如何使用tf.keras.utils.image_dataset_from_directory应用kfold交叉验证

Question

我的目标是应用k次交叉验证来训练VGG19模型。为了做到这一点，我使用以下代码从目录中读取图像：

DIR = "/Images"
data_dir = pathlib.Path(os.getcwd() + '\Images')
train_ds = tf.keras.utils.image_dataset_from_directory(
data_dir,
validation_split=0.2,
subset="training",
seed=123,
image_size=(224, 224),
batch_size=32)
val_ds = tf.keras.utils.image_dataset_from_directory(
data_dir,
validation_split=0.2,
subset="validation",
seed=123,
image_size=(224, 224),
batch_size=32)

并且它在没有使用kfold交叉验证的情况下正常工作。但当我想使用K-fold交叉验证时，我必须分别获得train_ds的标签和图像，我找不到解决方案，除非我需要使用另一种方法读取图像。因此，我决定使用ImageDataGenerator和flow_from_directory来读取图像。但据我所知，为了使用flow_from_directory加载图像，我必须在图像中有两个独立的子集traning和test，而在我的情况下，我没有traning和test文件夹。这两种方法都有解决方案吗？

此外，使用第一种方法，即tf.keras.utils.image_dataset_from_directory，将找到的图像数量与flow_from_directory不同。以下是第一种方法的输出：

找到属于4个类的1060个文件。使用848个文件进行培训

以下是第二种方法的输出：

img_gen = tf.keras.preprocessing.image.ImageDataGenerator(
rescale=1.0 / 255,
rotation_range=20,
width_shift_range=0.2,
height_shift_range=0.2,
vertical_flip=True)
Wheat_data = img_gen.flow_from_directory(data_dir,
subset="training",
seed=123)

发现属于4个类别的849个图像

Answer 1

您可以将数据集转换为numpy阵列，它应该像往常一样工作：

import tensorflow as tf
import pathlib
import numpy as np
from sklearn.model_selection import KFold
dataset_url = "https://storage.googleapis.com/download.tensorflow.org/example_images/flower_photos.tgz"
data_dir = tf.keras.utils.get_file('flower_photos', origin=dataset_url, untar=True)
data_dir = pathlib.Path(data_dir)
batch_size = 32
train_ds = tf.keras.utils.image_dataset_from_directory(
data_dir,
validation_split=0.2,
subset="training",
seed=123,
image_size=(180, 180),
batch_size=batch_size)
val_ds = tf.keras.utils.image_dataset_from_directory(
data_dir,
validation_split=0.2,
subset="validation",
seed=123,
image_size=(180, 180),
batch_size=batch_size)
train_images = np.concatenate(list(train_ds.map(lambda x, y:x)))
train_labels = np.concatenate(list(train_ds.map(lambda x, y:y)))
val_images = np.concatenate(list(val_ds.map(lambda x, y:x)))
val_labels = np.concatenate(list(val_ds.map(lambda x, y:y)))
inputs = np.concatenate((train_images, val_images), axis=0)
targets = np.concatenate((train_labels, val_labels), axis=0)
kfold = KFold(n_splits=5, shuffle=True)
for train, test in kfold.split(inputs, targets):

model = tf.keras.Sequential([
tf.keras.layers.Rescaling(1./255, input_shape=(180, 180, 3)),
tf.keras.layers.Conv2D(16, 3, padding='same', activation='relu'),
tf.keras.layers.MaxPooling2D(),
tf.keras.layers.Conv2D(32, 3, padding='same', activation='relu'),
tf.keras.layers.MaxPooling2D(),
tf.keras.layers.Conv2D(64, 3, padding='same', activation='relu'),
tf.keras.layers.MaxPooling2D(),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(128, activation='relu'),
tf.keras.layers.Dense(5)])
model.compile(optimizer='adam',
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
history = model.fit(inputs[train], targets[train],
batch_size=batch_size,
epochs=2)
scores = model.evaluate(inputs[test], targets[test], verbose=0)

