)
# gan.perform_training(epochs=GAN_EPOCHS, checkpoint_freq=50)
gan.load_latest_checkpoint()
print("GAN training finished")
time.sleep(1)
train_folders = omniglot_database.train_folders
keys = list(train_folders.keys())
sample_size = np.max([
N_WAY * META_BATCH_SIZE,
int(len(train_folders.keys()) * labeled_percentage)
])
keys = np.random.choice(keys, sample_size, replace=False)
train_folders = {k: v for (k, v) in train_folders.items() if k in keys}
omniglot_database.train_folders = train_folders
L = None
ssml_maml = SSMLMAML(perc=labeled_percentage,
accessible_labels=L,
database=omniglot_database,
network_cls=SimpleModel,
n=N_WAY,
k_ml=K,
k_val_ml=K,
k_val=K,
k_val_val=K,
k_test=K,
k_val_test=K,
meta_batch_size=META_BATCH_SIZE,
num_steps_ml=5,
experiment_name = prefix+str(labeled_percentage)
if GAN_N_ALT > 1:
experiment_name += "_alt"+str(GAN_N_ALT)
# Split labeled and not labeled
train_folders = omniglot_database.train_folders
keys = list(train_folders.keys())
labeled_keys = np.random.choice(keys, int(len(train_folders.keys())*labeled_percentage), replace=False)
train_folders_labeled = {k: v for (k, v) in train_folders.items() if k in labeled_keys}
train_folders_unlabeled = {k: v for (k, v) in train_folders.items() if k not in labeled_keys}
print("Number of unlabeled classes:", len(train_folders_unlabeled.keys()))
print("Number of labeled classes:" , len(train_folders_labeled.keys()))
# Start with unlabeled
omniglot_database.train_folders = train_folders_unlabeled
omniglot_discriminator = get_discriminator(len(train_folders_labeled))
ssgan = SSGAN(
gan_name='ssgan_omniglot_'+str(labeled_percentage),
SS=False,
image_shape=shape,
latent_dim=latent_dim,
database=omniglot_database,
parser=omniglot_parser,
generator=omniglot_generator,
discriminator=omniglot_discriminator,
visualization_freq=50,
d_learning_rate=0.0003,
g_learning_rate=0.0003,
)