Tensorflow2（预课程）---7.2、cifar10分类-层方式-卷积神经网络

Tensorflow2（预课程）---7.2、cifar10分类-层方式-卷积神经网络

一、总结

一句话总结：

卷积层构建非常简单，就是CBAPD，注意卷积层接全连接层的时候注意flatten打平

# 构建容器
model = tf.keras.Sequential()

# 卷积层
model.add(tf.keras.layers.Conv2D(filters=6, kernel_size=(3, 3), padding='same',input_shape=(32,32,3))) # 卷积层
model.add(tf.keras.layers.BatchNormalization()) # BN层
model.add(tf.keras.layers.Activation('relu')) # 激活层
model.add(tf.keras.layers.MaxPool2D(pool_size=(2, 2), strides=2, padding='same')) # 池化层
model.add(tf.keras.layers.Dropout(0.5)) # dropout层

# 全连接层
model.add(tf.keras.layers.Flatten())
model.add(tf.keras.layers.Dense(512,activation='relu'))
model.add(tf.keras.layers.Dense(256,activation='relu'))
model.add(tf.keras.layers.Dense(128,activation='relu'))
# 输出层
model.add(tf.keras.layers.Dense(10,activation='softmax'))
# 模型的结构
model.summary()

 

 

 

二、cifar10分类-层方式-卷积神经网络（65）（dropuot=0.5）

博客对应课程的视频位置：

 

步骤

1、读取数据集
2、拆分数据集（拆分成训练数据集和测试数据集）
3、构建模型
4、训练模型
5、检验模型

需求

cifar10（物品分类）

---

该数据集共有60000张彩色图像，这些图像是32*32，分为10个类，每类6000张图。这里面有50000张用于训练，构成了5个训练批，每一批10000张图；另外10000用于测试，单独构成一批。测试批的数据里，取自10类中的每一类，每一类随机取1000张。抽剩下的就随机排列组成了训练批。注意一个训练批中的各类图像并不一定数量相同，总的来看训练批，每一类都有5000张图。

---

 

 

In [1]:

import pandas as pd
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt

1、读取数据集

直接从tensorflow的dataset来读取数据集即可

In [2]:

 (train_x, train_y), (test_x, test_y) = tf.keras.datasets.cifar10.load_data()
print(train_x.shape, train_y.shape)

(50000, 32, 32, 3) (50000, 1)

这是32*32的彩色图，rgb三个通道如何处理呢

In [3]:

plt.imshow(train_x[0])
plt.show()

In [4]:

plt.figure()
plt.imshow(train_x[1])
plt.figure()
plt.imshow(train_x[2])
plt.show()

In [5]:

print(test_y)

[[3]
 [8]
 [8]
 ...
 [5]
 [1]
 [7]]

In [6]:

# 像素值 RGB
np.max(train_x[0])

Out[6]:

255

2、拆分数据集（拆分成训练数据集和测试数据集）

上一步做了拆分数据集的工作

In [7]:

# 图片数据如何归一化
# 直接除255即可
train_x = train_x/255
test_x = test_x/255

In [8]:

# 像素值 RGB
np.max(train_x[0])

Out[8]:

1.0

In [9]:

train_y=train_y.flatten()
test_y=test_y.flatten()
train_y = tf.one_hot(train_y, depth=10)
test_y = tf.one_hot(test_y, depth=10)
print(test_y.shape)

(10000, 10)

3、构建模型

应该构建一个怎么样的模型：

In [10]:

# 构建容器
model = tf.keras.Sequential()

# 卷积层
model.add(tf.keras.layers.Conv2D(filters=6, kernel_size=(3, 3), padding='same',input_shape=(32,32,3))) # 卷积层
model.add(tf.keras.layers.BatchNormalization()) # BN层
model.add(tf.keras.layers.Activation('relu')) # 激活层
model.add(tf.keras.layers.MaxPool2D(pool_size=(2, 2), strides=2, padding='same')) # 池化层
model.add(tf.keras.layers.Dropout(0.5)) # dropout层

# 全连接层
model.add(tf.keras.layers.Flatten())
model.add(tf.keras.layers.Dense(512,activation='relu'))
model.add(tf.keras.layers.Dense(256,activation='relu'))
model.add(tf.keras.layers.Dense(128,activation='relu'))
# 输出层
model.add(tf.keras.layers.Dense(10,activation='softmax'))
# 模型的结构
model.summary()

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 32, 32, 6)         168       
_________________________________________________________________
batch_normalization (BatchNo (None, 32, 32, 6)         24        
_________________________________________________________________
activation (Activation)      (None, 32, 32, 6)         0         
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 16, 16, 6)         0         
_________________________________________________________________
dropout (Dropout)            (None, 16, 16, 6)         0         
_________________________________________________________________
flatten (Flatten)            (None, 1536)              0         
_________________________________________________________________
dense (Dense)                (None, 512)               786944    
_________________________________________________________________
dense_1 (Dense)              (None, 256)               131328    
_________________________________________________________________
dense_2 (Dense)              (None, 128)               32896     
_________________________________________________________________
dense_3 (Dense)              (None, 10)                1290      
=================================================================
Total params: 952,650
Trainable params: 952,638
Non-trainable params: 12
_________________________________________________________________

4、训练模型

In [11]:

# 配置优化函数和损失器
model.compile(optimizer='adam',loss='categorical_crossentropy',metrics=['acc'])
# 开始训练
history = model.fit(train_x,train_y,epochs=50,validation_data=(test_x,test_y))

Epoch 1/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.6209 - acc: 0.4125 - val_loss: 1.3494 - val_acc: 0.5211
Epoch 2/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.3799 - acc: 0.5090 - val_loss: 1.2441 - val_acc: 0.5570
Epoch 3/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.2797 - acc: 0.5454 - val_loss: 1.2381 - val_acc: 0.5648
Epoch 4/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.2087 - acc: 0.5712 - val_loss: 1.1817 - val_acc: 0.5834
Epoch 5/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.1527 - acc: 0.5936 - val_loss: 1.1657 - val_acc: 0.5954
Epoch 6/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.1070 - acc: 0.6111 - val_loss: 1.1013 - val_acc: 0.6192
Epoch 7/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.0516 - acc: 0.6297 - val_loss: 1.1230 - val_acc: 0.6097
Epoch 8/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.0123 - acc: 0.6422 - val_loss: 1.1698 - val_acc: 0.5970
Epoch 9/50
1563/1563 [==============================] - 7s 4ms/step - loss: 0.9770 - acc: 0.6541 - val_loss: 1.0683 - val_acc: 0.6303
Epoch 10/50
1563/1563 [==============================] - 7s 4ms/step - loss: 0.9436 - acc: 0.6680 - val_loss: 1.1084 - val_acc: 0.6166
Epoch 11/50
1563/1563 [==============================] - 7s 4ms/step - loss: 0.9142 - acc: 0.6798 - val_loss: 1.0627 - val_acc: 0.6303
Epoch 12/50
1563/1563 [==============================] - 7s 4ms/step - loss: 0.8823 - acc: 0.6897 - val_loss: 1.0670 - val_acc: 0.6315
Epoch 13/50
1563/1563 [==============================] - 7s 4ms/step - loss: 0.8527 - acc: 0.6975 - val_loss: 1.1186 - val_acc: 0.6186
Epoch 14/50
1563/1563 [==============================] - 7s 5ms/step - loss: 0.8319 - acc: 0.7058 - val_loss: 1.0376 - val_acc: 0.6405
Epoch 15/50
1563/1563 [==============================] - 7s 5ms/step - loss: 0.8099 - acc: 0.7131 - val_loss: 1.0556 - val_acc: 0.6372
Epoch 16/50
1563/1563 [==============================] - 7s 4ms/step - loss: 0.7912 - acc: 0.7213 - val_loss: 1.1238 - val_acc: 0.6216
Epoch 17/50
1563/1563 [==============================] - 7s 5ms/step - loss: 0.7657 - acc: 0.7313 - val_loss: 1.0372 - val_acc: 0.6452
Epoch 18/50
1563/1563 [==============================] - 7s 4ms/step - loss: 0.7461 - acc: 0.7388 - val_loss: 1.0493 - val_acc: 0.6516
Epoch 19/50
1563/1563 [==============================] - 7s 4ms/step - loss: 0.7268 - acc: 0.7445 - val_loss: 1.0913 - val_acc: 0.6377
Epoch 20/50
1563/1563 [==============================] - 7s 4ms/step - loss: 0.7088 - acc: 0.7484 - val_loss: 1.0899 - val_acc: 0.6356
Epoch 21/50
1563/1563 [==============================] - 7s 5ms/step - loss: 0.6907 - acc: 0.7579 - val_loss: 1.2129 - val_acc: 0.6060
Epoch 22/50
1563/1563 [==============================] - 7s 4ms/step - loss: 0.6715 - acc: 0.7648 - val_loss: 1.1440 - val_acc: 0.6253
Epoch 23/50
1563/1563 [==============================] - 7s 4ms/step - loss: 0.6609 - acc: 0.7677 - val_loss: 1.0904 - val_acc: 0.6430
Epoch 24/50
1563/1563 [==============================] - 7s 5ms/step - loss: 0.6349 - acc: 0.7769 - val_loss: 1.1141 - val_acc: 0.6409
Epoch 25/50
1563/1563 [==============================] - 7s 4ms/step - loss: 0.6267 - acc: 0.7805 - val_loss: 1.1753 - val_acc: 0.6198
Epoch 26/50
1563/1563 [==============================] - 7s 4ms/step - loss: 0.6104 - acc: 0.7839 - val_loss: 1.1174 - val_acc: 0.6442
Epoch 27/50
1563/1563 [==============================] - 7s 5ms/step - loss: 0.6064 - acc: 0.7878 - val_loss: 1.1045 - val_acc: 0.6407
Epoch 28/50
1563/1563 [==============================] - 7s 4ms/step - loss: 0.5871 - acc: 0.7925 - val_loss: 1.1561 - val_acc: 0.6374
Epoch 29/50
1563/1563 [==============================] - 7s 4ms/step - loss: 0.5778 - acc: 0.7961 - val_loss: 1.1731 - val_acc: 0.6310
Epoch 30/50
1563/1563 [==============================] - 7s 5ms/step - loss: 0.5635 - acc: 0.8009 - val_loss: 1.1782 - val_acc: 0.6277
Epoch 31/50
1563/1563 [==============================] - 7s 4ms/step - loss: 0.5512 - acc: 0.8078 - val_loss: 1.1345 - val_acc: 0.6416
Epoch 32/50
1563/1563 [==============================] - 7s 4ms/step - loss: 0.5442 - acc: 0.8089 - val_loss: 1.1450 - val_acc: 0.6440
Epoch 33/50
1563/1563 [==============================] - 7s 4ms/step - loss: 0.5341 - acc: 0.8142 - val_loss: 1.1412 - val_acc: 0.6436
Epoch 34/50
1563/1563 [==============================] - 7s 5ms/step - loss: 0.5210 - acc: 0.8182 - val_loss: 1.3549 - val_acc: 0.6052
Epoch 35/50
1563/1563 [==============================] - 7s 5ms/step - loss: 0.5159 - acc: 0.8193 - val_loss: 1.1278 - val_acc: 0.6539
Epoch 36/50
1563/1563 [==============================] - 7s 5ms/step - loss: 0.5070 - acc: 0.8217 - val_loss: 1.1707 - val_acc: 0.6441
Epoch 37/50
1563/1563 [==============================] - 8s 5ms/step - loss: 0.4952 - acc: 0.8256 - val_loss: 1.2057 - val_acc: 0.6345
Epoch 38/50
1563/1563 [==============================] - 9s 6ms/step - loss: 0.4828 - acc: 0.8305 - val_loss: 1.1531 - val_acc: 0.6506
Epoch 39/50
1563/1563 [==============================] - 8s 5ms/step - loss: 0.4757 - acc: 0.8326 - val_loss: 1.1730 - val_acc: 0.6413
Epoch 40/50
1563/1563 [==============================] - 8s 5ms/step - loss: 0.4709 - acc: 0.8364 - val_loss: 1.2210 - val_acc: 0.6418
Epoch 41/50
1563/1563 [==============================] - 8s 5ms/step - loss: 0.4600 - acc: 0.8392 - val_loss: 1.2046 - val_acc: 0.6400
Epoch 42/50
1563/1563 [==============================] - 8s 5ms/step - loss: 0.4563 - acc: 0.8418 - val_loss: 1.3474 - val_acc: 0.6066
Epoch 43/50
1563/1563 [==============================] - 8s 5ms/step - loss: 0.4467 - acc: 0.8451 - val_loss: 1.3222 - val_acc: 0.6239
Epoch 44/50
1563/1563 [==============================] - 8s 5ms/step - loss: 0.4381 - acc: 0.8478 - val_loss: 1.4079 - val_acc: 0.6078
Epoch 45/50
1563/1563 [==============================] - 8s 5ms/step - loss: 0.4295 - acc: 0.8510 - val_loss: 1.1937 - val_acc: 0.6558
Epoch 46/50
1563/1563 [==============================] - 8s 5ms/step - loss: 0.4242 - acc: 0.8518 - val_loss: 1.2182 - val_acc: 0.6504
Epoch 47/50
1563/1563 [==============================] - 8s 5ms/step - loss: 0.4172 - acc: 0.8550 - val_loss: 1.2157 - val_acc: 0.6465
Epoch 48/50
1563/1563 [==============================] - 8s 5ms/step - loss: 0.4085 - acc: 0.8585 - val_loss: 1.2362 - val_acc: 0.6506
Epoch 49/50
1563/1563 [==============================] - 8s 5ms/step - loss: 0.4136 - acc: 0.8562 - val_loss: 1.2187 - val_acc: 0.6505
Epoch 50/50
1563/1563 [==============================] - 8s 5ms/step - loss: 0.4109 - acc: 0.8570 - val_loss: 1.1927 - val_acc: 0.6530