或者，您可以使用批量大小为1的tf.keras.utils.image_dataset_from_directory和shuffle=False，但效率不高：

import tensorflow as tf
import pathlib
import numpy as np
from sklearn.model_selection import KFold
dataset_url = "https://storage.googleapis.com/download.tensorflow.org/example_images/flower_photos.tgz"
data_dir = tf.keras.utils.get_file('flower_photos', origin=dataset_url, untar=True)
data_dir = pathlib.Path(data_dir)
batch_size = 1
train_ds = tf.keras.utils.image_dataset_from_directory(
data_dir,
validation_split=0.2,
subset="training",
seed=123,
image_size=(180, 180),
batch_size=batch_size,
shuffle = False)
val_ds = tf.keras.utils.image_dataset_from_directory(
data_dir,
validation_split=0.2,
subset="validation",
seed=123,
image_size=(180, 180),
batch_size=batch_size,
shuffle = False)
ds = train_ds.concatenate(val_ds)
kfold = KFold(n_splits=5, shuffle=True)
for train, test in kfold.split(np.arange(len(ds))):
train = [x+1 for x in train]
test = [x+1 for x in test]
train_ds = tf.data.Dataset.from_tensor_slices([ds.skip(t-1).take(t) for t in train]).flat_map(lambda x: x).map(lambda x, y: (x[0, ...], y[0, ...]))
test_ds = tf.data.Dataset.from_tensor_slices([ds.skip(t-1).take(t) for t in test]).flat_map(lambda x: x).map(lambda x, y: (x[0, ...], y[0, ...]))
train_ds = train_ds.take(len(train)).batch(64, drop_remainder=True)
test_ds = test_ds.take(len(test)).batch(64, drop_remainder=True)

model = tf.keras.Sequential([
tf.keras.layers.Rescaling(1./255, input_shape=(180, 180, 3)),
tf.keras.layers.Conv2D(16, 3, padding='same', activation='relu'),
tf.keras.layers.MaxPooling2D(),
tf.keras.layers.Conv2D(32, 3, padding='same', activation='relu'),
tf.keras.layers.MaxPooling2D(),
tf.keras.layers.Conv2D(64, 3, padding='same', activation='relu'),
tf.keras.layers.MaxPooling2D(),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(128, activation='relu'),
tf.keras.layers.Dense(5)])
model.compile(optimizer='adam',loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
history = model.fit(train_ds,
epochs=2)
scores = model.evaluate(test_ds, verbose=0)

另一种选择是使用字典来存储索引和张量：

#...
ds = train_ds.concatenate(val_ds)
lookup_images = {}
lookup_labels = {}
for i, (x, y) in enumerate(ds):
lookup_images[i] = x
lookup_labels[i] = y
kfold = KFold(n_splits=5, shuffle=True)
for train, test in kfold.split(np.arange(len(ds))):
images_train = np.concatenate(list(map(lookup_images.get, train)))
labels_train = np.concatenate(list(map(lookup_labels.get, train)))
images_test = np.concatenate(list(map(lookup_images.get, test)))
labels_test = np.concatenate(list(map(lookup_labels.get, test)))
model = tf.keras.Sequential([
tf.keras.layers.Rescaling(1./255, input_shape=(180, 180, 3)),
tf.keras.layers.Conv2D(16, 3, padding='same', activation='relu'),
tf.keras.layers.MaxPooling2D(),
tf.keras.layers.Conv2D(32, 3, padding='same', activation='relu'),
tf.keras.layers.MaxPooling2D(),
tf.keras.layers.Conv2D(64, 3, padding='same', activation='relu'),
tf.keras.layers.MaxPooling2D(),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(128, activation='relu'),
tf.keras.layers.Dense(5)])
model.compile(optimizer='adam',
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
history = model.fit(images_train, labels_train, epochs=2)
scores = model.evaluate(images_test, labels_test, verbose=0)