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

How to build your first deep learning classifier through TensorFlow, I believe that many inexperienced people are at a loss about this. Therefore, this article summarizes the causes and solutions of the problem. Through this article, I hope you can solve this problem.

Build your first deep learning classifier through TensorFlow (take dogs as an example)!

Convolution neural network (shown above) is a powerful tool for image classification.

The following shows you several different technologies. Help you develop algorithms that can be used for classic image classification problems: detecting dog breeds from images.

And at the end of this article, the author will provide users with a code that can enter any image and return estimates of different dog breeds. In addition, if humans are detected, the algorithm will provide estimates of the breeds that are most similar to each other.

Note: this project is completed as part of the Udacity Machine Learning Nanodegree (GitHub buyback).

1. What is convolution neural network?

Convolution neural network (also known as CNN or ConvNet) is a kind of depth neural network, which is widely used in computer vision and visual image applications.

Famous examples of CNN applications can be found in the research papers of the Stanford research team. They proved the level of skin damage through a single CNN. It should be noted here that the neural network only needs to use pixels and disease tags as inputs to train the image.

The convolution neural network is composed of multiple layers. Compared with other image classification algorithms, it requires less preprocessing.

They learn by using filters to apply convolution neural networks to images. The algorithm uses a small square (or "window") to apply it to the image. Filter the CNN to recognize certain patterns in the image. In order to achieve CNN through the part filter to find the image that matches the content of the original image.

Example of CNN layer architecture for image classification (source: https://bit.ly/2vwlegO)

The first few layers of the network can detect simple features, such as lines, circles, and edges. In each layer, the network can combine these findings and go deep into all levels of the neural network to learn more complex concepts.

1.1 what kind of layers are there?

The overall structure of CNN consists of input layer, hidden layer and output layer. They are type layers, such as convolution, activation, merge, differential pressure, dense and SoftMax layers.

Neural network consists of input layer, hidden layer and output layer (source: https://bit.ly/2Hxhjaw)

Convolution layer (or Conv layer) is the core of convolution neural network. The Conv layer consists of a set of filters. Each filter can be thought of as a small square (with fixed width and height).

In each process, the filter "convolution" the width and height of the input volume. This process produces a two-dimensional activation diagram. This diagram shows the response of the filter at each spatial location.

In order to avoid overfitting. In the active drawing, the Pooling layer is applied to nonlinear downsampling. In other words, the Pooling layer is aggressive when discarding information. But if used properly, it can be very useful. So in CNN architecture, Pooling layers usually follow one or two Conv layers.

In the active drawing, merge layers apply nonlinear downsampling (source: https://bit.ly/2Hxhjaw)

The Dropout layer is also used to reduce overfitting when some activation functions are randomly ignored. Dense layers are fully connected layers and are usually located at the end of the neural network.

1.2 what is the activation function?

The activation function is used to process the output of layers and neural networks. The activation function is the node added to the hidden layer and the output layer.

You will often find that the ReLu activation function is used to hide the layer, while the last layer usually consists of the SoftMax activation function. The author's idea is to carry out a large number of pattern detection and accurately predict the label of a given image by superimposing linear and nonlinear function layers.

As a standardizer, SoftMax is usually found at the last layer and produces a discrete probability distribution vector. This is also in line with the author's purpose, because the author wants the CNN output is the probability that the image corresponds to a particular class.

The most common activation features include ReLU and Sigmoid activation

Through model evaluation and performance evaluation, the loss function is selected. In the CNN of image classification, classification cross entropy is often selected. Use gradual descent to minimize errors-in this article, the author relies on "rmsprop" (adaptive learning rate method) as the optimizer and accuracy as a metric.

two。 Set the building block of the algorithm

To build the algorithm, the authors will use TensorFlow,Keras (neural network API, which runs on TensorFlow) and OpenCV (computer vision library).

2.1 detect whether the image contains a human face

In order to detect whether the provided image contains a human face, the author will use OpenCV's face detection algorithm (one). It is standard procedure to convert an image to grayscale before using any human face detector. The following are the specific steps to detect the storage in the face_cascade multi-scale function and take the grayscale image as a parameter.

2.2 detect whether the image contains a dog

In order to detect whether the provided image contains the dog's face, the author uses the ResNet-50 model of the ImageNet data set used in the pre-training. The model can classify objects from 1000 categories. Given an image, the pre-trained ResNet- 50 model will return the prediction of the containing object in the image.

When using TensorFlow as the backend, Keras CNN requires a 4D array as input. The following path_to_tensor function takes the file path of a string value as input and adjusts it to a square image of 224x224 pixels.

In addition, all pre-trained models have additional standardization steps. So you have to subtract the average pixel from each pixel of each image.

This is achieved through the imported function preprocess_input.

For the final prediction, as shown in the code above. The author obtains the whole corresponding to the prediction object class of the model by predicting the argmax of the probability vector. Here, you can identify object categories by using the ImageNet tag dictionary.

3. Use transfer learning to build your CNN classifier

With the ability to detect people and dogs in the image, the author now needs a method to predict the breeds in the image. In this section, the author will create a CNN for classifying dog breeds.

To reduce training time without sacrificing accuracy, the authors will use transfer learning to train CNN--, a method that allows the use of networks that have been pre-trained on large data sets. By maintaining the previous level and the newly added training layer, the author can use the knowledge gained by the pre-training algorithm to apply it to the program.

Keras includes several pre-trained deep learning models that can be used for prediction, feature extraction and fine tuning.

3.1 Model architecture

As mentioned earlier, the output of the ResNet-50 model is the input layer-also known as the bottleneck feature. In the following code block, the author extracts the corresponding bottleneck features by running the following.

The author will build the model architecture so that the final convolution output of ResNet-50 is returned to the model as input. The author only adds a global average pool and a full connection layer, where the latter contains one node for each dog category and has the Softmax activation function.

As you can see in the code output above, you end up with a neural network with 272517 parameters!

3.2 compile and test the model

Now, you can use CNN to test the degree of breed recognition in the dog image test dataset. In order to fine-tune the model, the author passes 20 iterations to reduce the loss function (classification cross entropy) optimized by using RMS Prop.

Test accuracy: 80.0239%

You can see that the test accuracy of the algorithm has reached 80%, no problem at all!

3.3 Prediction of dog breeds by model

Now there is an algorithm to write a function that takes the image path as input and returns the dog breed prediction model.

4. Test the CNN classifier

Now, you can write a function. In addition to determining the file path where it accepts the image, it is also necessary to determine whether the image contains humans, dogs, or both.

If the dog is detected in the image, the predicted breed is returned. If a human body is detected in the image, a similar breed of dog is returned. If none of them are detected in the image, an output indicating an error is provided.

The algorithm is tested on several sample images:

These predictions look accurate! Now see what happens when the author enters a face!

Obviously, according to the algorithm, the author (blue arrow) looks like a beagle.

After reading the above, have you learned how to build your first deep learning classifier through TensorFlow? If you want to learn more skills or want to know more about it, you are welcome to follow the industry information channel, thank you for reading!

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.