[KERAS] ImageDataGenerator images and masks

[KERAS] ImageDataGenerator images and masks

import numpy as np

from keras.preprocessing.image import ImageDataGenerator
import imgaug as ia
import imgaug.augmenters as iaa
ia.seed(1)

train_imgs = np.load('train_imgs.npy')
train_masks = np.load('train_masks.npy')
val_imgs = np.load('val_imgs.npy')
val_masks = np.load('val_masks.npy')

seq = iaa.SomeOf((0, None), [
#     iaa.CropAndPad(percent=(-0.25, 0.25))
    iaa.HorizontalFlip(1), # horizontally flip the images
    iaa.VerticalFlip(1), # vertical flip the images
    iaa.Rot90([1]), # 90
    iaa.Rot90([2]), # 180
#     iaa.Rot90([3]), # 270
#     iaa.Clouds(),
#     iaa.Fog(),
    iaa.Crop(px=(5, 16)), # 128 crop images from each side by 0 to 16px (randomly chosen)\n",
    iaa.Affine(
        scale={"x": (1, 1.4), "y": (1, 1.4)},
        #translate_percent={\"x\": (-0.2, 0.2), \"y\": (-0.2, 0.2)},
        rotate=(-5, 5),
        shear=(-10, 10)
    ),
    iaa.AddToHue((-20, 20)),
    iaa.AdditiveGaussianNoise(scale=0.02 * 255),
    iaa.GammaContrast((0.6, 1.7)),   
])
seq_det = seq.to_deterministic()

image_datagen_train = ImageDataGenerator(rescale=1./255, preprocessing_function=seq_det.augment_image)
mask_datagen_train = ImageDataGenerator(preprocessing_function=seq_det.augment_image)

image_datagen_val = ImageDataGenerator(rescale=1./255)
mask_datagen_val = ImageDataGenerator()

'''
data_gen_args = dict(rescale=1./255,
                     width_shift_range=0.1,
                     height_shift_range=0.1,
                     zoom_range=0.2,
                     rotation_range=5., 
                     horizontal_flip=True,
                     vertical_flip=True
                    )
'''

# image_datagen_train = ImageDataGenerator(**data_gen_args)
# mask_datagen_train = ImageDataGenerator(**data_gen_args)

# Provide the same seed and keyword arguments to the fit and flow methods
seed = 42
image_datagen_train.fit(train_imgs, augment=True, seed=seed)
mask_datagen_train.fit(train_masks, augment=True, seed=seed)

image_datagen_val.fit(test_imgs, augment=True, seed=seed)
mask_datagen_val.fit(test_masks, augment=True, seed=seed)

train_batch_size = 32
val_batch_size = 32

image_generator_train = image_datagen_train.flow(train_imgs, batch_size=train_batch_size, shuffle=True, seed=seed)
mask_generator_train = mask_datagen_train.flow(train_masks, batch_size=train_batch_size, shuffle=True, seed=seed)
# combine generators into one which yields image and masks
train_generator = zip(image_generator_train, mask_generator_train)

image_generator_val = image_datagen_val.flow(test_imgs, batch_size=val_batch_size, shuffle=False)
mask_generator_val = mask_datagen_val.flow(test_masks, batch_size=val_batch_size, shuffle=False)
val_generator = zip(image_generator_val, mask_generator_val)

model = some_model()

path = f'snapshots/'

if not os.path.exists(path):
    os.makedirs(path)

earlystopper = EarlyStopping(patience=15, verbose=1)

filepath = path + 'e{epoch:02d}_b64_val{val_loss:.4f}_iou{iou:.4f}_iouv{val_metric:.4f}.h5'

callbacks = [
        ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=10, min_lr=1e-9, epsilon=0.00001, verbose=1, mode='min'),
        # EarlyStopping(monitor='val_loss', patience=patience, verbose=0),
        ModelCheckpoint(filepath, monitor='val_loss', save_best_only=True, verbose=0),
    ]

results = model.fit_generator(train_generator,
                    epochs=100,
                    steps_per_epoch=train_imgs.shape[0] // train_batch_size,
                    validation_steps=val_imgs.shape[0] // val_batch_size,
                    validation_data=val_generator,
                    verbose=1, callbacks=callbacks
                   )