def load_mpiigaze_train_test_ds_one_eye_leave_one_out(dataset_name,
person_id,
batch_size=128,
grayscale=True,
val_split=None):
"""
val_split argument only for consistent function interface
"""
eye_train, headpose_train, y_train, train_subject_ids = _load_all_people(
dataset_name, grayscale, person_id)
eye_test, headpose_test, y_test = _load_one_person(dataset_name, person_id,
grayscale)
test_subject_ids = np.asarray([person_id] * len(y_test))
train_dataset = prepare_dataset((eye_train, headpose_train), y_train,
batch_size)
# don't shuffle test dataset to have consistent outcomes
# (when ploting error per person)
test_dataset = prepare_dataset((eye_test, headpose_test),
y_test,
batch_size,
shuffle=False)
return train_dataset, test_dataset, test_subject_ids
def load_mpiigaze_train_test_ds_both_leave_one_out_reject_suspicious(
dataset_name,
person_id,
val_split=0.2,
batch_size=128,
grayscale=True):
"""
val_split argument only for consistent function interface
"""
right_eye_train, left_eye_train, headpose_train, y_train, train_subject_ids = _load_all_people_reject_suspicious(
dataset_name, grayscale, person_id)
right_eye_test, left_eye_test, headpose_test, y_test = _load_one_person(
dataset_name, person_id, grayscale)
if person_id == 2:
right_eye_test, left_eye_test, headpose_test, y_test = _filter_subject_02(
right_eye_test, left_eye_test, headpose_test, y_test)
if person_id == 10:
right_eye_test, left_eye_test, headpose_test, y_test = _filter_subject_10(
right_eye_test, left_eye_test, headpose_test, y_test)
test_subject_ids = np.asarray([person_id] * len(y_test))
train_dataset = prepare_dataset(
(right_eye_train, left_eye_train, headpose_train), y_train, batch_size)
# don't shuffle test dataset to have consistent outcomes
test_dataset = prepare_dataset(
(right_eye_test, left_eye_test, headpose_test),
y_test,
batch_size,
shuffle=False)
return train_dataset, test_dataset, test_subject_ids
def load_train_test_ds_reject_suspicious(dataset_name,
person_id,
val_split=0.2,
batch_size=128,
grayscale=True,
all_subjects=None):
if all_subjects is None:
all_subjects = list(range(0, 7)) + list(range(8, 15)) + [24, 25]
test_subject_ids = None
if person_id is None:
right_images, left_images, poses, gazes, subject_ids = _load_all_people_reject_suspicious(
dataset_name, grayscale, subjects=all_subjects)
(right_eye_train, right_eye_test, left_eye_train, left_eye_test,
headpose_train, headpose_test, y_train, y_test, train_subject_ids,
test_subject_ids) = train_test_split(right_images,
left_images,
poses,
gazes,
subject_ids,
test_size=val_split,
random_state=42,
stratify=subject_ids)
else:
right_images, left_images, poses, gazes = _load_one_person(
dataset_name, person_id, grayscale)
(right_eye_train, right_eye_test, left_eye_train, left_eye_test,
headpose_train, headpose_test, y_train,
y_test) = train_test_split(right_images,
left_images,
poses,
gazes,
test_size=val_split,
random_state=42)
# filter suspicious
if person_id == 7:
raise ValueError("Person 7 is rejected.")
right_images, left_images, poses, gazes = _load_one_person(
dataset_name, i, grayscale)
if person_id == 2:
right_images, left_images, poses, gazes = _filter_subject_02(
right_images, left_images, poses, gazes)
if person_id == 10:
right_images, left_images, poses, gazes = _filter_subject_10(
right_images, left_images, poses, gazes)
train_dataset = prepare_dataset(
(right_eye_train, left_eye_train, headpose_train), y_train, batch_size)
# don't shuffle test dataset to have consistent outcomes
test_dataset = prepare_dataset(
(right_eye_test, left_eye_test, headpose_test),
y_test,
batch_size,
shuffle=False)
return train_dataset, test_dataset, test_subject_ids
def load_test_own_mpiigaze_full(dataset_name,
person_id,
val_split=0.2,
batch_size=128,
grayscale=True,
all_subjects=None):
if all_subjects is None:
all_subjects = list(range(0, 7)) + list(range(
8, 15)) + [24, 25] + list(range(30, 36))
right_images, left_images, poses, gazes, subject_ids = _load_all_people_reject_suspicious(
dataset_name, grayscale, subjects=all_subjects)
indices = np.isin(subject_ids, all_subjects)
right_images = right_images[indices]
left_images = left_images[indices]
poses = poses[indices]
gazes = gazes[indices]
test_subject_ids = subject_ids[indices]
# don't shuffle test dataset to have consistent outcomes
test_dataset = prepare_dataset((right_images, left_images, poses),
gazes,
batch_size,
shuffle=False)
return test_dataset, test_subject_ids
def load_hysts_mpiigaze_train_test_ds(person_id, val_split=0.2, batch_size=128):
test_subject_ids = None
if person_id is None:
images, poses, gazes, subject_ids = _load_all_people()
(eye_train, eye_test,
headpose_train, headpose_test,
y_train, y_test,
train_subject_ids, test_subject_ids) = train_test_split(images, poses, gazes, subject_ids, test_size=val_split,
random_state=42, stratify=subject_ids)
else:
images, poses, gazes = _load_one_person(person_id)
(eye_train, eye_test,
headpose_train, headpose_test,
y_train, y_test) = train_test_split(images, poses, gazes, test_size=val_split, random_state=42)
train_dataset = prepare_dataset((eye_train, headpose_train), y_train, batch_size)
# don't shuffle test dataset to have consistent outcomes
test_dataset = prepare_dataset((eye_test, headpose_test), y_test, batch_size, shuffle=False)
return train_dataset, test_dataset, test_subject_ids
def load_mpiigaze_train_test_ds_both_from_single(dataset_name,
person_id,
val_split=0.2,
batch_size=128,
grayscale=True):
test_subject_ids = None
if person_id is None:
right_images, left_images, poses, gazes, subject_ids = _load_all_people(
dataset_name, grayscale)
(right_eye_train, right_eye_test, left_eye_train, left_eye_test,
headpose_train, headpose_test, y_train, y_test, train_subject_ids,
test_subject_ids) = train_test_split(right_images,
left_images,
poses,
gazes,
subject_ids,
test_size=val_split,
random_state=42,
stratify=subject_ids)
else:
right_images, left_images, poses, gazes = _load_one_person(
dataset_name, person_id, grayscale)
(right_eye_train, right_eye_test, left_eye_train, left_eye_test,
headpose_train, headpose_test, y_train,
y_test) = train_test_split(right_images,
left_images,
poses,
gazes,
test_size=val_split,
random_state=42)
train_dataset = prepare_dataset(
(right_eye_train, left_eye_train, headpose_train), y_train, batch_size)
# don't shuffle test dataset to have consistent outcomes
test_dataset = prepare_dataset(
(right_eye_test, left_eye_test, headpose_test),
y_test,
batch_size,
shuffle=False)
return train_dataset, test_dataset, test_subject_ids
def load_test_mpii_face_gaze(dataset_name, person_id, val_split=0.2, batch_size=128,
grayscale=True, all_subjects=None):
if all_subjects is None:
all_subjects = list(range(0, 7)) + list(range(8, 15))
right_images, left_images, poses, gazes, test_subject_ids = load_all_full_data(dataset_name, grayscale)
indices = np.isin(test_subject_ids, all_subjects)
right_images = right_images[indices]
left_images = left_images[indices]
poses = poses[indices]
gazes = gazes[indices]
test_subject_ids = test_subject_ids[indices]
# don't shuffle test dataset to have consistent outcomes
test_dataset = prepare_dataset((right_images, left_images, poses), gazes, batch_size, shuffle=False)
return test_dataset, test_subject_ids