def open_sesemi():
args = parse_args()
network = args.network
dataset = args.dataset
nb_labels = args.nb_labels
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_id
arg2var = {
'convnet': convnet,
'wrn': wrn,
'nin': nin,
'svhn': svhn,
'cifar10': cifar10,
'cifar100': cifar100,
}
# Experiment- and dataset-dependent parameters.
zca = True
hflip = True
epochs = 50
if dataset in {'svhn', 'cifar10'}:
if dataset == 'svhn':
zca = False
hflip = False
epochs = 30
nb_classes = 10
elif dataset == 'cifar100':
nb_classes = 100
else:
raise ValueError('`dataset` must be "svhn", "cifar10", "cifar100".')
super_dropout = 0.2
in_network_dropout = 0.0
if network == 'convnet' and dataset == 'svhn':
super_dropout = 0.5
in_network_dropout = 0.5
elif network == 'wrn' and dataset == 'svhn':
super_dropout = 0.5
# Prepare the dataset.
(x_train, y_train), (x_test, y_test) = arg2var[dataset].load_data()
x_test = global_contrast_normalize(x_test)
x_train = global_contrast_normalize(x_train)
if zca:
zca_whiten = zca_whitener(x_train)
x_train = zca_whiten(x_train)
x_test = zca_whiten(x_test)
x_test = x_test.reshape((len(x_test), 32, 32, 3))
x_train = x_train.reshape((len(x_train), 32, 32, 3))
if nb_labels in {50000, 73257}:
x_labeled = x_train
y_labeled = y_train
else:
labels_per_class = nb_labels // nb_classes
sample_inds = stratified_sample(y_train, labels_per_class)
x_labeled = x_train[sample_inds]
y_labeled = y_train[sample_inds]
y_labeled = to_categorical(y_labeled)
# Shared training parameters.
base_lr = 0.05
batch_size = 16
lr_decay_power = 0.5
input_shape = (32, 32, 3)
max_iter = (len(x_train) // batch_size) * epochs
# Compile the SESEMI model.
sesemi_model, inference_model = compile_sesemi(arg2var[network],
input_shape, nb_classes,
base_lr, in_network_dropout,
super_dropout)
print(sesemi_model.summary())
lr_poly_decay = LRScheduler(base_lr, max_iter, lr_decay_power)
evaluate = DenseEvaluator(inference_model, (x_test, y_test),
hflip,
oversample=True)
super_datagen = ImageDataGenerator(
width_shift_range=[-2, -1, 0, 1, 2],
height_shift_range=[-2, -1, 0, 1, 2],
horizontal_flip=hflip,
preprocessing_function=gaussian_noise,
fill_mode='reflect',
)
self_datagen = ImageDataGenerator(
width_shift_range=[-2, -1, 0, 1, 2],
height_shift_range=[-2, -1, 0, 1, 2],
horizontal_flip=False,
preprocessing_function=gaussian_noise,
fill_mode='reflect',
)
super_data = super_datagen.flow(x_labeled,
y_labeled,
shuffle=True,
batch_size=1,
seed=None)
self_data = self_datagen.flow(x_train,
shuffle=True,
batch_size=1,
seed=None)
train_data_loader = datagen(super_data, self_data, batch_size)
# Fit the SESEMI model on mini-batches with data augmentation.
print('Run configuration:')
print('network=%s,' % network, 'dataset=%s,' % dataset, \
'horizontal_flip=%s,' % hflip, 'ZCA=%s,' % zca, \
'nb_epochs=%d,' % epochs, 'batch_size=%d,' % batch_size, \
'nb_labels=%d,' % len(y_labeled), 'gpu_id=%s' % args.gpu_id)
sesemi_model.fit_generator(
train_data_loader,
epochs=epochs,
verbose=1,
steps_per_epoch=len(x_train) // batch_size,
callbacks=[lr_poly_decay, evaluate],
)
def open_sesemi():
args = parse_args()
network = args.network
nb_extra = args.nb_extra
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_id
arg2var = {
'convnet': convnet,
'wrn': wrn,
}
# Load Tiny Images.
# Code adapted from https://github.com/smlaine2/tempens
with open('./datasets/tiny-images/tiny_index.pkl', 'rb') as f:
tinyimg_index = pickle.load(f, encoding='latin1')
if nb_extra == 237203:
print("Using all classes common with CIFAR-100.")
