【Tensorflow】多层感知器，一维卷积和二维卷积在分类问题的使用区别

1.摘要

Fashion-MNIST 的目的是要成为 MNIST 数据集的一个直接替代品。作为算法作者，你不需要修改任何的代码，就可以直接使用这个数据集。Fashion-MNIST 的图片大小，训练、测试样本数及类别数与经典 MNIST 完全相同。每个训练和测试样本都按照以下类别进行了标注：标注编号描述0T-shirt/top（T恤）1Trouser（裤子）2Pullover（套衫）3Dress（裙子）4Coat（外套）5Sandal（凉鞋）6Shirt（汗衫）7Sneaker（运动鞋）8Bag（包）9Ankle boot（踝靴）
本文，主要是基于fashion_mnist数据集，分别采用多层感知器，一维卷积神经网络，二维卷积神经网络对数据集进行分类，为了直观的发现他们的使用区别。

2.数据预处理

导入相应的包

import warnings
warnings.filterwarnings("ignore")
import tensorflow as tf
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from tensorflow.keras import layers,models,losses
from tensorflow.keras.preprocessing.image import ImageDataGenerator

导入数据并进行可视化

fashion_mnist = tf.keras.datasets.fashion_mnist
(train_images,train_labels),(test_images,test_labels) = fashion_mnist.load_data()
# 类别名称
class_names = ['T-shirt/top', 'Trouser', 'Pullover', 'Dress', 'Coat', 'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle boot'] 
print(train_images.shape,train_labels.shape,test_images.shape,test_labels.shape)

# 归一化
train_images = train_images/255.0
test_images = test_images/255.0

plt.figure(figsize=(10, 10))
for i in range(16):
    ax = plt.subplot(4,4,i+1)
    plt.subplots_adjust(hspace=0.5)
    ax.imshow(train_images[i])
    ax.set_title(class_names[train_labels[i]])
plt.show()

3.二维卷积神经网络

改变数据维度，方便模型的导入

Xtrain = train_images.reshape(train_images.shape[0],train_images.shape[1],train_images.shape[2],1)
Xtest = np.expand_dims(test_images,axis=3)
print(Xtrain.shape,Xtest.shape)

搭建模型

# 数据增强
datagen = ImageDataGenerator(height_shift_range=3,horizontal_flip=True) 
train_generator = datagen.flow(Xtrain,train_labels,batch_size=128)
# 搭建网络
model_CNN = models.Sequential()
model_CNN.add(layers.Conv2D(32,(3,3),activation='relu',input_shape=(28,28,1)))
model_CNN.add(layers.MaxPool2D((2,2)))
model_CNN.add(layers.Conv2D(64,(3,3),activation="relu"))
model_CNN.add(layers.MaxPool2D((2,2)))
model_CNN.add(layers.Conv2D(128,(3,3),activation="relu"))
model_CNN.add(layers.Flatten())
model_CNN.add(layers.Dense(64,activation="relu"))
model_CNN.add(layers.Dense(10,activation="softmax"))
model_CNN.compile(optimizer="adam",loss=losses.sparse_categorical_crossentropy,metrics=["accuracy"])
model_CNN.summary()

训练数据

history = model_CNN.fit(train_generator,epochs=5,batch_size=1280,validation_data=(Xtest,test_labels))

test_loss,test_acc = model_CNN.evaluate(Xtest,test_labels,verbose=1)
print("测试集的准确率为{}".format(test_acc))

4.一维卷积神经网络

改变数据维度，方便模型的导入

Xtrain_1 = train_images.reshape(train_images.shape[0],train_images.shape[1]*train_images.shape[2],1)
Xtest_1 = test_images.reshape(test_images.shape[0],test_images.shape[1]*test_images.shape[2],1)
print(Xtrain_1.shape,Xtest_1.shape)

搭建模型

# 搭建模型
model_CNN1 = models.Sequential()
model_CNN1.add(layers.Conv1D(32, 3, activation='relu', input_shape=(784, 1), padding="same"))
model_CNN1.add(layers.MaxPool1D(pool_size=3, strides=3))
model_CNN1.add(layers.Conv1D(64, 3, strides=1, activation='relu', padding='same'))
model_CNN1.add(layers.MaxPool1D(pool_size=3, strides=3))
# 神经元随机失活
model_CNN1.add(layers.Dropout(0.25))
model_CNN1.add(layers.Flatten())
model_CNN1.add(layers.Dense(1024,activation="relu"))
model_CNN1.add(layers.Dense(10,activation="softmax"))
model_CNN1.compile(optimizer="adam",loss=losses.sparse_categorical_crossentropy,metrics=["accuracy"])
model_CNN1.summary()

训练数据

history1 = model_CNN1.fit(Xtrain_1,train_labels,epochs=5,batch_size=1280,validation_data=(Xtest_1,test_labels))

test_loss1,test_acc1 = model_CNN1.evaluate(Xtest_1,test_labels,verbose=1)
print("测试集的准确率为{}".format(test_acc1))

5.多层感知器分类

改变数据维度，方便模型的导入

Xtrain_2 = train_images.reshape(train_images.shape[0],train_images.shape[1]*train_images.shape[2])
Xtest_2 = test_images.reshape(test_images.shape[0],test_images.shape[1]*test_images.shape[2])
print(Xtrain_2.shape,Xtest_2.shape)

搭建模型

# 搭建模型
model_MLP = models.Sequential()
# 神经元随机失活
model_MLP.add(layers.Dense(1024,activation="relu",input_dim=784))
model_MLP.add(layers.Dense(512,activation="relu"))
model_MLP.add(layers.Dense(100,activation="relu"))
model_MLP.add(layers.Dense(10,activation="softmax"))

model_MLP.compile(optimizer="adam",loss=losses.sparse_categorical_crossentropy,metrics=["accuracy"])
model_MLP.summary()

训练数据

history2 = model_MLP.fit(Xtrain_2,train_labels,epochs=5,batch_size=1280,validation_data=(Xtest_2,test_labels))

test_loss2,test_acc2 = model_MLP.evaluate(Xtest_2,test_labels,verbose=1)
print("测试集的准确率为{}".format(test_acc2))