Train an Image Classifier using Keras.

栏目: IT技术 · 发布时间: 4年前

内容简介:I have used sequential API to stack multiple layers of convolution, max-pooling layers with a flatten as well as some dense layers.I strongly recommend you to first build your basics of the working of CNN and get used to all the terminologies that are comm

I have used sequential API to stack multiple layers of convolution, max-pooling layers with a flatten as well as some dense layers.

I strongly recommend you to first build your basics of the working of CNN and get used to all the terminologies that are common in CNNs.

Let’s look at the architecture in detail.

The first layer is a Convolutional Layer that will take an image as an input of size (300 * 300* 3). Then using the activation argument we can apply different activation functions like ‘relu’, ‘tanh’, ‘sigmoid’, ‘softmax’ e.t.c, here I have applied ‘relu’ activation function.

Let’s talk about the number and size of the filters in the first layer we have used.

Through a Kernel size parameter, we could set the size of the filter as we have set it as (5*5), and using filters parameter we could set the number of filters we want to use and in our architecture, it is set to 32.

Now the second layer is a Maxpooling layer that can be applied simply by calling the MaxPooling2D function.

Similarly the third and the fourth layers are the Convolutional and the Maxpooling layers respectively.

Now we’ll flatten out the feature map using flatten function so that we could use those extracted features as an input to the Multi-Layer Perceptron for the classification.

After flattening out the feature map I have used 3 hidden layers that have 128, 64, and 32 number of neurons respectively.

You could build these hidden layers easily with the help of the Dense() method, also you could pass the activation function that you want to apply.

In the end, the last layer also called an Output layer has 2 neurons that give out the likelihood of two classes.

Check that I have used the Softmax activation function in the last layer as it converts the final information in the probability distribution for multiple classes.

Now let’s see the summary of our model using a summary() method.


以上就是本文的全部内容,希望本文的内容对大家的学习或者工作能带来一定的帮助,也希望大家多多支持 码农网

查看所有标签

猜你喜欢:

本站部分资源来源于网络,本站转载出于传递更多信息之目的,版权归原作者或者来源机构所有,如转载稿涉及版权问题,请联系我们

C++面向对象程序设计

C++面向对象程序设计

萨维奇 (Walter Savitch) / 周靖 / 清华大学出版社 / 2003-12 / 59.0

《C++面向对象程序设计》具备良好的编排体系,适合打算涉足编程领域的读者阅读,尤其适合大一学生。它最大的特色是Savitch教授最受欢迎的写作风格,这一风格非常适合初学者,能迅速引导他们开始编程实践。《C++面向对象程序设计》包括全面的习题、项目、编程提示、编程示例、编程陷阱以及有用的小结,以帮助初学者更清楚地了解C++。一起来看看 《C++面向对象程序设计》 这本书的介绍吧!

HTML 压缩/解压工具
HTML 压缩/解压工具

在线压缩/解压 HTML 代码

MD5 加密
MD5 加密

MD5 加密工具

RGB HSV 转换
RGB HSV 转换

RGB HSV 互转工具