摘要：分類器項(xiàng)目自定義數(shù)據(jù)讀入在照著官網(wǎng)的敲了一遍分類器項(xiàng)目的代碼后，運(yùn)行倒是成功了，結(jié)果也不錯(cuò)。

在照著Tensorflow官網(wǎng)的demo敲了一遍分類器項(xiàng)目的代碼后，運(yùn)行倒是成功了，結(jié)果也不錯(cuò)。但是最終還是要訓(xùn)練自己的數(shù)據(jù)，所以嘗試準(zhǔn)備加載自定義的數(shù)據(jù)，然而demo中只是出現(xiàn)了fashion_mnist.load_data()并沒有詳細(xì)的讀取過程，隨后我又找了些資料，把讀取的過程記錄在這里。
首先提一下需要用到的模塊：

import os

import keras
import matplotlib.pyplot as plt
from PIL import Image
from keras.preprocessing.image import ImageDataGenerator
from sklearn.model_selection import train_test_split

圖片分類器項(xiàng)目，首先確定你要處理的圖片分辨率將是多少，這里的例子為30像素：

IMG_SIZE_X = 30
IMG_SIZE_Y = 30

其次確定你圖片的方式目錄：

image_path = r"D:ProjectsImageClassifierdataset"
path = ".data"
# 你也可以使用相對(duì)路徑的方式
# image_path =os.path.join(path, "set")

目錄下的結(jié)構(gòu)如下：

相應(yīng)的label.txt如下：

動(dòng)漫
風(fēng)景
美女
物語
櫻花

接下來是接在labels.txt，如下：

label_name = "labels.txt"
label_path = os.path.join(path, label_name)
class_names = np.loadtxt(label_path, type(""))

這里簡便起見，直接利用了numpy的loadtxt函數(shù)直接加載。

之后便是正式處理圖片數(shù)據(jù)了，注釋就寫在里面了：

re_load = False
re_build = False
# re_load = True
re_build = True

data_name = "data.npz"
data_path = os.path.join(path, data_name)
model_name = "model.h5"
model_path = os.path.join(path, model_name)

count = 0

# 這里判斷是否存在序列化之后的數(shù)據(jù)，re_load是一個(gè)開關(guān)，是否強(qiáng)制重新處理，測試用，可以去除。
if not os.path.exists(data_path) or re_load:
    labels = []
    images = []
    print("Handle images")
    # 由于label.txt是和圖片防止目錄的分類目錄一一對(duì)應(yīng)的，即每個(gè)子目錄的目錄名就是labels.txt里的一個(gè)label，所以這里可以通過讀取class_names的每一項(xiàng)去拼接path后讀取
    for index, name in enumerate(class_names):
        # 這里是拼接后的子目錄path
        classpath = os.path.join(image_path, name)
        # 先判斷一下是否是目錄
        if not os.path.isdir(classpath):
            continue
        # limit是測試時(shí)候用的這里可以去除
        limit = 0
        for image_name in os.listdir(classpath):
            if limit >= max_size:
                break
            # 這里是拼接后的待處理的圖片path
            imagepath = os.path.join(classpath, image_name)
            count = count + 1
            limit = limit + 1
            # 利用Image打開圖片
            img = Image.open(imagepath)
            # 縮放到你最初確定要處理的圖片分辨率大小
            img = img.resize((IMG_SIZE_X, IMG_SIZE_Y))
            # 轉(zhuǎn)為灰度圖片，這里彩色通道會(huì)干擾結(jié)果，并且會(huì)加大計(jì)算量
            img = img.convert("L")
            # 轉(zhuǎn)為numpy數(shù)組
            img = np.array(img)
            # 由（30，30）轉(zhuǎn)為（1，30，30）（即`channels_first`），當(dāng)然你也可以轉(zhuǎn)換為（30，30，1）（即`channels_last`）但為了之后預(yù)覽處理后的圖片方便這里采用了（1，30，30）的格式存放
            img = np.reshape(img, (1, IMG_SIZE_X, IMG_SIZE_Y))
            # 這里利用循環(huán)生成labels數(shù)據(jù)，其中存放的實(shí)際是class_names中對(duì)應(yīng)元素的索引
            labels.append([index])
            # 添加到images中，最后統(tǒng)一處理
            images.append(img)
            # 循環(huán)中一些狀態(tài)的輸出，可以去除
            print("{} class: {} {} limit: {} {}"
                  .format(count, index + 1, class_names[index], limit, imagepath))
    # 最后一次性將images和labels都轉(zhuǎn)換成numpy數(shù)組
    npy_data = np.array(images)
    npy_labels = np.array(labels)
    # 處理數(shù)據(jù)只需要一次，所以我們選擇在這里利用numpy自帶的方法將處理之后的數(shù)據(jù)序列化存儲(chǔ)
    np.savez(data_path, x=npy_data, y=npy_labels)
    print("Save images by npz")
else:
    # 如果存在序列化號(hào)的數(shù)據(jù)，便直接讀取，提高速度
    npy_data = np.load(data_path)["x"]
    npy_labels = np.load(data_path)["y"]
    print("Load images by npz")
image_data = npy_data
labels_data = npy_labels