In [12]:

plt.plot(history.epoch,history.history.get('loss'))
plt.title("train data loss")
plt.show()

In [13]:

plt.plot(history.epoch,history.history.get('val_loss'))
plt.title("test data loss")
plt.show()

In [14]:

plt.plot(history.epoch,history.history.get('acc'))
plt.title("train data acc")
plt.show()

In [15]:

plt.plot(history.epoch,history.history.get('val_acc'))
plt.title("test data acc")
plt.show()

5、检验模型

In [16]:

# 看一下模型的预测能力
pridict_y=model.predict(test_x)
print(pridict_y)
print(test_y)

[[4.5043899e-04 1.1276284e-02 7.4143500e-05 ... 1.0536104e-03
  1.8133764e-03 6.7020673e-04]
 [4.5525166e-03 5.5896789e-02 1.5721616e-05 ... 6.7627825e-06
  9.3855476e-01 8.3672808e-04]
 [6.8573037e-04 2.2152183e-04 8.1032022e-06 ... 2.9030971e-05
  9.9802029e-01 8.5997221e-04]
 ...
 [9.7019289e-04 2.5262563e-05 1.4840290e-01 ... 1.3067307e-01
  2.1098384e-03 5.0971273e-04]
 [5.0079293e-04 9.8580730e-01 2.0990559e-04 ... 1.4797639e-03
  1.3322505e-03 1.7820521e-03]
 [1.2427025e-06 1.1827400e-06 1.1778239e-05 ... 9.7277403e-01
  4.7926852e-07 5.0128792e-06]]
tf.Tensor(
[[0. 0. 0. ... 0. 0. 0.]
 [0. 0. 0. ... 0. 1. 0.]
 [0. 0. 0. ... 0. 1. 0.]
 ...
 [0. 0. 0. ... 0. 0. 0.]
 [0. 1. 0. ... 0. 0. 0.]
 [0. 0. 0. ... 1. 0. 0.]], shape=(10000, 10), dtype=float32)

In [17]:

# 在pridict_y中找最大值的索引，横向
pridict_y = tf.argmax(pridict_y, axis=1)
print(pridict_y)
#
test_y = tf.argmax(test_y, axis=1)
print(test_y)

tf.Tensor([3 8 8 ... 3 1 7], shape=(10000,), dtype=int64)
tf.Tensor([3 8 8 ... 5 1 7], shape=(10000,), dtype=int64)

In [18]:

plt.figure()
plt.imshow(test_x[0])
plt.figure()
plt.imshow(test_x[1])
plt.figure()
plt.imshow(test_x[2])
plt.figure()
plt.imshow(test_x[3])
plt.show()

In [ ]: