如何有效地训练具有大图像数据集的 CNN 模型

Question

我是机器学习的初学者。我正在使用 keras 制作一个 CNN 模型来检测叶子图像中的害虫。在训练数据期间，内存超出，我无法训练。我使用过 kaggle/Google Collab，但两者都有记忆问题。 有人建议我使用数据生成器，但是在尝试这样做时，我无法做到。是否有任何其他有效训练的方法或是否使用数据生成器的任何示例（看过很多示例但添加时遇到问题。

import numpy as np
import pickle
import cv2
from os import listdir
from sklearn.preprocessing import LabelBinarizer
from keras.models import Sequential
from keras.layers.normalization import BatchNormalization
from keras.layers.convolutional import Conv2D
from keras.layers.convolutional import MaxPooling2D
from keras.layers.core import Activation, Flatten, Dropout, Dense
from keras import backend as K
from keras.preprocessing.image import ImageDataGenerator
from keras.optimizers import Adam
from keras.preprocessing import image
from keras.preprocessing.image import img_to_array
from sklearn.preprocessing import MultiLabelBinarizer
from sklearn.model_selection import train_test_split
import matplotlib.pyplot as plt



EPOCHS = 25
INIT_LR = 1e-3
BS = 32
default_image_size = tuple((256, 256))
image_size = 0
directory_root = 'PlantVillage/'
width=256
height=256
depth=3

#Function to convert images to array

def convert_image_to_array(image_dir):
    try:
        image = cv2.imread(image_dir)
        if image is not None:
            image = cv2.resize(image,default_image_size)
            return img_to_array(image)
        else:
            return np.array([])
    except Exception as e:
        print(f"Error : {e}")
        return None


image_list, label_list = [], []
try:
    print("[INFO] Loading images ...")
    root_dir = listdir(directory_root)

    #Looping inside root_directory
    for directory in root_dir :
        # remove .DS_Store from list
        if directory == ".DS_Store" :
            root_dir.remove(directory)

    for plant_folder in root_dir :
        plant_disease_folder_list = listdir(f"{directory_root}/{plant_folder}")
        print(f"[INFO] Processing {plant_folder} ...")

        #looping in images
        for disease_folder in plant_disease_folder_list :
            # remove .DS_Store from list
            if disease_folder == ".DS_Store" :
                plant_disease_folder_list.remove(plant_folder)

        #If all data taken not able to train
        for images in plant_disease_folder_list:
            image_directory = f"{directory_root}/{plant_folder}/{images}"
            if image_directory.endswith(".jpg") == True or image_directory.endswith(".JPG") == True:
                image_list.append(convert_image_to_array(image_directory))
                label_list.append(plant_folder)



    print("[INFO] Image loading completed")  
except Exception as e:
    print(f"Error : {e}")



#Get Size of Processed Image
image_size = len(image_list)

#Converting multi-class labels to binary labels(belong or doesnot belong in the class)

label_binarizer = LabelBinarizer()
image_labels = label_binarizer.fit_transform(label_list)

#Saving label binarizer instance using pickle
pickle.dump(label_binarizer,open('label_transform.pkl','wb'))
n_classes = len(label_binarizer.classes_)

print(label_binarizer.classes_)

#Normalizing image from [0,255] to [0,1]
np_image_list = np.array(image_list, dtype = np.float)/255.0

#Splitting data into training and test set 80:20
print('Splitting data to train,test')
x_train, x_test, y_train, y_test = train_test_split(np_image_list, image_labels, test_size=0.2, random_state = 42)

#Creating image generator object which performs random rotations, shifs,flips,crops,sheers
aug = ImageDataGenerator(
        rotation_range = 25, width_shift_range=0.1,
        height_shift_range=0.1, shear_range=0.2,
        zoom_range=0.2, horizontal_flip = True,
        fill_mode="nearest")

model = Sequential()

inputShape = (height, width, depth)
chanDim = -1
if K.image_data_format() == "channels_first":
    inputShape = (depth, height, width)
    chanDim = 1

model.add(Conv2D(32, (3, 3), padding="same",input_shape=inputShape))
model.add(Activation("relu"))
model.add(BatchNormalization(axis=chanDim))
model.add(MaxPooling2D(pool_size=(3, 3)))
model.add(Dropout(0.25))
model.add(Conv2D(64, (3, 3), padding="same"))
model.add(Activation("relu"))
model.add(BatchNormalization(axis=chanDim))
model.add(Conv2D(64, (3, 3), padding="same"))
model.add(Activation("relu"))
model.add(BatchNormalization(axis=chanDim))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Conv2D(128, (3, 3), padding="same"))
model.add(Activation("relu"))
model.add(BatchNormalization(axis=chanDim))
model.add(Conv2D(128, (3, 3), padding="same"))
model.add(Activation("relu"))
model.add(BatchNormalization(axis=chanDim))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Flatten())

model.add(Dense(32))
model.add(Activation("relu"))
model.add(BatchNormalization())
model.add(Dropout(0.5))
model.add(Dense(n_classes))
model.add(Activation("softmax"))

#model.summary()
#Compiling the CNN
opt = Adam(lr= INIT_LR, decay= INIT_LR/EPOCHS)

#distribution
model.compile(loss="binary_crossentropy", optimizer = opt, metrics=["accuracy"])

#training the Model
print("Training Model.....")

history = model.fit_generator(
        aug.flow(x_train, y_train, batch_size= BS),
        validation_data = (x_test, y_test),
        steps_per_epoch = len(x_train) // BS,
        epochs = EPOCHS, verbose = 1
        )

您也可以在 link 中找到代码。

Answer 1

这里的问题是您已经在工作区中加载了完整的数据，这会占用大量内存并在进程上产生大量额外负载。

您可以使用data-generator 和flow_from_directory，它允许您定义augmentation 和pre-processing 管道以及data on the fly。这里的优点是工作区没有额外的数据负载。你可以找到一个例子here。

欢迎提问。

Answer 2

您可以将图像转换为tensorflow可以理解的二进制格式，称为“tfrecord”格式。

请参考以下链接

https://www.tensorflow.org/guide/datasetshttp://www.machinelearninguru.com/deep_learning/tensorflow/basics/tfrecord/tfrecord.html