Tensorflow2（预课程）---7.1、cifar10分类-层方式

Tensorflow2（预课程）---7.1、cifar10分类-层方式

一、总结

一句话总结：

可以看到，用全连接层神经网络来对cifar10分类的时候，准确率只有50%左右，可以用卷积神经网络，准确率应该高很多

 

 

1、rgb3通道图片处理？

也是一样打平：model.add(tf.keras.layers.Flatten(input_shape=(32,32,3)))

# 将多维数据（60000, 32, 32, 3）变成一维
# 把图像扁平化成一个向量
model.add(tf.keras.layers.Flatten(input_shape=(32,32,3))) 

 

 

2、提高神经网络准确率的方法？

甲、增加层

乙、增加每层节点数

 

 

 

二、cifar10分类-层方式

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：

输入是32*32*3维，输出是一个label，是一个10分类问题，

需要one_hot编码么，如果是one_hot编码，那么输出是10维

也就是 32*32*3->n->10，可以试下3072->1024->256->128->64->32->10

In [10]:

# 构建容器
model = tf.keras.Sequential()
# 输入层
# 将多维数据（60000, 32, 32, 3）变成一维
# 把图像扁平化成一个向量
model.add(tf.keras.layers.Flatten(input_shape=(32,32,3))) 
# 中间层
model.add(tf.keras.layers.Dense(256,activation='relu'))
model.add(tf.keras.layers.Dense(128,activation='relu'))
model.add(tf.keras.layers.Dense(64,activation='relu'))
model.add(tf.keras.layers.Dense(32,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 #   
=================================================================
flatten (Flatten)            (None, 3072)              0         
_________________________________________________________________
dense (Dense)                (None, 256)               786688    
_________________________________________________________________
dense_1 (Dense)              (None, 128)               32896     
_________________________________________________________________
dense_2 (Dense)              (None, 64)                8256      
_________________________________________________________________
dense_3 (Dense)              (None, 32)                2080      
_________________________________________________________________
dense_4 (Dense)              (None, 10)                330       
=================================================================
Total params: 830,250
Trainable params: 830,250
Non-trainable params: 0
_________________________________________________________________

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 [==============================] - 6s 4ms/step - loss: 1.9072 - acc: 0.3012 - val_loss: 1.7286 - val_acc: 0.3780
Epoch 2/50
1563/1563 [==============================] - 6s 4ms/step - loss: 1.7166 - acc: 0.3824 - val_loss: 1.6491 - val_acc: 0.4011
Epoch 3/50
1563/1563 [==============================] - 6s 4ms/step - loss: 1.6342 - acc: 0.4131 - val_loss: 1.6009 - val_acc: 0.4249
Epoch 4/50
1563/1563 [==============================] - 7s 5ms/step - loss: 1.5760 - acc: 0.4325 - val_loss: 1.5339 - val_acc: 0.4532
Epoch 5/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.5387 - acc: 0.4476 - val_loss: 1.5336 - val_acc: 0.4525
Epoch 6/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.5044 - acc: 0.4603 - val_loss: 1.5341 - val_acc: 0.4531
Epoch 7/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.4788 - acc: 0.4724 - val_loss: 1.4793 - val_acc: 0.4754
Epoch 8/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.4557 - acc: 0.4815 - val_loss: 1.4708 - val_acc: 0.4715
Epoch 9/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.4345 - acc: 0.4883 - val_loss: 1.4712 - val_acc: 0.4776
Epoch 10/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.4201 - acc: 0.4931 - val_loss: 1.4784 - val_acc: 0.4718
Epoch 11/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.3992 - acc: 0.5003 - val_loss: 1.4802 - val_acc: 0.4699
Epoch 12/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.3818 - acc: 0.5043 - val_loss: 1.4721 - val_acc: 0.4841
Epoch 13/50
1563/1563 [==============================] - 7s 5ms/step - loss: 1.3723 - acc: 0.5082 - val_loss: 1.4530 - val_acc: 0.4930
Epoch 14/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.3573 - acc: 0.5141 - val_loss: 1.4627 - val_acc: 0.4802
Epoch 15/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.3416 - acc: 0.5212 - val_loss: 1.4565 - val_acc: 0.4872
Epoch 16/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.3314 - acc: 0.5228 - val_loss: 1.4438 - val_acc: 0.4914
Epoch 17/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.3201 - acc: 0.5274 - val_loss: 1.4539 - val_acc: 0.4972
Epoch 18/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.3068 - acc: 0.5323 - val_loss: 1.4281 - val_acc: 0.4979
Epoch 19/50
1563/1563 [==============================] - 7s 5ms/step - loss: 1.2945 - acc: 0.5356 - val_loss: 1.4282 - val_acc: 0.4998
Epoch 20/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.2850 - acc: 0.5411 - val_loss: 1.4722 - val_acc: 0.4866
Epoch 21/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.2758 - acc: 0.5431 - val_loss: 1.4234 - val_acc: 0.5056
Epoch 22/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.2678 - acc: 0.5473 - val_loss: 1.4430 - val_acc: 0.4976
Epoch 23/50
1563/1563 [==============================] - 7s 5ms/step - loss: 1.2570 - acc: 0.5504 - val_loss: 1.4371 - val_acc: 0.5049
Epoch 24/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.2479 - acc: 0.5545 - val_loss: 1.4330 - val_acc: 0.4973
Epoch 25/50
1563/1563 [==============================] - 7s 5ms/step - loss: 1.2376 - acc: 0.5569 - val_loss: 1.4484 - val_acc: 0.4962
Epoch 26/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.2319 - acc: 0.5582 - val_loss: 1.4493 - val_acc: 0.4992
Epoch 27/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.2268 - acc: 0.5599 - val_loss: 1.4509 - val_acc: 0.4970
Epoch 28/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.2151 - acc: 0.5627 - val_loss: 1.4624 - val_acc: 0.4929
Epoch 29/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.2167 - acc: 0.5638 - val_loss: 1.4369 - val_acc: 0.5055
Epoch 30/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.2028 - acc: 0.5683 - val_loss: 1.4270 - val_acc: 0.5103
Epoch 31/50
1563/1563 [==============================] - 7s 5ms/step - loss: 1.1950 - acc: 0.5713 - val_loss: 1.4530 - val_acc: 0.5043
Epoch 32/50
1563/1563 [==============================] - 7s 5ms/step - loss: 1.1897 - acc: 0.5735 - val_loss: 1.4564 - val_acc: 0.5033
Epoch 33/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.1838 - acc: 0.5766 - val_loss: 1.4682 - val_acc: 0.4981
Epoch 34/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.1748 - acc: 0.5783 - val_loss: 1.5145 - val_acc: 0.4862
Epoch 35/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.1732 - acc: 0.5806 - val_loss: 1.4676 - val_acc: 0.4941
Epoch 36/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.1645 - acc: 0.5827 - val_loss: 1.4899 - val_acc: 0.4923
Epoch 37/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.1583 - acc: 0.5830 - val_loss: 1.4780 - val_acc: 0.5017
Epoch 38/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.1580 - acc: 0.5825 - val_loss: 1.4808 - val_acc: 0.4928
Epoch 39/50
1563/1563 [==============================] - 7s 5ms/step - loss: 1.1505 - acc: 0.5866 - val_loss: 1.4767 - val_acc: 0.5002
Epoch 40/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.1446 - acc: 0.5881 - val_loss: 1.4793 - val_acc: 0.4922
Epoch 41/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.1426 - acc: 0.5901 - val_loss: 1.5011 - val_acc: 0.4976
Epoch 42/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.1357 - acc: 0.5936 - val_loss: 1.4820 - val_acc: 0.5033
Epoch 43/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.1299 - acc: 0.5922 - val_loss: 1.5110 - val_acc: 0.4964
Epoch 44/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.1244 - acc: 0.5957 - val_loss: 1.5044 - val_acc: 0.4945
Epoch 45/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.1231 - acc: 0.5957 - val_loss: 1.4931 - val_acc: 0.4981
Epoch 46/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.1148 - acc: 0.6001 - val_loss: 1.4952 - val_acc: 0.4983
Epoch 47/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.1112 - acc: 0.6006 - val_loss: 1.5073 - val_acc: 0.4976
Epoch 48/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.1085 - acc: 0.6033 - val_loss: 1.5084 - val_acc: 0.5002
Epoch 49/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.0985 - acc: 0.6059 - val_loss: 1.5110 - val_acc: 0.5010
Epoch 50/50
1563/1563 [==============================] - 7s 4ms/step - loss: 1.0957 - acc: 0.6073 - val_loss: 1.5539 - val_acc: 0.4895

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)

[[6.2265161e-02 3.6133263e-02 2.9608410e-02 ... 1.4575418e-02
  2.3273245e-01 3.9462153e-02]
 [1.2520383e-03 2.8723732e-02 1.3520659e-05 ... 1.4830942e-05
  2.0657738e-01 7.6337880e-01]
 [4.4282374e-01 2.5962757e-02 4.9018036e-03 ... 5.0854594e-03
  4.7830233e-01 3.1182669e-02]
 ...
 [2.5298037e-02 1.5645087e-03 4.5591959e-01 ... 2.0820173e-02
  4.5593046e-02 5.5354494e-03]
 [4.8573306e-03 1.7568754e-02 2.7253026e-01 ... 2.8846616e-01
  3.5799999e-02 1.9565571e-02]
 [1.8785211e-04 9.9887984e-06 4.1551115e-03 ... 9.8020053e-01
  2.0824935e-05 2.1540496e-04]]
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([8 9 8 ... 2 7 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 [ ]: