Defining a ConvNet to use complex images

Loading...

Из курса от партнера deeplearning.ai

Introduction to TensorFlow for Artificial Intelligence, Machine Learning, and Deep Learning

1500 оценки

deeplearning.ai

Introduction to TensorFlow for Artificial Intelligence, Machine Learning, and Deep Learning

1500 оценки

If you are a software developer who wants to build scalable AI-powered algorithms, you need to understand how to use the tools to build them. This course is part of the upcoming Machine Learning in Tensorflow Specialization and will teach you best practices for using TensorFlow, a popular open-source framework for machine learning. The Machine Learning course and Deep Learning Specialization from Andrew Ng teach the most important and foundational principles of Machine Learning and Deep Learning. This new deeplearning.ai TensorFlow Specialization teaches you how to use TensorFlow to implement those principles so that you can start building and applying scalable models to real-world problems. To develop a deeper understanding of how neural networks work, we recommend that you take the Deep Learning Specialization. The full deeplearning.ai TensorFlow Specialization will be available later this year, but you can get started with Course 1, Introduction to Tensorflow for AI, ML and DL, available now on Coursera.

Из урока

Using Real-world Images

Last week you saw how to improve the results from your deep neural network using convolutions. It was a good start, but the data you used was very basic. What happens when your images are larger, or if the features aren’t always in the same place? Andrew and Laurence discuss this to prepare you for what you’ll learn this week: handling complex images!

A conversation with Andrew Ng2:20

Understanding ImageGenerator4:15

Defining a ConvNet to use complex images2:36

Training the ConvNet with fit_generator2:55

Walking through developing a ConvNet2:09

Walking through training the ConvNet with fit_generator3:29

Adding automatic validation to test accuracy4:44

Exploring the impact of compressing images3:09

Познакомьтесь с преподавателями

Laurence Moroney
AI Advocate
Google Brain

So let's now take a look at the definition of

the neural network that we'll use to

classify horses versus humans.

It's very similar to what you

just used for the fashion items,

but there are a few minor differences based on this data,

and the fact that we're using generators.

So here's the code.

As you can see, it's the sequential as before with

convolutions and pooling before

we get to the dense layers at the bottom.

But let's highlight some of the differences.

First of all, you'll notice that there are

three sets of convolution pooling layers at the top.

This reflects the higher complexity

and size of the images.

Remember our earlier our 28 by

28.5 to 13 and then five before flattening,

well, now we have 300 by 300.

So we start at 298 by 298 and then have that etc., etc.

until by the end, we're at a 35 by 35 image.

We can even add another couple of layers to this if

we wanted to get to the same ballpark size as previously,

but we'll keep it at three for now.

Another thing to pay attention to is the input shape.

We resize their images to be

300 by 300 as they were loaded,

but they're also color images.

So there are three bytes per pixel.

One byte for the red, one for green,

and one for the blue channel, and that's

a common 24-bit color pattern.

If you're paying really close attention,

you can see that the output layer has also changed.

Remember before when you created the output layer,

you had one neuron per class,

but now there's only one neuron for two classes.

That's because we're using

a different activation function where

sigmoid is great for binary classification,

where one class will tend towards

zero and the other class tending towards one.

You could use two neurons here if you want,

and the same softmax function as before,

but for binary this is a bit more efficient.

If you want you can experiment with

the workbook and give it a try yourself.

Now, if we take a look at our model summary,

we can see the journey of

the image data through the convolutions

The 300 by 300 becomes 298 by

298 after the three by three filter,

it gets pulled to 149 by

149 which in turn gets reduced to

73 by 73 after

the filter that then gets pulled to 35 by 35,

this will then get flattened,

so 64 convolutions that are 35

squared and shape will get fed into the DNN.

If you multiply 35 by 35 by 64,

you get 78,400, and that's

the shape of the data once

it comes out of the convolutions.

If we had just fed raw

300 by 300 images without the convolutions,

that would be over 900,000 values.

So we've already reduced it quite a bit.

Ознакомьтесь с нашим каталогом

Присоединяйтесь бесплатно и получайте персонализированные рекомендации, обновления и предложения.

Coursera делает лучшее в мире образование доступным каждому, предлагая онлайн-курсы от ведущих университетов и организаций.

© Coursera Inc., 2019 Все права защищены.

Загрузить из App Store

Загрузить в Google Play