There are many transfer learning methods to solve classification problems with respect to the accuracy, i.e 1.VGG (e.g. VGG16 or VGG19) 2.GoogLeNet (e.g. InceptionV3).3.Residual Network (e.g. ResNet50).These are some pre-trained models, but here we are going to build an end to end model as per the convolutional neural networks architecture.

有很多转移学习方法可以解决关于准确性的分类问题，例如1.VGG(例如VGG16或VGG19)2.GoogLeNet(例如InceptionV3).3.Residual Network(例如ResNet50)。这些是一些预先训练的模型，但是在这里，我们将根据卷积神经网络架构来构建端到端模型。

卷积神经网络 (Convolutional Neural Network)

There are many definitions to describe CNN but in simple terms, convolutional refers to the mathematical combination of two functions to produce a third function. It merges two sets of information. in the case of a CNN is used to analyze visual imaginary. They are also known as shift invariant or space invariant artificial neural networks, based on their shared-weights architecture and translation invariance characteristics.

描述CNN的定义很多，但简单来说，卷积是指两个函数的数学组合以产生第三个函数。 它合并了两组信息。 在使用CNN的情况下，可以分析视觉虚像。 基于它们的共享权重架构和平移不变性特征，它们也被称为位移不变或空间不变的人工神经网络。

CNN中的图层 (Layers in CNN)

There are three layers in CNN, convolutional layer, pooling layer, and fully connected layer. Each of these layers has different parameters that can be optimized and performs a different task on the input data.

CNN分为三层：卷积层 ，池化层和完全连接层 。 这些层中的每个层都有可以优化的不同参数，并且对输入数据执行不同的任务。

池化层 (Pooling Layer)

The pooling layer is a building block on CNN. its function progressively reduces the spatial size of the presentation to reduce the number of parameters and computation in network. The pooling layer operates on each feature map independently. max-pooling is the most common approach used in the convolutional neural networks.

池化层是CNN的基础。 它的功能逐渐减小了演示文稿的空间大小，从而减少了网络中的参数和计算量。 池化层在每个要素地图上独立运行。 最大池化是卷积神经网络中最常用的方法。

数据集 (Dataset)

We have data for training:

我们有用于培训的数据：

500 horse images and 527 humans (male & Female) images

500张马图像和527位人类(男性和女性)图像

For Validation:

验证：

122 Horse Images and 123 Human(male & Female) images

122张马图像和123张人类(男性和女性)图像

Dataset link

数据集链接

实作 (Implementation)

Importing necessary libraries and packages.

导入必要的库和包。

import kerasfrom keras.preprocessing.image import ImageDataGeneratorimport matplotlib.pyplot as plt

资料载入 (Data Load)

train_data = "train/"validation_data = "validation/"

Here we have train and validation datasets, we will use train data for training and validation data to prevent overfitting problem. also here we have a small size of the dataset as we before building the CNN model the data should have huge images to make the most accurate model. so we will generate more images using image augmentation.

这里有训练和验证数据集，我们将训练数据用于训练和验证数据以防止过拟合问题。 同样，在这里，我们的数据集很小，因为在构建CNN模型之前，数据应该具有巨大的图像以构成最准确的模型。 因此我们将使用图像增强生成更多图像。

数据预处理 (Data Preprocessing)

#For tariningtrain_data_gen = ImageDataGenerator(rescale=1./255,                                   rotation_range=40,                                   width_shift_range=0.2,                                   height_shift_range=0.2,                                   shear_range=0.2,                                   zoom_range=0.2,                                   horizontal_flip=True,                                   fill_mode='nearest')

Generating more training data for model training using image data generator parameters like rotation and width, height, zoom range.

使用图像数据生成器参数(例如旋转和宽度，高度，缩放范围)为模型训练生成更多训练数据。

train_data1 = train_data_gen.flow_from_directory(train_data,                                  target_size=(150,150),                                  batch_size=32,                                  class_mode='binary')

Let’s Check the image classes

让我们检查一下图像类

train_data1.class_indices

The classes are divided as {‘horses’: 0, ‘humans’: 1}.

这些类别分为{'horses'：0，'humans'：1}。

Now Generating for validation.

现在生成以进行验证。

validation_data_gen = ImageDataGenerator(rescale=1./255)validation_data1 = validation_data_gen.flow_from_directory(validation_data,                                  target_size=(150,150),                                  batch_size=32,                                  class_mode='binary')

Now plotting the generated images.

现在绘制生成的图像。

def plotImages(images_arr):    fig, axes = plt.subplots(1, 5, figsize=(20, 20))    axes = axes.flatten()    for img, ax in zip(images_arr, axes):        ax.imshow(img)    plt.tight_layout()    plt.show()

Let’s plots some training image datasets.

让我们绘制一些训练图像数据集。

images = [train_data1[0][0][0] for i in range(5)]plotImages(images)

建立CNN (Building CNN)

We are using Keras here so it makes code much simpler form because it uses TensorFlow API. so let’s start to build a model step by step.