到了這里原始數(shù)據(jù)的加工預(yù)處理便已經(jīng)完成，只需要最后一步，就和demo中fashion_mnist.load_data()返回的結(jié)果一樣了。代碼如下：

# 最后一步就是將原始數(shù)據(jù)分成訓(xùn)練數(shù)據(jù)和測試數(shù)據(jù)
train_images, test_images, train_labels, test_labels = 
    train_test_split(image_data, labels_data, test_size=0.2, random_state=6)

這里將相關(guān)信息打印的方法也附上：

print("_________________________________________________________________")
print("%-28s %-s" % ("Name", "Shape"))
print("=================================================================")
print("%-28s %-s" % ("Image Data", image_data.shape))
print("%-28s %-s" % ("Labels Data", labels_data.shape))
print("=================================================================")

print("Split train and test data,p=%")
print("_________________________________________________________________")
print("%-28s %-s" % ("Name", "Shape"))
print("=================================================================")
print("%-28s %-s" % ("Train Images", train_images.shape))
print("%-28s %-s" % ("Test Images", test_images.shape))
print("%-28s %-s" % ("Train Labels", train_labels.shape))
print("%-28s %-s" % ("Test Labels", test_labels.shape))
print("=================================================================")

之后別忘了歸一化喲：

print("Normalize images")
train_images = train_images / 255.0
test_images = test_images / 255.0

最后附上讀取自定義數(shù)據(jù)的完整代碼：

import os

import keras
import matplotlib.pyplot as plt
from PIL import Image
from keras.layers import *
from keras.models import *
from keras.optimizers import Adam
from keras.preprocessing.image import ImageDataGenerator
from sklearn.model_selection import train_test_split

os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"
# 支持中文
plt.rcParams["font.sans-serif"] = ["SimHei"]  # 用來正常顯示中文標(biāo)簽
plt.rcParams["axes.unicode_minus"] = False  # 用來正常顯示負(fù)號(hào)
re_load = False
re_build = False
# re_load = True
re_build = True
epochs = 50
batch_size = 5
count = 0
max_size = 2000000000
IMG_SIZE_X = 30
IMG_SIZE_Y = 30
np.random.seed(9277)
image_path = r"D:ProjectsImageClassifierdataset"
path = ".data"
data_name = "data.npz"
data_path = os.path.join(path, data_name)
model_name = "model.h5"
model_path = os.path.join(path, model_name)
label_name = "labels.txt"
label_path = os.path.join(path, label_name)
class_names = np.loadtxt(label_path, type(""))
print("Load class names")
if not os.path.exists(data_path) or re_load:
    labels = []
    images = []
    print("Handle images")
    for index, name in enumerate(class_names):
        classpath = os.path.join(image_path, name)
        if not os.path.isdir(classpath):
            continue
        limit = 0
        for image_name in os.listdir(classpath):
            if limit >= max_size:
                break
            imagepath = os.path.join(classpath, image_name)
            count = count + 1
            limit = limit + 1
            img = Image.open(imagepath)
            img = img.resize((30, 30))
            img = img.convert("L")
            img = np.array(img)
            img = np.reshape(img, (1, 30, 30))
            # img = skimage.io.imread(imagepath, as_grey=True)
            # if img.shape[2] != 3:
            #     print("{} shape is {}".format(image_name, img.shape))
            #     continue
            # data = transform.resize(img, (IMG_SIZE_X, IMG_SIZE_Y))
            labels.append([index])
            images.append(img)
            print("{} class: {} {} limit: {} {}"
                  .format(count, index + 1, class_names[index], limit, imagepath))
    npy_data = np.array(images)
    npy_labels = np.array(labels)
    np.savez(data_path, x=npy_data, y=npy_labels)
    print("Save images by npz")
else:
    npy_data = np.load(data_path)["x"]
    npy_labels = np.load(data_path)["y"]
    print("Load images by npz")
image_data = npy_data
labels_data = npy_labels
print("_________________________________________________________________")
print("%-28s %-s" % ("Name", "Shape"))
print("=================================================================")
print("%-28s %-s" % ("Image Data", image_data.shape))
print("%-28s %-s" % ("Labels Data", labels_data.shape))
print("=================================================================")
train_images, test_images, train_labels, test_labels = 
    train_test_split(image_data, labels_data, test_size=0.2, random_state=6)
print("Split train and test data,p=%")
print("_________________________________________________________________")
print("%-28s %-s" % ("Name", "Shape"))
print("=================================================================")
print("%-28s %-s" % ("Train Images", train_images.shape))
print("%-28s %-s" % ("Test Images", test_images.shape))
print("%-28s %-s" % ("Train Labels", train_labels.shape))
print("%-28s %-s" % ("Test Labels", test_labels.shape))
print("=================================================================")

# 歸一化
# 我們將這些值縮小到 0 到 1 之間，然后將其饋送到神經(jīng)網(wǎng)絡(luò)模型。為此，將圖像組件的數(shù)據(jù)類型從整數(shù)轉(zhuǎn)換為浮點(diǎn)數(shù)，然后除以 255。以下是預(yù)處理圖像的函數(shù)：
# 務(wù)必要以相同的方式對(duì)訓(xùn)練集和測試集進(jìn)行預(yù)處理：
print("Normalize images")
train_images = train_images / 255.0
test_images = test_images / 255.0