本週工作內容依然是繼續調整模型結構，試圖優化結果。在上週設計完成新的數據集，並在本週進行了數據的採集之後，我們在初步數據集上對兩個新的模型進行了嘗試。
一個是減少特徵，只有一層卷積、一層池化、一層全連接結構的簡單模型；一個是增加特徵，包括臉部卷積、池化、全連接，眼睛部位卷積、池化、全連接以及總的全連接等結構的複雜模型。
相關代碼如下
簡單模型
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
@author: vali
"""
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Sat Apr 27 19:54:34 2019
@author: vali
"""
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Sun Mar 31 21:02:14 2019
@author: vali
"""
# coding:utf8
import tensorflow as tf
import numpy as np
import loadData2 as ld
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
def weight_variable(shape):
'''
使用卷積神經網絡會有很多權重和偏置需要創建,我們可以定義初始化函數便於重複使用
這裏我們給權重製造一些隨機噪聲避免完全對稱,使用截斷的正態分佈噪聲,標準差爲0.1
:param shape: 需要創建的權重Shape
:return: 權重Tensor
'''
initial = tf.random_normal(shape,stddev=0.01)
return tf.Variable(initial)
def bias_variable(shape):
'''
偏置生成函數,因爲激活函數使用的是ReLU,我們給偏置增加一些小的正值(0.1)避免死亡節點(dead neurons)
:param shape:
:return:
'''
initial = tf.constant(0.1, shape=shape)
return tf.Variable(initial)
def conv2d(x, W):
'''
卷積層接下來要重複使用,tf.nn.conv2d是Tensorflow中的二維卷積函數,
:param x: 輸入 例如[5, 5, 1, 32]代表 卷積核尺寸爲5x5,1個通道,32個不同卷積核
:param W: 卷積的參數
strides:代表卷積模板移動的步長,都是1代表不遺漏的劃過圖片的每一個點.
padding:代表邊界處理方式,SAME代表輸入輸出同尺寸
:return:
'''
return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding="SAME")
def max_pool_2x2(x):
'''
tf.nn.max_pool是TensorFLow中最大池化函數.我們使用2x2最大池化
因爲希望整體上縮小圖片尺寸,因而池化層的strides設爲橫豎兩個方向爲2步長
:param x:
:return:
'''
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME")
def generator():
eye_img,img_info,img_label = ld.load_data('dataset_300.json')
length = len(eye_img)
index = 0
while (True):
input_eye = eye_img[index]
input_img_info = img_info[index]
input_label = img_label[index]
yield(input_eye,input_img_info,input_label)
index+=1
if(index==length):
index=0
def train(val_eye,val_data_info,val_label):
data = tf.data.Dataset.from_generator(generator,(tf.float32,tf.float32,tf.float32),(tf.TensorShape([25,50,3]),
tf.TensorShape([4]),tf.TensorShape([2])))
data = data.shuffle(158).batch(25)#當shuffle爲數據集大小的時候是完全打亂，batch是一次的大小
data = data.repeat()#代數
data = data.make_one_shot_iterator()
gen = data.get_next()
# 使用佔位符
x0 = tf.placeholder(tf.float32, [None, 25,50,3],'x0')# x爲特徵
x_info = tf.placeholder(tf.float32,[None,4],'x_info')#x_info
y_ = tf.placeholder(tf.float32, [None,2],'y0')# y_爲label
#眼睛部分網絡
# 第一個卷積層 [5, 5, 3, 32]代表 卷積核尺寸爲5x5,3個通道,32個不同卷積核
# 創建濾波器權值-->加偏置-->卷積-->池化
W_conv_eye = weight_variable([5, 5, 3, 20])
b_conv_eye = bias_variable([20])
h_conv_eye = tf.nn.relu(conv2d(x0, W_conv_eye)+b_conv_eye) #25x50x3 與32個5x5x1濾波器 --> 25x50x32
