Kaggle比赛NCFM图像分类任务简介

为了保护和监控海洋环境及生态平衡，大自然保护协会（The Nature Conservancy）邀请Kaggle社区的参赛者们开发能够出机器学习算法，自动分类和识别远洋捕捞船上的摄像头拍摄到的图片中鱼类的品种，例如不同种类的吞拿鱼和鲨鱼。大自然保护协会一共提供了3777张标注的图片作为训练集，这些图片被分为了8类，其中7类是不同种类的海鱼，剩余1类则是不含有鱼的图片，每张图片只属于8类中的某一类别。

图片中待识别的海鱼所占整张图片的一小部分，这就给识别带来了很大的挑战性。此外，为了衡量算法的有效性，还提供了额外的1000张图片作为测试集，参赛者们需要设计出一种图像识别的算法，尽可能地识别出这1000张测试图片属于8类中的哪一类别。Kaggle平台为每一个竞赛都提供了一个榜单（Leaderboard），识别的准确率越高的竞赛者在榜单上的排名越靠前。

split_train_val.py

import os
import numpy as np
import shutil

np.random.seed(2016)

root_train = 'data\train_split'
root_val = 'data\val_split'

root_total = 'data\train'

FishNames = ['ALB', 'BET', 'DOL', 'LAG', 'NoF', 'OTHER', 'SHARK', 'YFT']

nbr_train_samples = 0
nbr_val_samples = 0
# 训练集所占比例
split_proportion = 0.8

for fish in FishNames:
    if not os.path.exists(root_train):
        os.mkdir(root_train)
    if not os.path.exists(root_val):
        os.mkdir(root_val)

    # 建立各个类别的文件夹
    if fish not in os.listdir(root_train):
        os.mkdir(os.path.join(root_train, fish))
    # 当前类别所有的图片
    total_images = os.listdir(os.path.join(root_total, fish))
    # 训练集中图片的数量
    nbr_train = int(len(total_images) * split_proportion)
    # 打乱数据集
    np.random.shuffle(total_images)
    # 划分出训练集
    train_images = total_images[:nbr_train]
    # 划分出测试集
    val_images = total_images[nbr_train:]

    for img in train_images:
        # 从train文件夹将图片拷贝至train_split文件夹下
        source = os.path.join(root_total, fish, img)
        target = os.path.join(root_train, fish, img)
        shutil.copy(source, target)
        nbr_train_samples += 1

    if fish not in os.listdir(root_val):
        os.mkdir(os.path.join(root_val, fish))

    for img in val_images:
        # 从train文件夹将图片拷贝至val_split文件夹下
        source = os.path.join(root_total, fish, img)
        target = os.path.join(root_val, fish, img)
        shutil.copy(source, target)
        nbr_val_samples += 1

print('Finish splitting train and val images!')
print('# training samples: {}, # val samples: {}'.format(nbr_train_samples, nbr_val_samples))

ImageDataGenerator()是keras.preprocessing.image模块中的图片生成器，同时也可以在batch中对数据进行增强，扩充数据集大小，增强模型的泛化能力。比如进行旋转，变形，归一化等等。

参数:

• rescale: rescaling factor. Defaults to None.If None or 0, no rescaling is applied,otherwise we multiply the data by the value provided
• featurewise_center: Boolean. 对输入的图片每个通道减去每个通道对应均值。
• samplewise_center: Boolan. 每张图片减去样本均值, 使得每个样本均值为0。
• featurewise_std_normalization(): Boolean()
• samplewise_std_normalization(): Boolean()
• zca_epsilon(): Default 12-6
• zca_whitening: Boolean. 去除样本之间的相关性
• rotation_range(): 旋转范围
• width_shift_range(): 水平平移范围
• height_shift_range(): 垂直平移范围
• shear_range(): float, 透视变换的范围
• zoom_range(): 缩放范围
• fill_mode: 填充模式, constant, nearest, reflect
• cval: fill_mode == 'constant'的时候填充值
• horizontal_flip(): 水平反转
• vertical_flip(): 垂直翻转
• preprocessing_function(): user提供的处理函数
• data_format(): channels_first或者channels_last
• validation_split(): 多少数据用于验证集

方法:

• apply_transform(x, transform_parameters):根据参数对x进行变换
• fit(x, augment=False, rounds=1, seed=None): 将生成器用于数据x,从数据x中获得样本的统计参数, 只有featurewise_center, featurewise_std_normalization或者zca_whitening为True才需要
• flow(x, y=None, batch_size=32, shuffle=True, sample_weight=None, seed=None, save_to_dir=None, save_prefix='', save_format='png', subset=None) ):按batch_size大小从x,y生成增强数据
• flow_from_directory()从路径生成增强数据,和flow方法相比最大的优点在于不用一次将所有的数据读入内存当中,这样减小内存压力，这样不会发生OOM，血的教训。
• get_random_transform(img_shape, seed=None): 返回包含随机图像变换参数的字典
• random_transform(x, seed=None): 进行随机图像变换, 通过设置seed可以达到同步变换。
• standardize(x): 对x进行归一化

train.py

from keras.applications.inception_v3 import InceptionV3
from keras.layers import Flatten, Dense, AveragePooling2D
from keras.models import Model
from keras.optimizers import RMSprop, SGD
from keras.callbacks import ModelCheckpoint
from keras.preprocessing.image import ImageDataGenerator

# 超参数
learning_rate = 0.0001
img_width = 299
img_height = 299
nbr_train_samples = 3019
nbr_validation_samples = 758
# nbr_epochs = 25
nbr_epochs = 1
batch_size = 32

# 训练集和测试集路径
train_data_dir = 'data\train_split'
val_data_dir = 'data\val_split'

# 类别
FishNames = ['ALB', 'BET', 'DOL', 'LAG', 'NoF', 'OTHER', 'SHARK', 'YFT']

# 加载InceptionV3模型
print('Loading InceptionV3 Weights ...')
InceptionV3_notop = InceptionV3(include_top=False, weights='imagenet', input_tensor=None, input_shape=(299, 299, 3))
# Note that the preprocessing of InceptionV3 is: (x / 255 - 0.5) x 2

# 添加平均池化层和Softmax输出层
print('Adding Average Pooling Layer and Softmax Output Layer ...')
# Shape: (8, 8, 2048)
output = InceptionV3_notop.get_layer(index=-1).output
output = AveragePooling2D((8, 8), strides=(8, 8), name='avg_pool')(output)
output = Flatten(name='flatten')(output)
output = Dense(8, activation='softmax', name='predictions')(output)

InceptionV3_model = Model(InceptionV3_notop.input, output)
print(InceptionV3_model.summary())

# 使用梯度下降优化模型
optimizer = SGD(lr=learning_rate, momentum=0.9, decay=0.0, nesterov=True)
# 模型编译
InceptionV3_model.compile(loss='categorical_crossentropy', optimizer=optimizer, metrics=['accuracy'])
# 自动保存最佳模型
best_model_file = "./weights.h5"
best_model = ModelCheckpoint(best_model_file, monitor='val_acc', verbose=1, save_best_only=True)

# 训练集数据扩增配置
train_datagen = ImageDataGenerator(
    rescale=1. / 255,
    shear_range=0.1,
    zoom_range=0.1,
    rotation_range=10.,
    width_shift_range=0.1,
    height_shift_range=0.1,
    horizontal_flip=True)

# 验证集数据扩增配置
val_datagen = ImageDataGenerator(rescale=1. / 255)

train_generator = train_datagen.flow_from_directory(
    train_data_dir,
    target_size=(img_width, img_height),
    batch_size=batch_size,
    shuffle=True,
    classes=FishNames,
    class_mode='categorical')

validation_generator = val_datagen.flow_from_directory(
    val_data_dir,
    target_size=(img_width, img_height),
    batch_size=batch_size,
    shuffle=True,
    classes=FishNames,
    class_mode='categorical')

InceptionV3_model.fit_generator(
    train_generator,
    samples_per_epoch=nbr_train_samples,
    nb_epoch=nbr_epochs,
    validation_data=validation_generator,
    nb_val_samples=nbr_validation_samples,
    callbacks=[best_model])

predict.py

from keras.models import load_model
import os
from keras.preprocessing.image import ImageDataGenerator
import numpy as np

# 超参数
img_width = 299
img_height = 299
batch_size = 32
nbr_test_samples = 1000

# 类别
FishNames = ['ALB', 'BET', 'DOL', 'LAG', 'NoF', 'OTHER', 'SHARK', 'YFT']

# 模型文件路径
weights_path = os.path.join('weights.h5')

# 测试集路径
test_data_dir = os.path.join('data/test_stg1/')
if not os.path.exists(test_data_dir):
    os.mkdir(test_data_dir)

# 测试集数据生成器
test_datagen = ImageDataGenerator(rescale=1. / 255)

test_generator = test_datagen.flow_from_directory(
    test_data_dir,
    target_size=(img_width, img_height),
    batch_size=batch_size,
    shuffle=False,  # Important !!!
    classes=None,
    class_mode=None)

test_image_list = test_generator.filenames

# 加载模型
print('Loading model and weights from training process ...')
InceptionV3_model = load_model(weights_path)

# 预测
print('Begin to predict for testing data ...')
predictions = InceptionV3_model.predict_generator(test_generator, nbr_test_samples)

# 保存预测结果
np.savetxt(os.path.join('predictions.txt'), predictions)

# 写入提交文件
print('Begin to write submission file ..')
f_submit = open(os.path.join('submit.csv'), 'w')
f_submit.write('image,ALB,BET,DOL,LAG,NoF,OTHER,SHARK,YFT
')
for i, image_name in enumerate(test_image_list):
    pred = ['%.6f' % p for p in predictions[i, :]]
    if i % 100 == 0:
        print('{} / {}'.format(i, nbr_test_samples))
    f_submit.write('%s,%s
' % (os.path.basename(image_name), ','.join(pred)))

f_submit.close()

print('Submission file successfully generated!')

predict_average_augmentation.py

from keras.models import load_model
import os
from keras.preprocessing.image import ImageDataGenerator
import numpy as np

os.environ["CUDA_VISIBLE_DEVICES"] = "0"

# 超参数
img_width = 299
img_height = 299
batch_size = 32
nbr_test_samples = 1000
nbr_augmentation = 5

# 类别
FishNames = ['ALB', 'BET', 'DOL', 'LAG', 'NoF', 'OTHER', 'SHARK', 'YFT']

# 模型文件路径
weights_path = os.path.join('weights.h5')

# 测试集文件路径
test_data_dir = os.path.join('data/test_stg1/')

# 测试集数据生成器
test_datagen = ImageDataGenerator(
    rescale=1. / 255,
    shear_range=0.1,
    zoom_range=0.1,
    width_shift_range=0.1,
    height_shift_range=0.1,
    horizontal_flip=True)

# 加载模型
print('Loading model and weights from training process ...')
InceptionV3_model = load_model(weights_path)

for idx in range(nbr_augmentation):
    print('{}th augmentation for testing ...'.format(idx))
    random_seed = np.random.random_integers(0, 100000)

    test_generator = test_datagen.flow_from_directory(
        test_data_dir,
        target_size=(img_width, img_height),
        batch_size=batch_size,
        shuffle=False,  # Important !!!
        seed=random_seed,
        classes=None,
        class_mode=None)

    test_image_list = test_generator.filenames
    # print('image_list: {}'.format(test_image_list[:10]))
    print('Begin to predict for testing data ...')
    if idx == 0:
        predictions = InceptionV3_model.predict_generator(test_generator, nbr_test_samples)
    else:
        predictions += InceptionV3_model.predict_generator(test_generator, nbr_test_samples)

# 同一个模型，平均多个测试样例
predictions /= nbr_augmentation

# 写入提交文件
print('Begin to write submission file ..')
f_submit = open(os.path.join('submit.csv'), 'w')
f_submit.write('image,ALB,BET,DOL,LAG,NoF,OTHER,SHARK,YFT
')
for i, image_name in enumerate(test_image_list):
    pred = ['%.6f' % p for p in predictions[i, :]]
    if i % 100 == 0:
        print('{} / {}'.format(i, nbr_test_samples))
    f_submit.write('%s,%s
' % (os.path.basename(image_name), ','.join(pred)))

f_submit.close()

print('Submission file successfully generated!')