深度学习笔记——分类模型（七）DenseNet121

# -*- coding: utf-8 -*-
import datetime
from keras.models import Model
from keras.layers import *
from keras.preprocessing.image import ImageDataGenerator

def dense_block(x, blocks, name):
    for i in range(blocks):
        x = conv_block(x, 32, name=name + '_block' + str(i + 1))
    return x

def conv_block(x, growth_rate, name):
    bn_axis = 3
    x1 = BatchNormalization(axis=bn_axis, epsilon=1.001e-5, name=name + '_0_bn')(x)
    x1 = Activation('relu', name=name + '_0_relu')(x1)

    x1 = Conv2D(4 * growth_rate, 1, use_bias=False, name=name + '_1_conv')(x1)
    x1 = BatchNormalization(axis=bn_axis, epsilon=1.001e-5, name=name + '_1_bn')(x1)
    x1 = Activation('relu', name=name + '_1_relu')(x1)

    x1 = Conv2D(growth_rate, 3, padding='same', use_bias=False, name=name + '_2_conv')(x1)
    x = Concatenate(axis=bn_axis, name=name + '_concat')([x, x1])
    return x

def transition_block(x, reduction, name):
    bn_axis = 3
    x = BatchNormalization(axis=bn_axis, epsilon=1.001e-5, name=name + '_bn')(x)
    x = Activation('relu', name=name + '_relu')(x)

    x = Conv2D(int(K.int_shape(x)[bn_axis] * reduction), 1, use_bias=False, name=name + '_conv')(x)
    x = AveragePooling2D(2, strides=2, name=name + '_pool')(x)
    return x

def DenseNet121(input_shape=(224, 224, 3), classes=1000):
    blocks = [6, 12, 24, 16]
    img_input = Input(shape=input_shape)

    bn_axis = 3

    x = ZeroPadding2D(padding=((3, 3), (3, 3)))(img_input)
    x = Conv2D(64, 7, strides=2, use_bias=False, name='conv1/conv')(x)
    x = BatchNormalization(axis=bn_axis, epsilon=1.001e-5, name='conv1/bn')(x)
    x = Activation('relu', name='conv1/relu')(x)
    x = ZeroPadding2D(padding=((1, 1), (1, 1)))(x)
    x = MaxPooling2D(3, strides=2, name='pool1')(x)

    x = dense_block(x, blocks[0], name='conv2')
    x = transition_block(x, 0.5, name='pool2')

    x = dense_block(x, blocks[1], name='conv3')
    x = transition_block(x, 0.5, name='pool3')

    x = dense_block(x, blocks[2], name='conv4')
    x = transition_block(x, 0.5, name='pool4')

    x = dense_block(x, blocks[3], name='conv5')

    x = BatchNormalization(axis=bn_axis, epsilon=1.001e-5, name='bn')(x)

    x = GlobalAveragePooling2D(name='avg_pool')(x)
    if classes == 2:
        x = Dense(1, activation='sigmoid', name='pred')(x)
    else:
        x = Dense(classes, activation='softmax', name='pred')(x)

    model = Model(img_input, x, name='densenet121')

    return model

def main():
    width = 224
    height = 224
    batch_size = 8

    generator = ImageDataGenerator(horizontal_flip=True,
                                   vertical_flip=True,
                                   validation_split=0.2)

    train_generator = generator.flow_from_directory(directory="datasets/train",
                                                    target_size=(width, height),
                                                    batch_size=batch_size,
                                                    class_mode="binary",
                                                    subset="training")

    val_generator = generator.flow_from_directory(directory="datasets/train",
                                                  target_size=(width, height),
                                                  batch_size=batch_size,
                                                  class_mode="binary",
                                                  subset="validation")

    model = DenseNet121(classes=2)
    model.summary()
    # Compile model
    model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])

    model.fit_generator(train_generator, validation_data=val_generator, epochs=10, verbose=1)

if __name__ == '__main__':
    tic = datetime.datetime.now()
    main()
    toc = datetime.datetime.now()
    print("\nThis model took ", (toc - tic))