Learn>
Data Science>
Machine Learning>
Convolutional Neural Networks in TensorFlow course>
Coding augmentation with ImageDataGenerator

Loading...

Coding augmentation with ImageDataGenerator

You've heard the term overfitting a number of times to this point. Overfitting is simply the concept of being over specialized in training -- namely that your model is very good at classifying what it is trained for, but not so good at classifying things that it hasn't seen. In order to generalize your model more effectively, you will of course need a greater breadth of samples to train it on. That's not always possible, but a nice potential shortcut to this is Image Augmentation, where you tweak the training set to potentially increase the diversity of subjects it covers. You'll learn all about that this week!

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. In Course 2 of the deeplearning.ai TensorFlow Specialization, you will learn advanced techniques to improve the computer vision model you built in Course 1. You will explore how to work with real-world images in different shapes and sizes, visualize the journey of an image through convolutions to understand how a computer “sees” information, plot loss and accuracy, and explore strategies to prevent overfitting, including augmentation and dropout. Finally, Course 2 will introduce you to transfer learning and how learned features can be extracted from models. 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 enroll in the first two courses today. We recommend starting with Course 1: Introduction to TensorFlow for AI, ML, and DL.

Skills You'll Learn

Inductive Transfer, Augmentation, Dropouts, Machine Learning, Tensorflow

Taught By

Laurence Moroney
AI Advocate

Transcript

So if you remember the image

generator class that we used earlier,

it actually has the ability to do this for you.

Indeed, you have already done

a little image augmentation

with it when you're re-scaled upon loading.

That saved you from converting all of

your images on the file system and then loading them in,

you just re-scaled on the fly.

So let's take a look at some of the other options.

Here's how you could use a whole bunch of

image augmentation options with

the image generator adding onto re-scale.

Rotation range is a range from

0-180 degrees with which to randomly rotate images.

So in this case, the image will rotate

by random amount between 0 and 40 degrees.

Shifting, moves the image around inside its frame.

Many pictures have the subject centered.

So if we train based on those kind of images,

we might over-fit for that scenario.

These parameters specify, as

a proportion of the image size,

how much we should randomly move the subject around.

So in this case, we might offset it by

20 percent vertically or horizontally.

Shearing is also quite powerful.

So for example, consider the image on the right.

We know that it's a person.

But in our training set,

we don't have any images of a person in that orientation.

However, we do have an image like this one,

where the person is oriented similarly.

So if we shear that person by skewing along the x-axis,

we'll end up in a similar pose.

That's what the shear_range parameter gives us.

It will shear the image by

random amounts up to the specified portion in the image.

So in this case, it will shear

up to 20 percent of the image.

Zoom can also be very effective.

For example, consider the image on the right.

It's obviously a woman facing to the right.

Our image on the left is from

the humans or horses training set.

It's very similar but it

zoomed out to see the full person.

If we zoom in on the training image,

we could end up with a very similar image

to the one on the right.

Thus, if we zoom while training,

we could spot more generalized examples like this one.

So you zoom with code like this.

The 0.2 is a relative portion

of the image you will zoom in on.

So in this case,

zooms will be a random amount

up to 20 percent of the size of the image.

Another useful tool is horizontal flipping.

So for example, if you consider the picture on the right,

we might not be able to classify it correctly as

our training data doesn't have the image

of a woman with her left hand raised,

it does have the image on the left,

where the subjects right arm is raised.

So if the image were flipped horizontally,

then it becomes more structurally similar to the image on

the right and we might not over-fit to right arm raisers.

To turn on random horizontal flipping,

you just say horizontal_flip equals true

and the images will be flipped at random.

Finally, we just specify the fill mode.

This fills in any pixels that

might have been lost by the operations.

I'm just going to stick with nearest here,

which uses neighbors of

that pixel to try and keep uniformity.

Check the carets documentation for some other options.

So that's the concept of image augmentation.

Let's now take a look at cats versus dogs that are

trained with and without augmentation,

so that we can see the impact that this has.

About Coursera

Courses, Specializations, and Online Degrees taught by top instructors from the world's best universities and educational institutions.

Join a community of 40 million learners from around the world

Earn a skill-based course certificate to apply your knowledge

Gain confidence in your skills and further your career

Coursera provides universal access to the world’s best education, partnering with top universities and organizations to offer courses online.

© 2019 Coursera Inc. All rights reserved.

Download on the App Store

Get it on Google Play