with open('./datasets/cifar-100/meta', 'rb') as f:
cifar_labels = pickle.load(f,
encoding='latin1')['fine_label_names']
cifar_to_tinyimg = {'maple_tree': 'maple', 'aquarium_fish': 'fish'}
cifar_labels = [
l if l not in cifar_to_tinyimg else cifar_to_tinyimg[l]
for l in cifar_labels
]
load_indices = sum(
[list(range(*tinyimg_index[label])) for label in cifar_labels], [])
elif nb_extra == 500000:
print("Using %d random images." % nb_extra)
nb_tinyimages = max(e for s, e in tinyimg_index.values())
load_indices = np.arange(nb_tinyimages)
rng.shuffle(load_indices)
load_indices = load_indices[:nb_extra]
load_indices.sort() # sorted for faster seeks.
else:
raise ValueError('`--extra` must be integer 237203 or 500000.')
print("Loading %d auxiliary unlabeled Tiny Images." % len(load_indices))
z_train = load_tinyimages(load_indices)
# Load CIFAR-100.
(x_train, y_train), (x_test, y_test) = cifar100.load_data()
x_test = global_contrast_normalize(x_test)
x_train = global_contrast_normalize(x_train)
z_train = global_contrast_normalize(z_train)
zca_whiten = zca_whitener(np.concatenate([x_train, z_train], axis=0))
x_test = zca_whiten(x_test)
x_train = zca_whiten(x_train)
z_train = zca_whiten(z_train)
x_test = x_test.reshape((len(x_test), 32, 32, 3))
x_train = x_train.reshape((len(x_train), 32, 32, 3))
z_train = z_train.reshape((len(z_train), 32, 32, 3))
y_train = to_categorical(y_train)
# Shared training parameters.
zca = True
hflip = True
epochs = 50
base_lr = 0.05
batch_size = 8
nb_classes = 100
lr_decay_power = 0.5
super_dropout = 0.2
in_network_dropout = 0.0
input_shape = (32, 32, 3)
max_iter = (len(x_train) // batch_size) * epochs
# Compile the SESEMI model.
sesemi_model, inference_model = compile_sesemi(arg2var[network],
input_shape, nb_classes,
base_lr, in_network_dropout,
super_dropout)
print(sesemi_model.summary())
lr_poly_decay = LRScheduler(base_lr, max_iter, lr_decay_power)
evaluate = DenseEvaluator(inference_model, (x_test, y_test),
hflip,
oversample=True)
super_datagen = ImageDataGenerator(
width_shift_range=[-2, -1, 0, 1, 2],
height_shift_range=[-2, -1, 0, 1, 2],
horizontal_flip=hflip,
preprocessing_function=gaussian_noise,
fill_mode='reflect',
)
self_datagen = ImageDataGenerator(
width_shift_range=[-2, -1, 0, 1, 2],
height_shift_range=[-2, -1, 0, 1, 2],
horizontal_flip=False,
preprocessing_function=gaussian_noise,
fill_mode='reflect',
)
super_data = super_datagen.flow(x_train,
y_train,
shuffle=True,
batch_size=1,
seed=None)
self_data = self_datagen.flow(x_train,
shuffle=True,
batch_size=1,
seed=None)
extra_data = self_datagen.flow(z_train,
shuffle=True,
batch_size=1,
seed=None)
train_data_loader = datagen_tinyimages(super_data, self_data, extra_data,
batch_size)