我们在这里使用Keras，因为它使用TensorFlow API，所以它使代码的形式更简单。 因此，让我们开始逐步构建模型。

cnn_model = keras.models.Sequential([    keras.layers.Conv2D(filters=32, kernel_size=3, input_shape=[150,150,3]),    keras.layers.MaxPooling2D(pool_size=(2,3)),    keras.layers.Conv2D(filters=64, kernel_size=3),    keras.layers.MaxPooling2D(pool_size=(2,2)),    keras.layers.Conv2D(filters=128, kernel_size=3),    keras.layers.MaxPooling2D(pool_size=(2,2)),    keras.layers.Conv2D(filters=256, kernel_size=3),    keras.layers.MaxPooling2D(pool_size=(2,2)),

    keras.layers.Dropout(0.5),    keras.layers.Flatten(), #after this we will go for neural network building    keras.layers.Dense(units=128, activation='relu'), #inputlayers    keras.layers.Dropout(0.1),    keras.layers.Dense(units=256, activation='relu'), #Hidden layer    keras.layers.Dropout(0.25),    keras.layers.Dense(units=2, activation='softmax') #output layer with 2 neurons

])

Basically we took 128 for input and 256 neurons for hidden, and we used relu and softmax activation functions, with respect to classes we used softmax activation function. Now let’s compile the model.

基本上，我们使用128个输入作为输入，使用256个神经元进行隐藏，并使用relu和softmax激活函数，就使用softmax激活函数的类而言。 现在让我们编译模型。

from keras.optimizers import Adamfrom keras.callbacks import ModelCheckpoint

Let’s set the optimizer and loss function now.

现在设置优化器和损失函数。

cnn_model.compile(optimizer=Adam(lr=0.0001), loss='sparse_categorical_crossentropy', metrics=['accuracy'])

Before going to build a model we can set a path to save our model, so in my case, I have set my local systems path.

在构建模型之前，我们可以设置保存模型的路径，因此，就我而言，我已经设置了本地系统路径。

model_path='human_horse_predict.h5'checkpoint = ModelCheckpoint(model_path, monitor='val_accuracy', verbose=1, save_best_only=True, mode='max')callbacks_list = [checkpoint]

模型火车 (Model Train)

history = cnn_model.fit(train_data1,                       epochs=100,                       verbose=1,                       validation_data=validation_data1,                       callbacks=callbacks_list)

I have trained this model by using Cuda and it took less time for training, better you can train using Google COLAB.

我已经使用Cuda训练了该模型，并且花了更少的时间进行训练，更好的是可以使用Google COLAB进行训练。

Now the model is trained, let’s see the summary of the model.

现在已经对模型进行了训练，让我们看看模型的摘要。

#Summarize history for accuracyplt.plot(history.history['accuracy'])plt.plot(history.history['val_accuracy'])plt.title('Model Loss')plt.ylabel('accuracy')plt.xlabel('epoch')plt.legend(['train', 'test'], loc='upper left')plt.show()

Let’s look at our model accuracy on train and test data.

让我们看看我们在训练和测试数据上的模型准确性。

You can see how our model is overfitted as per epoch. Let’s test the model.

您可以看到我们的模型在每个时期都是过拟合的。 让我们测试一下模型。

模型测试 (Model Test)

model1 = keras.models.load_model(model_path)

Give the saved model path.

给出保存的模型路径。

Let’s take some human and horse images and assing in some variables.

让我们拍摄一些人和马的图像，并评估一些变量。

##Horse imagesh1 = 'train/horses/horse01-1.png'h2 = 'train/horses/horse01-2.png'h3 = 'train/horses/horse01-5.png'#human images hu1 = 'train/humans/human01-02.png'hu2 = 'train/humans/human01-05.png'hu3 = 'train/humans/human01-08.png

Creating one function which will prepress our image in an array format.

创建一个将以数组格式预印图像的功能。

import numpy as npfrom keras.preprocessing import image

Defining function.

定义功能。

def pred_human_horse(model, horse_or_human):    test_image = image.load_img(horse_or_human, target_size = (155,155))    test_image = image.img_to_array(test_image)/255    test_image = np.expand_dims(test_image, axis=0)

    result = model.predict(test_image).round(3)

    pred = np.argmax(result)

    if pred==0:        print("Horse")    else:        print("Human")

预测 (Prediction)

For using three images we are going to predict.

对于使用三个图像，我们将进行预测。

for horse_or_human in [h1,h2,h3]:    pred_human_horse(model1, horse_or_human)HorseHorseHorse

So, we got a good prediction here.

因此，我们在这里得到了很好的预测。

Hope you like this article, we can improve this model for using hyperparameter tuning and you can also try the pre-trained models likeVGG (e.g. VGG16 or VGG19).

希望您喜欢本文，我们可以改进此模型以使用超参数调整，还可以尝试使用预训练的模型，例如VGG(例如VGG16或VGG19)。

IPython notebook.

IPython 笔记本 。

Github Link for all machine learning and Deep Learning Resources, and you can also see my machine learning deployments using Flask API.

Github 链接提供了所有机器学习和深度学习资源，您还可以使用Flask API查看我的机器学习部署。