h_pool_eye = max_pool_2x2(h_conv_eye) # 25x50x32 -->13x25x32
h_pool_eye_flat = tf.reshape(h_pool_eye,[-1,13*25*20])
keep_prob_eye = tf.placeholder(tf.float32)
h_eye_drop = tf.nn.dropout(h_pool_eye_flat,keep_prob_eye)
h_test = tf.concat([h_eye_drop,x_info],1)
W_fc_eye = weight_variable([13*25*20+4,2])
b_fc_eye = bias_variable([2])
y_conv = tf.nn.tanh(tf.matmul(h_test,W_fc_eye)+b_fc_eye)
# 定義損失函數,使用均方誤差 同時定義優化器 learning rate = 1e-4
cross_entropy = tf.losses.mean_squared_error(y_conv,y_)
train_step = tf.train.AdamOptimizer(0.0005).minimize(cross_entropy)
# 定義評測準確率
#開始訓練
#writer = tf.summary.FileWriter("./", tf.get_default_graph())
#saver = tf.train.Saver()
with tf.Session() as sess:
init_op = tf.global_variables_initializer() #初始化所有變量
sess.run(init_op)
STEPS = 600
for i in range(STEPS):
#batch = 10
batch_eye,batch_img_info ,batch_y= sess.run(gen)
if i % 2 == 0:
train_cross_entropy= sess.run(cross_entropy,feed_dict={x0:batch_eye , y_:batch_y,
x_info:batch_img_info, keep_prob_eye:1.0})
print(i,train_cross_entropy)
sess.run(train_step, feed_dict={x0: batch_eye,y_: batch_y,
x_info: batch_img_info, keep_prob_eye:0.2})
#saver.save(sess,'./model/model',global_step=STEPS)
#validation
for i in range(10):
batch_eye=val_eye[i].reshape((1,25,50,3))
batch_img_info =val_data_info[i].reshape((1,4))
batch_y= val_label[i].reshape((1,2))
res=sess.run([cross_entropy,y_conv], feed_dict={x0: batch_eye, y_: batch_y,
x_info: batch_img_info, keep_prob_eye:1.0})
print('--------------validation---------------')
print(i,res[0])
print(batch_y,res[1])
if __name__=="__main__":
val_eye,val_data_info,val_label = ld.load_data('dataset_300.json',1)
train(val_eye,val_data_info,val_label )
複雜模型
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
@author: vali
"""
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Sun Mar 31 21:02:14 2019
@author: vali
"""
# coding:utf8
import tensorflow as tf
import numpy as np
import loadData as ld
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
def weight_variable(shape):
'''
使用卷積神經網絡會有很多權重和偏置需要創建,我們可以定義初始化函數便於重複使用
這裏我們給權重製造一些隨機噪聲避免完全對稱,使用截斷的正態分佈噪聲,標準差爲0.1
:param shape: 需要創建的權重Shape
:return: 權重Tensor
'''
initial = tf.random_normal(shape,stddev=0.01)
return tf.Variable(initial)
def bias_variable(shape):
'''
偏置生成函數,因爲激活函數使用的是ReLU,我們給偏置增加一些小的正值(0.1)避免死亡節點(dead neurons)
:param shape:
:return:
'''
initial = tf.constant(0.1, shape=shape)
return tf.Variable(initial)
def conv2d(x, W):
'''
卷積層接下來要重複使用,tf.nn.conv2d是Tensorflow中的二維卷積函數,
:param x: 輸入 例如[5, 5, 1, 32]代表 卷積核尺寸爲5x5,1個通道,32個不同卷積核
:param W: 卷積的參數
strides:代表卷積模板移動的步長,都是1代表不遺漏的劃過圖片的每一個點.
padding:代表邊界處理方式,SAME代表輸入輸出同尺寸
:return:
'''
return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding="SAME")
def max_pool_2x2(x):
'''
tf.nn.max_pool是TensorFLow中最大池化函數.我們使用2x2最大池化
因爲希望整體上縮小圖片尺寸,因而池化層的strides設爲橫豎兩個方向爲2步長
:param x:
:return:
'''
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME")
def generator():
eye_img,face_img,img_info,img_label = ld.load_data('dataset_300.json')
length = len(eye_img)
index = 0
while (True):
input_eye = eye_img[index]
input_face = face_img[index]
input_img_info = img_info[index]
input_label = img_label[index]
yield(input_eye,input_face,input_img_info,input_label)
index+=1
if(index==length):
index=0
def train(val_eye,val_face,val_data_info,val_label):
data = tf.data.Dataset.from_generator(generator,(tf.float32,tf.float32,tf.float32,tf.float32),(tf.TensorShape([25,50,3]),
tf.TensorShape([50,50,3]),tf.TensorShape([4]),tf.TensorShape([2])))
data = data.shuffle(241).batch(25)#當shuffle爲數據集大小的時候是完全打亂，batch是一次的大小
data = data.repeat()#代數
data = data.make_one_shot_iterator()
gen = data.get_next()
# 使用佔位符
x0 = tf.placeholder(tf.float32, [None, 25,50,3],'x0')# x爲特徵
x_face = tf.placeholder(tf.float32, [None, 50,50,3],'x_face')
x_info = tf.placeholder(tf.float32,[None,4],'x_info')#x_info
y_ = tf.placeholder(tf.float32, [None,2],'y0')# y_爲label
#眼睛部分網絡
# 第一個卷積層 [5, 5, 3, 32]代表 卷積核尺寸爲5x5,3個通道,32個不同卷積核
# 創建濾波器權值-->加偏置-->卷積-->池化
W_conv_eye = weight_variable([5, 5, 3, 32])
b_conv_eye = bias_variable([32])
h_conv_eye = tf.nn.relu(conv2d(x0, W_conv_eye)+b_conv_eye) #25x50x3 與32個5x5x1濾波器 --> 25x50x32
h_pool_eye = max_pool_2x2(h_conv_eye) # 25x50x32 -->13x25x32
h_pool_eye_flat = tf.reshape(h_pool_eye,[-1,13*25*32])
keep_prob_eye = tf.placeholder(tf.float32)
h_eye_drop = tf.nn.dropout(h_pool_eye_flat,keep_prob_eye)
h_test = tf.concat([h_eye_drop,x_info],1)
W_fc_eye = weight_variable([13*25*32,128])
b_fc_eye = bias_variable([128])
h_fc_eye = tf.nn.tanh(tf.matmul(h_eye_drop,W_fc_eye)+b_fc_eye)
#臉部分，兩層卷積
W_conv_face1 = weight_variable([5,5,3,32])
b_conv_face1 = bias_variable([32])
h_conv_face1 = tf.nn.relu(conv2d(x_face,W_conv_face1)+b_conv_face1)
h_pool_face1 = max_pool_2x2(h_conv_face1)
W_conv_face2 = weight_variable([5,5,32,64])
b_conv_face2 = bias_variable([64])
h_conv_face2 = tf.nn.relu(conv2d(h_pool_face1,W_conv_face2)+b_conv_face2)
h_pool_face2 = max_pool_2x2(h_conv_face2)
h_pool_face_flat= tf.reshape(h_pool_face2,[-1,13*13*64])
W_fc_face1 = weight_variable([13*13*64,128])
b_fc_face1 = bias_variable([128])
h_fc_face1 = tf.nn.tanh(tf.matmul(h_pool_face_flat,W_fc_face1)+b_fc_face1)
keep_prob_face = tf.placeholder(tf.float32)
h_fc_face1_drop = tf.nn.dropout(h_fc_face1,keep_prob_face)
W_fc_face2 = weight_variable([128,64])
