters=10

filtersize=(5,5) epochs =10

batchsize=128

input_shape=(32,32,3)

from keras.models import Sequential model = Sequential() model.add(keras.layers.InputLayer(input_shape=input_shape)) model.add(keras.layers.convolutional.Conv2D(filters, filtersize, strides=(1, 1), padding='valid', data_format="channels_last", activation='relu')) model.add(keras.layers.MaxPooling2D(pool_size=(2, 2))) model.add(keras.layers.Flatten()) model.add(keras.layers.Dense(units=3, input_dim=50,activation='softmax')) model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy']) model.fit(train_x, train_y, epochs=epochs, batch_size=batchsize,validation_split=0.3) model.summary()

Train on 13934 samples, validate on 5972 samples Epoch 1/1

013934/13934 [==============================] - 9s - loss: 0.8986 - acc: 0.5884 - val_loss: 0.8352 - val_acc: 0.6271

Epoch 2/1

013934/13934 [==============================] - 9s - loss: 0.8141 - acc: 0.6281 - val_loss: 0.7886 - val_acc: 0.6474

Epoch 3/1

013934/13934 [==============================] - 9s - loss: 0.7788 - acc: 0.6504 - val_loss: 0.7706 - val_acc: 0.6551

Epoch 4/1

013934/13934 [==============================] - 9s - loss: 0.7638 - acc: 0.6577 - val_loss: 0.7559 - val_acc: 0.6626

Epoch 5/1

013934/13934 [==============================] - 9s - loss: 0.7484 - acc: 0.6679 - val_loss: 0.7457 - val_acc: 0.6710

Epoch 6/1

013934/13934 [==============================] - 9s - loss: 0.7346 - acc: 0.6723 - val_loss: 0.7490 - val_acc: 0.6780

Epoch 7/1

013934/13934 [==============================] - 9s - loss: 0.7217 - acc: 0.6804 - val_loss: 0.7298 - val_acc: 0.6795

Epoch 8/1

013934/13934 [==============================] - 9s - loss: 0.7162 - acc: 0.6826 - val_loss: 0.7248 - val_acc: 0.6792

Epoch 9/1

013934/13934 [==============================] - 9s - loss: 0.7082 - acc: 0.6892 - val_loss: 0.7202 - val_acc: 0.6890

Epoch 10/1

013934/13934 [==============================] - 9s - loss: 0.7001 - acc: 0.6940 - val_loss: 0.7226 - val_acc: 0.6885

_________________________________________________________________

 Layer (type)                 Output Shape              Param #  

========================================

 input_6 (InputLayer)         (None, 32, 32, 3)         0        

_______________________________________________________________ 

conv2d_6 (Conv2D)            (None, 28, 28, 10)        760    

  _______________________________________________________________ 

max_pooling2d_6 (MaxPooling2 (None, 14, 14, 10)        0    

_______________________________________________________________ 

flatten_6 (Flatten)          (None, 1960)              0        

_________________________________________________________________ 

dense_6 (Dense)              (None, 3)                 5883      

========================================

Total params: 6,643

Trainable params: 6,643

Non-trainable params: 0

_________________________________________________________________

12

optimize2 增加神经网络的层数

我们在模型中多添加几层并且提高卷几层的输出维度，这次结果得到显著提升：0.750904

#参数初始化

filters1=50

filters2=100

filters3=100

filtersize=(5,5) epochs =10

batchsize=128

input_shape=(32,32,3)

from keras.models import Sequential

model = Sequential() model.add(keras.layers.InputLayer(input_shape=input_shape)) model.add(keras.layers.convolutional.Conv2D(filters1, filtersize, strides=(1, 1), padding='valid', data_format="channels_last", activation='relu'))

 model.add(keras.layers.MaxPooling2D(pool_size=(2, 2))) 

 model.add(keras.layers.convolutional.Conv2D(filters2, filtersize, strides=(1, 1), padding='valid', data_format="channels_last", activation='relu')) 

model.add(keras.layers.MaxPooling2D(pool_size=(2, 2))) model.add(keras.layers.convolutional.Conv2D(filters3, filtersize, strides=(1, 1), padding='valid', data_format="channels_last", activation='relu')) 

model.add(keras.layers.Flatten()) 

 model.add(keras.layers.Dense(units=3, input_dim=50,activation='softmax')) model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy']) model.fit(train_x, train_y, epochs=epochs, batch_size=batchsize,validation_split=0.3) model.summary()

Train on 13934 samples, validate on 5972 samples

Epoch 1/1

013934/13934 [==============================] - 44s - loss: 0.8613 - acc: 0.5985 - val_loss: 0.7778 - val_acc: 0.6586

Epoch 2/1

013934/13934 [==============================] - 44s - loss: 0.7493 - acc: 0.6697 - val_loss: 0.7545 - val_acc: 0.6808

Epoch 3/1

013934/13934 [==============================] - 43s - loss: 0.7079 - acc: 0.6877 - val_loss: 0.7150 - val_acc: 0.6947

Epoch 4/1

013934/13934 [==============================] - 43s - loss: 0.6694 - acc: 0.7061 - val_loss: 0.6496 - val_acc: 0.7261

Epoch 5/1

013934/13934 [==============================] - 43s - loss: 0.6274 - acc: 0.7295 - val_loss: 0.6683 - val_acc: 0.7125

Epoch 6/1

013934/13934 [==============================] - 43s - loss: 0.5950 - acc: 0.7462 - val_loss: 0.6194 - val_acc: 0.7400

Epoch 7/1

013934/13934 [==============================] - 43s - loss: 0.5562 - acc: 0.7655 - val_loss: 0.5981 - val_acc: 0.7465

Epoch 8/1

013934/13934 [==============================] - 43s - loss: 0.5165 - acc: 0.7852 - val_loss: 0.6458 - val_acc: 0.7354

Epoch 9/1

013934/13934 [==============================] - 46s - loss: 0.4826 - acc: 0.7986 - val_loss: 0.6206 - val_acc: 0.7467

Epoch 10/1

013934/13934 [==============================] - 45s - loss: 0.4530 - acc: 0.8130 - val_loss: 0.5984 - val_acc: 0.7569

_________________________________________________________________ Layer (type)                 Output Shape              Param #  

========================================== input_15 (InputLayer)        (None, 32, 32, 3)         0        

_________________________________________________________________ conv2d_31 (Conv2D)           (None, 28, 28, 50)        3800      

_________________________________________________________________ max_pooling2d_23 (MaxPooling (None, 14, 14, 50)        0        

_________________________________________________________________ conv2d_32 (Conv2D)           (None, 10, 10, 100)       125100    

_________________________________________________________________ max_pooling2d_24 (MaxPooling (None, 5, 5, 100)         0        

_________________________________________________________________ conv2d_33 (Conv2D)           (None, 1, 1, 100)         250100    

_________________________________________________________________ flatten_15 (Flatten)         (None, 100)               0        

_________________________________________________________________ dense_7 (Dense)              (None, 3)                 303      

========================================== Total params: 379,303

Trainable params: 379,303

Non-trainable params: 0

_________________________________________________________________

13

输出结果

pred=model.predict_classes(test_x) pred=lb.inverse_transform(pred) print(pred) test['Class']=pred test.to_csv('sub02.csv',index=False)

6636/6636 [==============================] - 7s     ['MIDDLE' 'YOUNG' 'MIDDLE' ..., 'MIDDLE' 'MIDDLE' 'YOUNG']

i = random.choice(train.index) img_name = train.ID[i] img=Image.open(os.path.join(root_dir,'Train',img_name)) img.show() pred = model.predict_classes(train_x) print('Original:', train.Class[i], 'Predicted:', lb.inverse_transform(pred[i]))

19872/19906 [============================>.] - ETA: 0sOriginal: MIDDLE Predicted: MIDDLE

14

结果

image.png

原文链接：https://www.jianshu.com/p/2354f3424377

查阅更为简洁方便的分类文章以及最新的课程、产品信息，请移步至全新呈现的“LeadAI学院官网”：

www.leadai.org

请关注人工智能LeadAI公众号，查看更多专业文章

大家都在看

LSTM模型在问答系统中的应用

基于TensorFlow的神经网络解决用户流失概览问题

最全常见算法工程师面试题目整理（一）

最全常见算法工程师面试题目整理（二）

TensorFlow从1到2 | 第三章 深度学习革命的开端：卷积神经网络

装饰器 | Python高级编程

今天不如来复习下Python基础