如何使用 keras 和 tensorflow 的 ImageDataGenerator 执行数据增强

Question

I'm having a hard time understanding how to implement data augmentation with tensorflow. I have a dataset (images), that is divided into two subsets; training and testing. After I have called the ImageDataGenerator functions with various parameters, do I need to save the images (like using the flow() ) or will Tensorflow augment my data while the model is training?

Here is my implemented code:

# necessary imports

train_datagen = ImageDataGenerator(
    rescale=1. / 255,
    width_shift_range=0.2,
    height_shift_range=0.2,
    shear_range=0.2,
    zoom_range=0.2,
    brightness_range=(0.3, 1.0),
    horizontal_flip=True,
    vertical_flip=True,
    fill_mode='nearest',
    validation_split=0.2
)

training_directory = '/tmp/dataset/training'
testing_directory = '/tmp/dataset/testing'

training_set = train_datagen.flow_from_directory(
    training_directory,
    target_size=(150, 150),
    batch_size=32,
    class_mode='binary',
    subset='training'
)

test_set = train_datagen.flow_from_directory(
    testing_directory,
    target_size=(150, 150),
    batch_size=32,
    class_mode='binary',
    subset='validation'
)

# creating a sequential model
...
# fitting and data plotting

The model summary:

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
conv2d (Conv2D)              (None, 148, 148, 32)      896
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 74, 74, 32)        0
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 72, 72, 64)        18496
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 36, 36, 64)        0
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 34, 34, 128)       73856
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 17, 17, 128)       0
_________________________________________________________________
dropout (Dropout)            (None, 17, 17, 128)       0
_________________________________________________________________
flatten (Flatten)            (None, 36992)             0
_________________________________________________________________
dense (Dense)                (None, 512)               18940416
_________________________________________________________________
dense_1 (Dense)              (None, 1)                 513
=================================================================
Total params: 19,034,177
Trainable params: 19,034,177
Non-trainable params: 0
_________________________________________________________________

Answer 1

You don't have to save new data.

When calling flow method, data are augmented on the fly and are served as input to the model.

So, data are being generated in real time and feed immediately into your model.

Answer 2

You don't need to save the data. The augmented data (train/test) is fed directly into the model for training or evaluation steps using the train and test data generators.

Here is your code updated with all the steps using the created data generators train_generator and test_generator .

 datagenerator = ImageDataGenerator(
    rescale=1. / 255,
    width_shift_range=0.2,
    height_shift_range=0.2,
    shear_range=0.2,
    zoom_range=0.2,
    brightness_range=(0.3, 1.0),
    horizontal_flip=True,
    vertical_flip=True,
    fill_mode='nearest',
    validation_split=0.2
)
 
training_directory = '/tmp/dataset/training'
testing_directory = '/tmp/dataset/testing'

train_generator = datagenerator.flow_from_directory(
    training_directory,
    target_size=(150, 150),
    batch_size=32,
    class_mode='binary',
    subset='training'
)

test_generator = datagenerator.flow_from_directory(
    testing_directory,
    target_size=(150, 150),
    batch_size=32,
    class_mode='binary',
    subset='validation'
)

# Build and compile the model
....
# Get the number of steps per epoch for each of the data generators
train_steps_per_epoch = train_generator.n // train_generator.batch_size
test_steps_per_epoch = test_generator.n // test_generator.batch_size

# Fit the model
model.fit_generator(train_generator, steps_per_epoch=train_steps_per_epoch, epochs=your_nepochs)

# Evaluate the model
model.evaluate_generator(test_generator, steps=test_steps_per_epoch)