Shulou(Shulou.com)06/01 Report--

How to use TensorFlow for Poets to train image classifiers, many novices are not very clear about this, in order to help you solve this problem, the following small series will explain in detail for everyone, people who have this need can learn, I hope you can gain something.

background introduction

Today we learn how to train image classifiers, just by going through an image catalog. Let's say you want to build a classifier to distinguish between pictures of Decepticons and Triceratops:

Or you want to distinguish between Monet and Picasso:

To do this we need to use a code lab called TensorFlow for Poets, which is a good way to start learning and doing image classification related work.

Now there are two things that need to be done: first, this code lab is very advanced, in order to train our classifier we just need to run a few scripts, impressive is that the classifier creates better scripts than we write ourselves, and then we show you what the code lab looks like in operation, and we add context and background to explain why it's so good, so let's get started. To train an image classifier with TensorFlow for Poets, we only need to provide one thing-training data. It's an image catalog with lots of images:

My plan is to create a picture classifier that can distinguish between five different flowers, roses, sunflowers, etc.:

So here's my training data. Notice that I have five categories, one for each flower, and in each category there are a lot of pictures:

The reason we use flowers is because we provide this data in the coding lab, so you can get started quickly. If you want to use your own images, like dinosaur images or famous paintings, you just create a catalog and put all the images you find on the Internet into a catalog, and each catalog needs about 100 images:

Once we have the training data we can start training the classifier, which we will do using TensorFlow. TensorFlow is an open source machine learning library that is particularly powerful in the field of deep learning, which has grown rapidly in recent years, especially in the field of image classification:

That's what we're going to do today. In previous articles we discussed how to distinguish apples from oranges, and we found it impossible to do this manually. Because there are so many different apples and oranges in the world, but now we also know that classifiers take features as input, it's hard to extract useful information from images with code.

For example, you don't want to write code that detects the texture of a piece of fruit. To solve this problem we use deep learning, because it has a huge advantage in image processing, which is that you don't have to manually extract features, you can use pixel image features. The classifier will solve the rest of the problem, but in order to distinguish between our training materials, let's compare the iris dataset with the image catalog in iris. Each column in the dataset describes a feature of the flower, and you can imagine that we extracted these features manually, like measuring flowers by length, in contrast to our training data.

In TensorFlow for Poets this is a list of labeled images, and the image classifier is simply a function f(x)=y where x is a 2D matrix of image pixels and y is the rose label. When we talk about deep learning, the classifiers we use are called neural networks:

At a high level this is just another classifier. The difference is that neural networks can learn a lot of complex functions, and TensorFlow for Poets in this coding lab is responsible for preparing and training neural networks for you behind the scenes, which does not mean that TensorFlow coding is harder to write than what we have seen so far. We'll run this script first, and all the details are available in the code lab (link at the end), so I won't go into them here:

But I'll just give you background on two things you might want to know, first of all, the script takes about 20 minutes to train the classifier. Now you can ask yourself, is this a long time? It's not very long, of course, and TensorFlow for Poets doesn't actually train classifiers from scratch, it trains classifiers from an existing classifier called Inception, which is one of Google's best image classifiers, and it's open source.

With only a few thousand images in our training data, Inception has been trained on more than 1,000 categories and 1.2 million images. Training Inception takes about two weeks on a fast computer with eight graphics processors. In TensorFlow for Poets we build on Inception and then use a feature called retraining to tune it to better distinguish our images.

Let's also redefine some of the parameters we learned before Inception, so that we can create a high-accuracy classifier with very little training data, and now the training is over:

Once we have a trained classifier, we can test it. I will download this picture of roses from the internet:

Use our classifier to predict what kind of flower it is. We can see that the answer is correct, we can see the confidence distribution for other flowers:

Remember, our classifier only knows the pictures we show it, so if we ask it to classify a picture like the Colosseum:

He'll say it's a flower. We can only hope that the confidence interval will be low, so let me say one or two more things: training an image classifier is a game of diversity and quantity. In terms of diversity, we have more images of different roses, and the predictions are more accurate. For example, our training data contains red, white and yellow roses, we also have pictures taken at different angles, such as from above or from the side, and we also have roses as foreground or background images. Quantitatively, I mean the more training data we have, the better classifiers we'll create.

There are hundreds of photos in the Rose folder, which is enough to retrain Inception. You might be able to use fewer images but the accuracy might be reduced.

Did reading the above help you? If you still want to have further understanding of related knowledge or read more related articles, please pay attention to the industry information channel, thank you for your support.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.