神经网络和keras

一、图片基础和keras介绍

1、图片三要素

2、keras常用api

3、keras读取图片

使用image前需安装pillow

pip install Pillow

from tensorflow.python.keras.preprocessing.image import load_img, img_to_array


def main():
    img = load_img("./images/img01.jpg", target_size=[300,300])
    img = img_to_array(img)
    print(img)

if __name__ == "__main__":
    main()

4、NHWC和NCHW

 

 

二、神经网络基础

1、神经网络

 

2、创建神经网络模型

 

 #创建神经网络模型
    model = Sequential([
        Flatten(input_shape=(28,28)),
        Dense(128,activation=tf.nn.relu),
        Dense(10,activation=tf.nn.softmax)
    ])

model的一些属性

 

3、模型的训练和评估

1）compile

loss:

2)fit

3)evaluate

4、案例：实现多层神经网络进行时装分类

import tensorflow as tf
from tensorflow.python import keras
from tensorflow.python.keras import Sequential
from tensorflow.python.keras.layers import Flatten, Dense
import os
import numpy as np

class SingleNN(object):

    #神经网络模型
    model = Sequential([
        Flatten(input_shape=(28,28)),
        Dense(128,activation=tf.nn.relu),
        Dense(10,activation=tf.nn.softmax)
    ])

    #初始化获取数据集
    def __init__(self):
        (self.x_train, self.y_train), (self.x_test, self.y_test) = keras.datasets.fashion_mnist.load_data()
        #进行数据归一化
        self.x_train = self.x_train / 255.0
        self.x_test = self.x_test / 255.0


    #编译模型优化器
    def singlenn_compile(self):
        SingleNN.model.compile(
            optimizer=keras.optimizers.SGD(lr=0.01), #优化器
            loss= keras.losses.sparse_categorical_crossentropy, #损失函数 数据非one-hot编码
            metrics=['accuracy']
        )
        return None

    #训练模型
    def singlenn_fit(self):
        #fit当中添加回调函数，记录训练模型过程
        modelcheck = keras.callbacks.ModelCheckpoint(
            filepath="./ckpt/singlenn_{epoch:02d}-{val_acc:.2f}.h5",
            monitor='val_acc', #保存损失率还是准确率
            save_weights_only=True,
            mode='auto',
            period=1
        )

        #添加tensorboard观察
        tensorboard = keras.callbacks.TensorBoard(
            log_dir="./graph",
            histogram_freq=1,
            write_grads=True,
            write_images=True
        )

        SingleNN.model.fit(self.x_train, self.y_train, epochs=1, batch_size=128, callbacks=[modelcheck, tensorboard]) #epochs迭代次数,batch_size每次迭代选择的样本
        return None

    #评估模型
    def single_evalute(self):
        test_loss, test_acc = SingleNN.model.evaluate(self.x_test, self.y_test) #返回损失率和精确率
        print(test_loss,test_acc)
        return None

    #用保存的模型预测
    def single_predict(self):
        if os.path.exists("./tmp/ckpt/checkpoint"):
            SingleNN.model.load_weights("./tmp/ckpt/SingleNN.h5")
            return SingleNN.model.predict(self.x_test)


if __name__ == '__main__':
    singlenn = SingleNN()
    # singlenn.singlenn_compile()
    # singlenn.singlenn_fit()
    # singlenn.single_evalute()

    # #保存模型
    # SingleNN.model.save_weights("./tmp/ckpt/SingleNN.h5")

    #使用模型预测
    predictions = singlenn.single_predict()
    print(predictions)
    print(np.argmax(predictions,axis=1))