keras入门(四) ResNet网络 实现猫狗大战

先来一张resnet网络结构图，本文采用的是resnet50

核心思想就是引入skip connect，阻止深层网络的退化。

                                                    

import keras
from keras import Model, layers
from keras.layers import Conv2D, BatchNormalization, Add, MaxPooling2D, AveragePooling2D, Input, Activation, Flatten, \
    Dense, GlobalAveragePooling2D, Dropout
from keras.callbacks import EarlyStopping
from keras.optimizers import SGD
import os
import argparse
import random
import numpy as np
from scipy.misc import imread, imresize, imsave

parser = argparse.ArgumentParser()
parser.add_argument('--train_dir', default='F:/kaggle/train/')
parser.add_argument('--test_dir', default='F:/kaggle/test/')
parser.add_argument('--log_dir', default='./')
parser.add_argument('--batch_size', default=32)
parser.add_argument('--gpu', type=int, default=0)
args = parser.parse_args()

os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu)
type_list = ['cat', 'dog']


def identity_block(input, stage, filters):
    conv_name = 'cn' + str(stage) + '_branch'
    bais_name = 'bn' + str(stage)  + '_branch'

    input_short_cut = input
    filter1, filter2, filter3 = filters
    input = Conv2D(filter1, (1, 1), strides=(1, 1), padding='same', name=conv_name + '_one')(input)
    input = BatchNormalization(name=bais_name + "_one_BN")(input)
    input = Activation("relu")(input)
    input = Conv2D(filter2, (3, 3), strides=(1, 1), padding='same', name=conv_name + '_two')(input)
    input = BatchNormalization(name=bais_name + "_two_BN")(input)
    input = Activation("relu")(input)
    input = Conv2D(filter3, (1, 1), strides=(1, 1), padding='same', name=conv_name + '_three')(input)
    input = BatchNormalization(name=bais_name + "_three_BN")(input)
    input = Add()([input, input_short_cut])
    input = Activation("relu")(input)
    return input

def convolutional_block(input, stage, strides, filters):
    conv_name = 'cn' + str(stage) + '_branch'
    bais_name = 'bn' + str(stage) + '_branch'

    input_short_cut = input
    filter1, filter2, filter3 = filters
    input = Conv2D(filter1, (1, 1), strides=strides, padding='same', name=conv_name + '_one')(input)
    input = BatchNormalization(name=bais_name + "_one_BN")(input)
    input = Activation('relu')(input)
    input = Conv2D(filter2, (3, 3), strides=(1, 1), padding='same', name=conv_name + '_two')(input)
    input = BatchNormalization(name=bais_name + "_two_BN")(input)
    input = Activation('relu')(input)
    input = Conv2D(filter3, (1, 1), strides=(1, 1), padding='same', name=conv_name + '_three')(input)
    input = BatchNormalization(name=bais_name + "_three_BN")(input)
    input_short_cut = Conv2D(filter3, (1, 1), strides=strides, padding='same', name=conv_name + '_four')(input_short_cut)
    input_short_cut = BatchNormalization(name=bais_name + "_four_BN")(input_short_cut)
    input = Add()([input, input_short_cut])
    input = Activation('relu')(input)
    return input



def ResNet50(input_shape=(224, 224, 3)):
    input = Input(shape=input_shape)
    outinput = Conv2D(64, (7, 7), strides=(2, 2), padding='same', name='conv1')(input)
    outinput = BatchNormalization(name="conv1_BN")(outinput)
    outinput = Activation("relu")(outinput)
    outinput = MaxPooling2D(pool_size=(3, 3), strides=(2, 2), name='pool1_1')(outinput)

