Keras介绍

Keras是一个用Python编写的开源神经网络库。它能够运行在TensorFlow，Microsoft Cognitive Toolkit，Theano或PlaidML之上。

时装分类数据集介绍

该数据集包含70000张灰度图像，一共有10个类别。

步骤分析及代码实现

读取数据集

from tensorflow.python.keras.datasets import fashion_mnist

class SingleNN(object):
    def __init__(self):
        (self.train,self.train_label),(self.test,self.test_label) = fashion_mnist.load_data()
        self.train = self.train / 255.0
        self.test = self.test / 255.0

训练数据，测试数据的形状如下：

train: (60000, 28, 28)
train_label: (60000,)
test: (10000, 28, 28)
test_label: (10000,)

模型构建

模型结构：双层神经网络

• 隐藏层有128个神经元，激活函数选择relu
• 全连接层有10个神经元，因为fashion_mnist具有10个类别，激活函数选择softmax

from tensorflow.python.keras.models import Sequential
from tensorflow.python.keras.layers import Flatten,Dense

import tensorflow as tf
tf.compat.v1.disable_eager_execution()

class SingleNN(object):
    model = Sequential([
        Flatten(input_shape=(28,28)),
        Dense(128,activation=tf.nn.relu),
        Dense(10,activation=tf.nn.softmax)
    ])

编译定义优化过程

优化器选择Adam，损失函数选择交叉熵损失**(标签数据是整型数据，需要先转换为one-hot编码)**

from tensorflow.python.keras.optimizer_v1 import Adam
from tensorflow.python.keras.losses import sparse_categorical_crossentropy

class SingleNN(object):
    def compile(self):
        SingleNN.model.compile(optimizer=Adam(),
        loss=sparse_categorical_crossentropy,
        metrics=['accuracy'])

定义训练函数

epochs设置为3次，batch_size设置为32

class SingleNN(object):
    def fit(self):
        SingleNN.model.fit(self.train,self.train_label,epochs=3,batch_size=32)
        return None

定义评估函数

记录测试集的损失函数值以及准确率

class SingleNN(object):
	    def evaluate(self):
        test_loss,test_acc = SingleNN.model.evaluate(self.test,self.test_label)
        print("test_loss:",test_loss)
        print("test_acc:",test_acc)
        return None

开启会话运行图

with tf.compat.v1.Session() as sess:
	cnn = SingleNN()
	cnn.compile()
	cnn.fit()
	cnn.evaluate()

运行结果

Train on 60000 samples
Epoch 1/3
60000/60000 [==============================] - 5s 81us/sample - loss: 0.4989 - accuracy: 0.8242
Epoch 2/3
60000/60000 [==============================] - 5s 77us/sample - loss: 0.3749 - accuracy: 0.8635
Epoch 3/3
60000/60000 [==============================] - 5s 78us/sample - loss: 0.3373 - accuracy: 0.8766
test_loss: 0.3732113081932068
test_acc: 0.8629