Deep learning image recognition 2 model implementation

Aidemy　2020/11/10

Introduction

Hello, it is Yope! I am a liberal arts student, but I was interested in the possibilities of AI, so I went to the AI-specialized school "Aidemy" to study. I would like to share the knowledge gained here with you, and I am summarizing it on Qiita. I am very happy that many people have read the previous summary article. Thank you! This is the second post for deep learning and image recognition. Nice to meet you.

• This article is a summary of what you learned in "Aidemy" "in your own words". It may contain mistakes and misunderstandings. Please note.

What to learn this time ・ About convolutional neural networks ・ About model implementation

(Review) Convolutional neural network

Overview

-__ Convolutional neural network (CNN) __ is a deep learning method mainly used in image processing. -The multi-layer perceptron (MLP) used in Chapter 1 received others by one-dimensional input, but CNN can receive data as it is __ two-dimensional, so two-dimensional data such as images can be used as information. You can learn without losing.

-The convolutional neural network consists of __ "convolutional layer" __ and __ "pooling layer" __. The convolutional layer is a layer that focuses on a part of the input data and examines the characteristics of the image. -The __ pooling layer __ is a layer that reduces the amount of data by reducing the __ output data from the convolution layer __. As a method of reduction, there are __ "Max pooling" __ which takes the maximum value of each feature and __ "Average pooling" __ which takes the average value.

・ Relationship diagram of convolutional layer

・ Relationship diagram of pooling layer (Max pooling)

Layer implementation

-This time, we will implement the layer with Sequential model. -One of the layer types is __ "Dense" __. This is a layer where all inputs and outputs are connected. In the model below, the fully connected layer is used twice.

-The "convolutional layer" in the previous section can be implemented with __ "Conv2D ()" __. __ "filters" __ is the number of filters, __ "kernel_size" __ is the size of the filter, __ "strides" __ is the distance to move the filters at once, __ "padding" __ is __'same' If __, the input size and output size are aligned, and if 'valid', do not.

-The pooling layer can be implemented with __ "MaxPooling2D ()" __ for Max pooling. __ "pool_size" __ specifies the size of the pooled range, __ "strides" __ specifies the distance to move the filter at one time (strides need not be specified).

-For these input layers, specify the size of the input data with __ "input_dim (one-dimensional)" and "input_shape (multidimensional)" __. For example, in the case of (28 × 28) RGB image data, “input_shape = (28,28,3)”.

-When combining a layer __ that outputs __2 dimensions such as Conv2D and MaxPooling2D and a layer __ that receives ___ 1 dimensional input such as Dense, it is necessary to add __ "flattening layer" __ There is. -The flattening layer can be added with __ "Flatten ()" __, and no __ argument is required __.

-The following code is an example of adding the layers so far. model.add(Conv2D(input_shape=(28, 28, 3), filters=32, kernel_size=(2, 2), strides=(1, 1), padding="same")) model.add(MaxPooling2D(pool_size=(2, 2))) model.add(Conv2D(filters=32, kernel_size=(2, 2), strides=(1, 1), padding="same")) model.add(MaxPooling2D(pool_size=(2, 2)))

Model learning

Specification of activation function

-Specify the activation function with __ "model_add (Activation ('function name'))" __. There are __ "sigmoid" __ for sigmoid functions, __ "relu" __ for ReLU functions, and __ "softmax" __ used in the output layer.

Compile, learn

-After adding a layer to the model, perform compile. In compilation, __ How to change the learning rate __, __ Loss function specification __, __ Accuracy index specification __, etc. are performed. More precisely, compiling is __ "parsing code written in a high-level programming language and pre-converting it into a machine language that can be directly executed by a computer" __. -Compile can be done with __ "model.compile ()" __. __ "optimizer" __ specifies the learning rate, __ "loss" __ specifies the loss function, __ "metrics" __ specifies the accuracy index. ・ There are __ [SGD, RMSprop, Adagrad, Adadelta, Adam] __ etc. for the learning rate, but which one to use is decided exploratory __. -Loss functions include __mean squared_error (mean_squared_error) __, __ cross entropy (category_crossentropy, binary_crossentropy) __, etc. -Metrics is often specified as __ "['accuracy']" __.

-Learning is done with __ "model.fit ()" __. In addition to inputting training data and passing a label to this, __ "batch_size" __ specifies the number of data to be passed at one time, and __ "epochs" __ specifies the number of learnings. The batch_size is specified to perform the __ "mini-batch learning" __ learned at the end of Chapter 1.

・ Code example

Model performance evaluation

-For the trained model, __ test data is used to evaluate the accuracy __. This is done with __ "model.evaluate ()" __. In addition to passing the input data and label of the test data, if verbose = 1, the calculation process is displayed.

Save and load models

-When saving the created model, do as __ "model.save ('model.h5')" __. -When loading this model, use __ "model.load_model ('model.h5')" __.

-Also, in order to allow __ people to view and edit this model, it is necessary to convert it to __ "json" __ format once. This is done with __ "model.to_json ()" __. -To load this, use __ "model_from_json (converted model)" __.

Summary

-In image processing, a "convolutional neural network" that can handle two-dimensional image data is used. -As a model implementation, implement by connecting the convolutional layer and the fully connected layer with a flattening layer, and after compiling and learning, evaluate the model.

This time is over. Thank you for reading until the end.