Question

我是机器学习的初学者。我正在制作一个使用keras从叶子图像中检测有害生物的CNN模型。在训练数据期间，内存超出限制，我无法训练。我曾经使用过kaggle / Google Collab，但是在这两者中我都有内存探测功能。建议我使用数据生成器，但是在尝试执行此操作时，我无法执行此操作。是否有其他方法可以有效地训练或是否使用了数据生成器的任何示例（已经看到了很多示例，但是添加时遇到了问题。

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
        )

您也可以在此link中找到代码。

Answer 1

这里的问题是，您已将完整的数据加载到工作空间中，这将占用大量内存，并在进程上创建大量额外负载。

您可以使用带有data-generator的{{1}}和flow_from_directory来定义augmentation和pre-processing管道。这样做的好处是工作空间没有额外的数据负载。您可以找到示例here。

随时提问。

Answer 2

您可以将图像转换为张量流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/datasets http://www.machinelearninguru.com/deep_learning/tensorflow/basics/tfrecord/tfrecord.html

如何有效地训练具有大图像数据集的CNN模型

2 个答案: