def prepare_batch(config):
"""Extract faces from the dataset."""
gl = utils.load_fun(config['get_label'])
stream = utils.load_fun(config['get_files'])(**config)
stream = ds.stream(ds.add_label(gl), stream)
stream = ds.stream(ds.apply_to_x(detect.load_img), stream)
stream = ds.stream(
ds.apply_to_x(adapters.resize(**config['full_image_size'])), stream
)
batches = ds.stream_batch(stream, config['batch_size'])
adapters_list = [
ds.apply_to_x(ds.foreach(adapters.astype('uint8')))
]
if config['has_faces']:
adapters_list.append(
ds.apply_to_x(ds.foreach(features.get_face()))
)
adapters_list.extend([
ds.apply_to_x(ds.foreach(adapters.resize(**config['image_size']))),
ds.apply_to_x(ds.foreach(adapters.astype('uint8')))
])
batches = ds.batch_adapt(batches, adapters_list)
return batches
def prepare_batch(filenames, config, epochs, *args, **kwargs):
"""Prepare a batch."""
shape = config['image_size']['img_x'], config['image_size']['img_y'], 1
filenames = dataset.epochs(filenames, epochs=epochs)
get_label = utils.load_fun(config['get_label'])
stream = dataset.get_data(filenames, get_label)
batches = dataset.stream_batch(stream, config['batch_size'])
adapters_list = [
dataset.apply_to_y(dataset.foreach(dataset.categorical)),
dataset.apply_to_x(dataset.foreach(adapters.rgb_to_bn)),
dataset.apply_to_x(dataset.foreach(
adapters.resize(**config['image_size'])
)),
dataset.apply_to_x(adapters.matrix_to_bn),
]
if 'distortions' in config:
adapters_list.append(
adapters.apply_distortion(
ImageDataGenerator(**config['distortions'])
)
)
adapters_list.append(dataset.apply_to_x(adapters.normalize(255))),
batches = dataset.batch_adapt(batches, adapters_list)
return batches, shape
def prepare_batch(filenames, config, epochs, *args, **kwargs):
"""Prepare a batch."""
shape = 68 * 2
norm_max = max(config['image_size'].values())
filenames = dataset.epochs(filenames, epochs=epochs)
get_label = load_fun(config['get_label'])
stream = dataset.get_data(filenames, get_label, loader=np.load)
batches = dataset.stream_batch(stream, config['batch_size'])
batches = dataset.batch_adapt(batches, [
dataset.apply_to_y(dataset.foreach(dataset.categorical)),
dataset.apply_to_x(dataset.foreach(dataset.flatten)),
dataset.apply_to_x(adapters.normalize(norm_max)),
])
return batches, shape
def _prepare_submodels(filenames, config, epochs, type_):
"""Prepare submodels."""
filenames = get_names_only(filenames)
output_shape = 0
batches_list = []
for subconf in config['submodels']:
# get filenames for the submodel
subtrain = deepcopy(filenames)
subtrain = attach_basepath(subconf['directory'], subtrain)
if 'files_types' in subconf:
subtrain = attach_filetype(subconf['files_types'][0], subtrain)
subtrain = dataset.epochs(subtrain, epochs=epochs)
# load keras submodel
submodel = models.load_model(subconf['model'])
# compute the output shape
(_, shape) = submodel.output_shape
output_shape += shape
# input shape if is not explicitly set
if 'image_size' not in subconf:
# input shape
(_, img_x, img_y, _) = submodel.input_shape
subconf['image_size'] = {'img_x': img_x, 'img_y': img_y}
# copy global configuration
subconf['batch_size'] = config['batch_size']
# check if defined a personalized prediction function
to_predict = dataset.to_predict
if 'to_predict' in subconf:
to_predict = utils.load_fun(subconf['to_predict'])(type_=type_)
# build prediction batches
prepare = utils.load_fun(subconf['prepare_batch'])
batches, _shape = prepare(subtrain, subconf, epochs)
batches = dataset.batch_adapt(batches, [
dataset.apply_to_x(to_predict(model=submodel)),
])
batches_list.append(batches)
todo = dataset.flatten
return dataset.merge_batches(batches_list,
adapters=[
dataset.apply_to_x(dataset.foreach(todo))
]), output_shape
directory = "data/KDEF-straight"
shape_predictor = "data/shape_predictor_68_face_landmarks.dat"
img_x, img_y = 100, 100
filenames = dataset.get_files(directory)
filenames = list(filenames)
stream = get_data(filenames)
batches = dataset.stream_batch(stream, 1)
batches = dataset.batch_adapt(
batches,
[
dataset.apply_to_y(dataset.foreach(dataset.categorical)),
# dataset.apply_to_x(dataset.foreach(
# adapters.resize(100, 100)
# )),
dataset.apply_to_x(dataset.foreach(adapters.astype('uint8'))),
# dataset.apply_to_x(dataset.foreach(
# features.extract_shape(shape_predictor)
# )),
dataset.apply_to_x(
dataset.foreach(
features.extract_part(
'nose', features.extract_shape(shape_predictor)))),
# dataset.apply_to_x(dataset.foreach(detect.shape2matrix)),
# dataset.apply_to_x(dataset.foreach(features.expand2image(100, 100)))
])