# Fit the SESEMI model on mini-batches with data augmentation.
print('Run configuration:')
print('network=%s,' % network, 'ZCA=%s,' % zca, 'nb_epochs=%d,' % epochs, \
'horizontal_flip=%s,' % hflip, 'nb_extra=%d,' % len(z_train), \
'batch_size=%d,' % batch_size, 'gpu_id=%s' % args.gpu_id)
sesemi_model.fit_generator(
train_data_loader,
epochs=epochs,
verbose=1,
steps_per_epoch=len(x_train) // batch_size,
callbacks=[lr_poly_decay, evaluate],
)
def main():
args = parse_args()
network = args.network
dataset = args.dataset
nb_labels = args.nb_labels
os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu_id)
arg2var = {'convnet': convnet,
'wrn': wrn,
'resnet50v2':resnet50v2,
'svhn': svhn,
'cifar10': cifar10,
'cifar100': cifar100,}
# Dataset-specific parameters
hflip = True
zca = True
epochs = 10
if dataset in ['svhn', 'cifar10']:
if dataset == 'svhn':
hflip = False
zca = False
epochs = 30
nb_classes = 10
elif dataset == 'cifar100':
nb_classes = 100
else:
raise ValueError('`dataset` must be "svhn", "cifar10", "cifar100".')
(x_train, y_train), (x_test, y_test) = arg2var[dataset].load_data()
x_train = global_contrast_normalize(x_train)
x_test = global_contrast_normalize(x_test)
if zca:
zca_whiten = zca_whitener(x_train)
x_train = zca_whiten(x_train)
x_test = zca_whiten(x_test)
x_train = x_train.reshape((len(x_train), 32, 32, 3))
x_test = x_test.reshape((len(x_test), 32, 32, 3))
labels_per_class = nb_labels // nb_classes
if nb_labels == 73257:
labels_per_class = 1000000
sample_inds = stratified_sample(y_train, labels_per_class)
x_labeled = x_train[sample_inds]
y_labeled = y_train[sample_inds]
y_labeled = to_categorical(y_labeled)
# Training parameters
input_shape = (32, 32, 3)
batch_size = 32
base_lr = 0.05
lr_decay_power = 0.5
dropout_rate = 0.2
max_iter = (len(x_train) // batch_size) * epochs
sesemi_model, inference_model = open_sesemi(
arg2var[network], input_shape, nb_classes, base_lr, dropout_rate)
print(sesemi_model.summary())
super_datagen = ImageDataGenerator(
width_shift_range=3,
height_shift_range=3,
horizontal_flip=hflip,
preprocessing_function=gaussian_noise,
fill_mode='reflect',
)
self_datagen = ImageDataGenerator(
width_shift_range=3,
height_shift_range=3,
horizontal_flip=False,
preprocessing_function=gaussian_noise,
fill_mode='reflect',
)
super_data = super_datagen.flow(
x_labeled, y_labeled, shuffle=True, batch_size=1, seed=None)
self_data = self_datagen.flow(
x_train, shuffle=True, batch_size=1, seed=None)
train_data_loader = datagen(super_data, self_data, batch_size)
lr_poly_decay = LRScheduler(base_lr, max_iter, lr_decay_power)
evaluate = DenseEvaluator(inference_model, (x_test, y_test), hflip)
# Fit the SESEMI model on mini-batches with data augmentation
print('Run configuration:')
print('network=%s,' % network, 'dataset=%s,' % dataset, \
'horizontal_flip=%s,' % hflip, 'ZCA=%s,' % zca, \
'nb_epochs=%d,' % epochs, 'batch_size=%d,' % batch_size, \
'nb_labels=%d,' % len(x_labeled), 'gpu_id=%d' % args.gpu_id)
sesemi_model.fit_generator(train_data_loader,
epochs=epochs, verbose=1,
steps_per_epoch=len(x_train) // batch_size,
callbacks=[lr_poly_decay, evaluate],)
return