def train_model(self):
from keras_video import VideoFrameGenerator
from keras.preprocessing.image import ImageDataGenerator
from keras_video.utils import show_sample
classes = ['Real', 'Deepfake']
classes.sort()
size = (224, 224)
channels = 3
nbframe = 10
bs = 2
glob_pattern = 'Dataset_Face_Extracted_redo/{classname}/*.mp4'
data_aug = ImageDataGenerator(zoom_range=.1,
horizontal_flip=True,
rotation_range=8,
width_shift_range=.2,
height_shift_range=.2)
train_generator = VideoFrameGenerator(classes=classes,
glob_pattern=glob_pattern,
nb_frames=nbframe,
split=.25,
shuffle=True,
batch_size=bs,
target_shape=size,
nb_channel=channels,
transformation=data_aug,
use_frame_cache=True)
valid = train_generator.get_validation_generator()
show_sample(train_generator)
input_shape = (nbframe, ) + size + (channels, )
model = self.timeModel(input_shape)
model.compile(Adam(0.001), 'categorical_crossentropy', metrics=['acc'])
epochs = 50
callbacks = [
ReduceLROnPlateau(verbose=1),
ModelCheckpoint(
filepath=os.path.join('dataLSTM', 'checkpoints', str(model) + \
'.{epoch:03d}-{val_loss:.3f}.hdf5'),
verbose=1, save_best_only=True
)
]
model.fit_generator(train_generator,
validation_data=valid,
verbose=1,
epochs=epochs,
callbacks=callbacks)
def create_train_dataset(self, model, prop_val_dataset = 0.33, do_data_aug = True, batch_size = 8, preprocess_input = None):
""" Retrieve the train, validation and test datasets
model: buildModel class
A keras model created by buildModel class.
prop_val_dataset: float, Optional
A float value between 0 and 1 to split the training set into train and
validation. Default value 0.33.
do_data_aug: boolean, Optional
Whether or not do data augmentation to the frames. Default value True.
"""
_, time_step, *size, channels = model.input_shape
self.size = tuple(size)
self.channels = channels
self.time_step = time_step
# for data augmentation
if do_data_aug:
data_aug = keras.preprocessing.image.ImageDataGenerator(
preprocessing_function = preprocess_input,
zoom_range=.1,
horizontal_flip = True,
rotation_range = 8,
width_shift_range = .2,
height_shift_range = .2)
else:
data_aug = None
# Create video frame generator
train_dataset = VideoFrameGenerator(
classes = self.categories,
glob_pattern = self.glob_pattern,
nb_frames = self.time_step,
split_val = prop_val_dataset,
shuffle = True,
batch_size = batch_size,
target_shape = self.size,
nb_channel = self.channels,
transformation = data_aug,
use_frame_cache = True)
# Create
validation_dataset = train_dataset.get_validation_generator()
return train_dataset, validation_dataset
def get_frames(filename):
SIZE = (128, 128)
CHANNELS = 3
NBFRAME = 20
data_aug = tf.keras.preprocessing.image.ImageDataGenerator(rescale=1./255)
test=VideoFrameGenerator(glob_pattern=filename,nb_frames=NBFRAME,shuffle=False, batch_size=1,
target_shape=SIZE, transformation=data_aug, use_frame_cache=False)
return test
def create_video_frames_generator():
glob_pattern = 'short_videos_2/*.avi'
gen = VideoFrameGenerator(glob_pattern=glob_pattern,
nb_frames=NBFRAME,
target_shape=SIZE,
batch_size=BS,
nb_channel=CHANNELS,
use_frame_cache=True)
return gen
NBFRAME = 5
BS = 8
EPOCHS=50
CONVSHAPE=SIZE + (CHANNELS,)
INSHAPE=(NBFRAME,) + CONVSHAPE
#pattern to get videos and classes
glob_pattern = '../../data/facerec/{classname}/*.mp4'
#Create video frame generator
train = VideoFrameGenerator(
classes=classes,
glob_pattern=glob_pattern,
nb_frames=NBFRAME,
split_val=.33,
shuffle=True,
batch_size=BS,
target_shape=SIZE,
nb_channel=CHANNELS,
use_frame_cache=False
)
valid = train.get_validation_generator()
from keras.layers import Conv2D, BatchNormalization, MaxPool2D, GlobalMaxPool2D
def build_convnet(shape=CONVSHAPE):
momentum = .9
model = keras.Sequential()
model.add(Conv2D(64, (3,3), input_shape=shape, padding='same', activation='relu'))
model.add(Conv2D(64, (3,3), padding='same', activation='relu'))
model.add(BatchNormalization(momentum=momentum))
mode='auto')
# for data augmentation
data_augmentation = keras.preprocessing.image.ImageDataGenerator(
zoom_range=.1,
rotation_range=8,
width_shift_range=.2,
height_shift_range=.2)
# Create video frame generator
train_set = VideoFrameGenerator(classes=class_ids,
glob_pattern=video_path,
nb_frames=no_of_frames,
split_test=.4,
split_val=.2,
shuffle=True,
batch_size=batch_size_value,
target_shape=(img_width, img_height),
nb_channel=channels,
transformation=data_augmentation,
use_frame_cache=True)
validation_set = train_set.get_validation_generator()
test_set = train_set.get_test_generator()
history = model.fit_generator(train_set,
validation_data=validation_set,
verbose=1,
epochs=no_of_epochs,
callbacks=[checkpoint, early_stopping_criteria])