b_fc_face2 = bias_variable([64])
h_fc_face2 = tf.nn.tanh(tf.matmul(h_fc_face1_drop,W_fc_face2)+b_fc_face2)
#合併
h_pool_merge1 = tf.concat([h_fc_eye,h_fc_face2],1)
# W_fc_feature = weight_variable([4,32])
# b_fc_feature = bias_variable([32])
# h_fc_feature = tf.nn.tanh(tf.matmul(x_info,W_fc_feature)+b_fc_feature)
h_pool_merge2 = tf.concat([h_pool_merge1,x_info],1)
W_fc1 = weight_variable([196, 128])
b_fc1 = bias_variable([128])
h_fc1 = tf.nn.tanh(tf.matmul(h_pool_merge2, W_fc1) + b_fc1) #FC層傳播 196--> 128
# 使用Dropout層減輕過擬合,通過一個placeholder傳入keep_prob比率控制
# 在訓練中,我們隨機丟棄一部分節點的數據來減輕過擬合,預測時則保留全部數據追求最佳性能
keep_prob = tf.placeholder(tf.float32)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
# 將Dropout層的輸出連接到一個Softmax層,得到最後的概率輸出
W_fc2 = weight_variable([128, 2]) #2種輸出可能
b_fc2 = bias_variable([2])
y_conv = tf.nn.tanh(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)
# 定義損失函數,使用均方誤差 同時定義優化器 learning rate = 1e-4
cross_entropy = tf.losses.mean_squared_error(y_conv,y_)
train_step = tf.train.AdamOptimizer(0.0005).minimize(cross_entropy)
# 定義評測準確率
accuracy = y_conv
#開始訓練
#writer = tf.summary.FileWriter("./", tf.get_default_graph())
#saver = tf.train.Saver()
with tf.Session() as sess:
init_op = tf.global_variables_initializer() #初始化所有變量
sess.run(init_op)
STEPS = 1000
for i in range(STEPS):
#batch = 10
batch_eye,batch_face,batch_img_info ,batch_y= sess.run(gen)
if i % 2 == 0:
train_cross_entropy= sess.run(cross_entropy,feed_dict={x0:batch_eye , x_face:batch_face,y_:batch_y,
x_info:batch_img_info, keep_prob_eye:1.0,keep_prob_face: 1.0,keep_prob:1.0})
print(i,train_cross_entropy)
sess.run(train_step, feed_dict={x0: batch_eye,x_face:batch_face, y_: batch_y,
x_info: batch_img_info, keep_prob_eye:0.2,keep_prob_face: 0.8,keep_prob:0.5})
#saver.save(sess,'./model/model',global_step=STEPS)
#validation
for i in range(10):
batch_eye=val_eye[i].reshape((1,25,50,3))
batch_face=val_face[i].reshape((1,50,50,3))
batch_img_info =val_data_info[i].reshape((1,4))
batch_y= val_label[i].reshape((1,2))
res=sess.run([cross_entropy,y_conv], feed_dict={x0: batch_eye,x_face:batch_face, y_: batch_y,
x_info: batch_img_info, keep_prob_eye:1.0,keep_prob_face: 1.0,keep_prob:1.0})
print('--------------validation---------------')
print(i,res[0])
print(batch_y,res[1])
if __name__=="__main__":
val_eye,val_face,val_data_info,val_label = ld.load_data('dataset.json',1)
train(val_eye,val_face,val_data_info,val_label )
經過反覆調整batch，learning rate，在模型上訓練400輪的結果如下
複雜模型訓練結果和驗證結果
簡單模型上的訓練驗證結果
在小數據集（251個樣本）上，從上圖中可以看出，同樣在400輪的基礎上，複雜模型在訓練集上表現很好，在驗證集上表現很差，可見有嚴重的過擬合；在簡單模型上，訓練集上的表現和複雜模型差不多，但在驗證集上表現就比複雜模型好很多了。
而且在訓練輪數增加之後，複雜模型loss能更進一步降到0.001左右，而簡單模型就基本在0.008左右。
對於這種現象，下一步我們決定適當減少複雜模型的特徵，並擴充數據集，再對比簡單模型看效果是否有改善。