def run_celeba():
celeba_database = CelebADatabase(input_shape=(224, 224, 3))
# celeba_database = LFWDatabase()
maml = FaceRecognition(
database=celeba_database,
network_cls=MiniImagenetModel,
# network_cls=VGGSmallModel,
n=5,
k_ml=1,
k_val_ml=5,
k_val=1,
k_val_val=15,
k_test=1,
k_val_test=15,
meta_batch_size=4,
num_steps_ml=5,
lr_inner_ml=0.01,
num_steps_validation=5,
save_after_iterations=20,
meta_learning_rate=0.001,
report_validation_frequency=250,
log_train_images_after_iteration=1000,
num_tasks_val=100,
clip_gradients=True,
experiment_name='celeba')
# maml.train(iterations=100)
# maml.evaluate(5, seed=42)
maml.evaluate_with_original_face_recognition(5, seed=42)
def run_transfer_learning():
isic_database = CelebADatabase()
transfer_learning = TransferLearning(
database=isic_database,
network_cls=get_network,
n=5,
k=1,
k_val_ml=5,
k_val_val=15,
k_val_test=15,
k_test=1,
meta_batch_size=4,
num_steps_ml=5,
lr_inner_ml=0.05,
num_steps_validation=5,
save_after_iterations=15000,
meta_learning_rate=0.001,
report_validation_frequency=250,
log_train_images_after_iteration=1000,
number_of_tasks_val=100,
number_of_tasks_test=1000,
clip_gradients=True,
experiment_name='fixed_vgg_19',
val_seed=42,
val_test_batch_norm_momentum=0.0,
)
transfer_learning.evaluate(50, seed=42, use_val_batch_statistics=True)
def run_celeba():
celeba_database = CelebADatabase()
base_model = tf.keras.applications.VGG19(weights='imagenet')
feature_model = tf.keras.models.Model(inputs=base_model.input,
outputs=base_model.layers[24].output)
sml = SML(
database=celeba_database,
network_cls=MiniImagenetModel,
n=5,
k=1,
k_val_ml=5,
k_val_val=15,
k_val_test=15,
k_test=15,
meta_batch_size=4,
num_steps_ml=5,
lr_inner_ml=0.05,
num_steps_validation=5,
save_after_iterations=15000,
meta_learning_rate=0.001,
n_clusters=20,
feature_model=feature_model,
# feature_size=288,
feature_size=4096,
input_shape=(224, 224, 3),
preprocess_function=tf.keras.applications.vgg19.preprocess_input,
log_train_images_after_iteration=1000,
number_of_tasks_val=100,
number_of_tasks_test=1000,
clip_gradients=True,
report_validation_frequency=250,
experiment_name='celeba_imagenet_features')
sml.train(iterations=60000)
sml.evaluate(iterations=50, seed=42)
def run_celeba():
celeba_database = CelebADatabase()
gan_sampling = GANSampling(
database=celeba_database,
network_cls=MiniImagenetModel,
n=5,
k=1,
k_val_ml=6,
k_val_train=1,
k_val_val=15,
k_val_test=15,
k_test=1,
meta_batch_size=4,
num_steps_ml=5,
lr_inner_ml=0.05,
num_steps_validation=5,
save_after_iterations=2000,
meta_learning_rate=0.001,
report_validation_frequency=250,
log_train_images_after_iteration=1000,
number_of_tasks_val=100,
number_of_tasks_test=1000,
clip_gradients=True,
experiment_name='celeba_noise_std_1.2',
val_seed=42,
val_test_batch_norm_momentum=0.0
)
gan_sampling.train(iterations=20000)
gan_sampling.evaluate(iterations=50, use_val_batch_statistics=True, seed=42)
def run_celeba():
celeba_database = CelebADatabase()
maml = ModelAgnosticMetaLearningModel(
database=celeba_database,
network_cls=MiniImagenetModel,
n=5,
k_ml=1,
k_val_ml=5,
k_val=1,
k_val_val=15,
k_test=1,
k_val_test=15,
meta_batch_size=4,
num_steps_ml=1,
lr_inner_ml=0.05,
num_steps_validation=5,
save_after_iterations=5000,
meta_learning_rate=0.0001,
report_validation_frequency=250,
log_train_images_after_iteration=1000,
num_tasks_val=100,
clip_gradients=True,
experiment_name='celeba'
)
maml.train(iterations=60000)
maml.evaluate(50, num_tasks=1000, seed=42)
def setUp(self):
def parse_function(example_address):
return example_address
self.parse_function = parse_function
self.omniglot_database = OmniglotDatabase(random_seed=-1,
num_train_classes=1200,
num_val_classes=100)
self.mini_imagenet_database = MiniImagenetDatabase()
self.celeba_database = CelebADatabase()
self.lfw_database = LFWDatabase()
self.euro_sat_database = EuroSatDatabase()
self.isic_database = ISICDatabase()
self.chest_x_ray_database = ChestXRay8Database()
def run_celeba():
celeba_database = CelebADatabase(random_seed=-1)
maml = MAMLAbstractLearner(database=celeba_database,
network_cls=MiniImagenetModel,
n=5,
k=4,
meta_batch_size=1,
num_steps_ml=5,
lr_inner_ml=0.01,
num_steps_validation=5,
save_after_epochs=20,
meta_learning_rate=0.001,
report_validation_frequency=1,
log_train_images_after_iteration=1000,
least_number_of_tasks_val_test=50,
clip_gradients=True,
experiment_name='celeba')
maml.train(epochs=100)
test_dataset = self.database.get_attributes_task_dataset(
partition='test',
k=self.k_test,
k_val=self.k_val_test,
meta_batch_size=1,
parse_fn=self.gan.parser.get_parse_fn(),
seed=seed)
test_dataset = test_dataset.repeat(-1)
test_dataset = test_dataset.take(num_tasks)
setattr(test_dataset, 'steps_per_epoch', num_tasks)
return test_dataset
if __name__ == '__main__':
celeba_database = CelebADatabase()
shape = (84, 84, 3)
latent_dim = 512
gan = hub.load(
"https://tfhub.dev/google/progan-128/1").signatures['default']
setattr(gan, 'parser', CelebAGANParser(shape=(84, 84, 3)))
maml_gan = MAMLGANProGAN(
gan=gan,
latent_dim=latent_dim,
generated_image_shape=shape,
database=celeba_database,
network_cls=MiniImagenetModel,
n=2, # n=2
k_ml=1,
k_val_ml=5,
x = layers.ReLU()(x)
decoder_outputs = layers.Conv2DTranspose(3, 4, activation="sigmoid", padding="same")(x)
decoder = keras.Model(latent_inputs, decoder_outputs, name="decoder")
decoder.summary()
return decoder
if __name__ == '__main__':
# import tensorflow as tf
# gpus = tf.config.experimental.list_physical_devices('GPU')
# tf.config.experimental.set_memory_growth(gpus[0], True)
# tf.config.experimental_run_functions_eagerly(True)
celebalot_database = CelebADatabase()
shape = (84, 84, 3)
latent_dim = 500
celebalot_encoder = get_encoder(latent_dim)
celebalot_decoder = get_decoder(latent_dim)
celebalot_parser = CelebAParser(shape=shape)
vae = VAE(
'celeba2',
image_shape=shape,
latent_dim=latent_dim,
database=celebalot_database,
parser=celebalot_parser,
encoder=celebalot_encoder,
decoder=celebalot_decoder,
visualization_freq=1,