    outinput = convolutional_block(outinput, 1, 1, [64, 64, 256])
    outinput = identity_block(outinput, 2, [64, 64, 256])
    outinput = identity_block(outinput, 3, [64, 64, 256])
    outinput = convolutional_block(outinput, 4, 2, [128, 128, 512])
    outinput = identity_block(outinput, 5, [128, 128, 512])
    outinput = identity_block(outinput, 6, [128, 128, 512])
    outinput = identity_block(outinput, 7, [128, 128, 512])
    outinput = convolutional_block(outinput, 8, 2, [256, 256, 1024])
    outinput = identity_block(outinput, 9, [256, 256, 1024])
    outinput = identity_block(outinput, 10, [256, 256, 1024])
    outinput = identity_block(outinput, 11, [256, 256, 1024])
    outinput = identity_block(outinput, 12, [256, 256, 1024])
    outinput = identity_block(outinput, 13, [256, 256, 1024])
    outinput = convolutional_block(outinput, 14, 2, [512, 512, 2048])
    outinput = identity_block(outinput, 15, [512, 512, 2048])
    outinput = identity_block(outinput, 16, [512, 512, 2048])
    outinput = GlobalAveragePooling2D(name='avg_pool')(outinput)
    outinput = (Dropout(0.5))(outinput)
    #outinput = AveragePooling2D(pool_size=(2, 2), padding="same")(outinput)
    #outinput = Flatten()(outinput)
    outinput = Dense(2, activation='softmax')(outinput)
    model = Model(inputs=input, outputs=outinput, name="ResNet50")
    return model


def create_generate(train_file_list, batch_size, input_size):
    while (True):
        random.shuffle(train_file_list)
        image_data = np.zeros((batch_size, input_size[0], input_size[1], 3), dtype='float32')
        label_data = np.zeros((batch_size, 1), dtype='int32')
        for index, file_name in enumerate(train_file_list):
            image = imresize(imread(args.train_dir + file_name), input_size)
            label = file_name.split('.')[0]
            image_data[index % batch_size] = np.reshape(image / 255, (input_size[0], input_size[1], 3))
            label_data[index % batch_size] = type_list.index(label)
            if (0 == (index + 1) % batch_size):
                # label_data = to_categorical(labei_data, 2)
                label_data = keras.utils.to_categorical(label_data, 2)
                yield image_data, label_data
                image_data = np.zeros((batch_size, input_size[0], input_size[1], 3), dtype='float32')
                label_data = np.zeros((batch_size, 1), dtype='int32')


def train(model):
    early_stop = EarlyStopping(monitor='val_loss', patience=5, mode='min')
    tensor_board = keras.callbacks.TensorBoard(log_dir='./logs', histogram_freq=0, batch_size=16, write_graph=True, write_grads=False, write_images=False, embeddings_freq=0, embeddings_layer_names=None, embeddings_metadata=None, embeddings_data=None, update_freq='epoch')
    sgd = SGD(lr=0.001, decay=1e-8, momentum=0.9, nesterov=True)
    model.compile(optimizer=sgd, loss='categorical_crossentropy', metrics=['accuracy'])
    #model.compile(optimizer=sgd, loss='binary_crossentropy', metrics=['accuracy'])
    train_generate, validation_genarate = prepare_data()
    model.fit_generator(generator=train_generate, steps_per_epoch=1250
                        , epochs=30, verbose=1,
                        validation_data=validation_genarate, validation_steps=30, max_queue_size=1,
                        shuffle=True,
                        callbacks=[early_stop, tensor_board])
    model.save_weights(args.log_dir + 'model.h5')


def predict(model):
    file_list = os.listdir(args.test_dir)
    cat_real_count = 0
    dog_real_count = 0
    cat_predict_count = 0
    dog_predict_count = 0
    for index, file_name in enumerate(file_list):
        print(index)
        label = file_name.split('.')[0]
        if 'cat' == label:
            cat_real_count = cat_real_count + 1
        else:
            dog_real_count = dog_real_count + 1
        image = imresize(imread(args.test_dir + file_name), (224, 224, 3))
        label = model.predict(np.reshape(image / 255, (1, 224, 224, 3)))
        label = np.argmax(label, 1)
        if 0 == label:
            imsave('./result/cat/' + file_name, imread(args.test_dir + file_name))
            cat_predict_count = cat_predict_count + 1
        else:
            imsave('./result/dog/' + file_name, imread(args.test_dir + file_name))
            dog_predict_count = dog_predict_count + 1
    # print(cat_predict_count / cat_real_count)


if __name__ == '__main__':
    try:
        model = ResNet50()
        train(model)
    except Exception  as err:
        print(err)

与VGG相比，在训练过程中能够快速收敛，而且准确率更高，能够达到97%，但是训练过程中遇到了过拟合的问题，验证集上的准确率在70%~80%，怎么调参都无法解决，训练了好几天，没有太多的头绪，等以后有想法了再解决一下