Keras - .flow_from_directory（目录）

Posted 2023-03-12

【中文标题】Keras - .flow_from_directory（目录）

【英文标题】：Keras - .flow_from_directory(directory)

【发布时间】：2018-08-07 14:39:56

【问题描述】：

我正在尝试使用.flow_from_directory(directory) 运行带有 cifar10 数据集的 Resnet 示例。以下代码如下：

from __future__ import print_function
from keras.datasets import cifar10
from keras.preprocessing.image import ImageDataGenerator
from keras.utils import np_utils
from keras.callbacks import ReduceLROnPlateau, CSVLogger, EarlyStopping

import numpy as np
import resnet
import os
import cv2
import csv
#import keras 

os.environ["CUDA_VISIBLE_DEVICES"] = "1"


# input image dimensions
img_rows, img_cols = 32, 32
# The CIFAR10 images are RGB.
img_channels = 3
nb_classes = 10


train_datagen = ImageDataGenerator(
        rescale=1./255,
        shear_range=0,
        zoom_range=0,
        horizontal_flip=False,
        width_shift_range=0.1,  # randomly shift images horizontally (fraction of total width)
        height_shift_range=0.1)  # randomly shift images vertically (fraction of total height))

test_datagen = ImageDataGenerator(rescale=1./255)  

train_generator = train_datagen.flow_from_directory(
        '/home/datasets/cifar10/train',
        target_size=(32, 32),
        batch_size=32,
        shuffle=False)

validation_generator = test_datagen.flow_from_directory(
        '/home/datasets/cifar10/test',
        target_size=(32, 32),
        batch_size=32,
        shuffle=False)

model = resnet.ResnetBuilder.build_resnet_18((img_channels, img_rows, img_cols), nb_classes)
model.compile(loss='categorical_crossentropy',
              optimizer='adam',
              metrics=['accuracy'])

model.fit_generator(
        train_generator,
        steps_per_epoch=500,
        epochs=50,
        validation_data=validation_generator,
        validation_steps=250)

但是，我获得了以下准确度值。

500/500 [==============================] - 22s - loss: 0.8139 - acc: 0.9254 - val_loss: 12.7198 - val_acc: 0.1250
Epoch 2/50
500/500 [==============================] - 19s - loss: 1.0645 - acc: 0.8856 - val_loss: 8.4179 - val_acc: 0.0560
Epoch 3/50
500/500 [==============================] - 19s - loss: 2.1014 - acc: 0.7492 - val_loss: 10.7770 - val_acc: 0.0956
Epoch 4/50
500/500 [==============================] - 19s - loss: 1.6806 - acc: 0.7772 - val_loss: 6.1023 - val_acc: 0.0741
Epoch 5/50
500/500 [==============================] - 19s - loss: 1.1798 - acc: 0.8669 - val_loss: 6.9016 - val_acc: 0.1253
Epoch 6/50
500/500 [==============================] - 19s - loss: 1.5448 - acc: 0.8369 - val_loss: 3.6371 - val_acc: 0.0370
Epoch 7/50
500/500 [==============================] - 19s - loss: 1.3763 - acc: 0.8599 - val_loss: 4.8012 - val_acc: 0.1204
Epoch 8/50
500/500 [==============================] - 19s - loss: 1.0186 - acc: 0.8891 - val_loss: 6.8395 - val_acc: 0.0912
Epoch 9/50
500/500 [==============================] - 19s - loss: 0.9477 - acc: 0.9081 - val_loss: 10.4287 - val_acc: 0.1253
Epoch 10/50
500/500 [==============================] - 19s - loss: 1.0689 - acc: 0.8686 - val_loss: 7.9931 - val_acc: 0.1253

我正在使用来自 link 的 Resnet。我尝试了许多示例来解决问题，包括官方文档中的示例。但是，我无法解决问题。训练精度正在发生变化，但 val 精度有些恒定。 有人能指出问题吗

【参考方案1】：

根据 Keras 文档。

flow_from_directory(directory)，描述：获取目录的路径，并生成批量增强/规范化的数据。在无限循环中无限产生批次。

使用shuffle = False，它会无限期地占用同一批次。导致这些精度值。我更改了shuffle = True，现在可以正常使用了。