def data_augment(x_train,
y_train,
batch_size,
val_split=0.25,
task='classification',
num_classes=None):
h, w = x_train.shape[1:3]
# Shuffle training data and split some for validation
seed = 1
m = len(x_train)
np.random.seed(seed)
index = np.arange(0, m)
random.shuffle(index)
x_train = x_train[index]
y_train = y_train[index]
split = int(m * (1 - val_split))
x_train, x_val = x_train[:split], x_train[split:]
y_train, y_val = y_train[:split], y_train[split:]
#Use shiftγflip and rotation to augment data
if task == 'classification':
datagen = ImageDataGenerator(rotation_range=10,
width_shift_range=0.1,
height_shift_range=0.1,
validation_split=val_split,
horizontal_flip=True)
train_flow = datagen.flow(x_train, y_train, batch_size=batch_size)
elif task == 'detection_single':
ssd_data_augmentation = SSDDataAugmentation(img_height=h, img_width=w)
normalize = Normalize(h, w)
gen = DataGenerator()
box = y_train[:, -4:] * [w, h, w, h]
class_id = np.argmax(y_train[:, :-4, np.newaxis], axis=1)
labels = np.hstack([class_id, box])
labels = labels[:, np.newaxis, :].tolist()
gen.images = x_train * 255
gen.labels = labels
gen.dataset_size = len(labels)
gen.dataset_indices = np.arange(gen.dataset_size)
gen.filenames = ['x' for i in range(gen.dataset_size)]
train_flow = gen.generate(
batch_size=batch_size,
shuffle=True,
transformations=[ssd_data_augmentation, normalize],
label_encoder=PetLabelEncoder(num_classes),
returns={'processed_images', 'encoded_labels'},
keep_images_without_gt=False)
datagen = ImageDataGenerator()
val_flow = datagen.flow(x_val, y_val, batch_size=batch_size)
return (train_flow, val_flow)
# option in the constructor, because in that cas the images are in memory already anyway. If you don't
# want to create HDF5 datasets, comment out the subsequent two function calls.
# train_dataset.create_hdf5_dataset(file_path='data/image_data/dataset_pascal_voc_07+12_trainval.h5',
# resize=False,
# variable_image_size=True,
# verbose=True)
# val_dataset.create_hdf5_dataset(file_path='data/image_data/dataset_pascal_voc_07_test.h5',
# resize=False,
# variable_image_size=True,
# verbose=True)
train_dataset.hdf5_dataset = h5py.File(
'data/image_data/dataset_pascal_voc_07+12_trainval.h5', 'r')
train_dataset.hdf5_dataset_path = 'data/image_data/dataset_pascal_voc_07+12_trainval.h5'
train_dataset.dataset_size = len(train_dataset.hdf5_dataset['images'])
train_dataset.dataset_indices = np.arange(train_dataset.dataset_size,
dtype=np.int32)
val_dataset.hdf5_dataset = h5py.File(
'data/image_data/dataset_pascal_voc_07_test.h5', 'r')
val_dataset.hdf5_dataset_path = 'data/image_data/dataset_pascal_voc_07_test.h5'
val_dataset.dataset_size = len(val_dataset.hdf5_dataset['images'])
val_dataset.dataset_indices = np.arange(val_dataset.dataset_size,
dtype=np.int32)
# 3: Set the batch size.
batch_size = 8 # Change the batch size if you like, or if you run into GPU memory issues.
# 4: Set the image transformations for pre-processing and data augmentation options.
