def train_mnist(features_mnist_train=gin.REQUIRED,
features_mnist_test=gin.REQUIRED,
data_path_mnist=gin.REQUIRED,
features_notmnist_test=gin.REQUIRED,
labels_notmnist_test=gin.REQUIRED,
model_dir=gin.REQUIRED,
dim_input=gin.REQUIRED,
num_classes=gin.REQUIRED):
"""Training function."""
# Define the paths
data_path_mnist = os.path.join(FLAGS.baseroute, data_path_mnist)
features_mnist_train = os.path.join(FLAGS.baseroute, features_mnist_train)
features_mnist_test = os.path.join(FLAGS.baseroute, features_mnist_test)
features_notmnist_test = os.path.join(FLAGS.baseroute,
features_notmnist_test)
labels_notmnist_test = os.path.join(FLAGS.baseroute, labels_notmnist_test)
model_dir = os.path.join(FLAGS.baseroute, model_dir)
# Load the features for mnist
(_, y_train), (_, y_test) = mnist_input.load_data(data_path_mnist)
with tf.gfile.Open(features_mnist_train, 'r') as f:
x_train = np.load(f)
with tf.gfile.Open(features_mnist_test, 'r') as f:
x_test = np.load(f)
dataset = (x_train, y_train), (x_test, y_test)
model = build_model(dataset, model_dir, dim_input, num_classes)
# Load notmnist features and labels test dataset
with tf.gfile.Open(features_notmnist_test, 'r') as f:
x_notmnist = np.load(f)
with tf.gfile.Open(labels_notmnist_test, 'r') as f:
y_notmnist = np.load(f)
y_notmnist = tf.keras.utils.to_categorical(y_notmnist, num_classes=10)
# Compute output probabilities on x_test and x_test_notmnist
# Warning: take time, pass over all the saved weights.
x = np.vstack((x_test, x_notmnist))
model.predict(x)
# Dictionary for y - Metrics
y_dic = {'mnist': y_test, 'notmnist': y_notmnist}
# Postprocessing and metrics
if FLAGS.algorithm in ['simple', 'dropout', 'precond']:
postprocess.postprocess_mnist(FLAGS.workdir)
for dataset_str in ['mnist', 'notmnist']:
path_postprocess = os.path.join(FLAGS.workdir, dataset_str)
metrics.Metrics(y_dic[dataset_str], path_postprocess)
# Write the gin config in the working directory
util.save_gin(os.path.join(FLAGS.workdir, 'gin_configuration.txt'))
util.write_gin(FLAGS.workdir)
def train_cifar100(features_cifar100_train=gin.REQUIRED,
features_cifar100_test=gin.REQUIRED,
data_path_cifar100=gin.REQUIRED,
features_cifar100_train_distorted=gin.REQUIRED,
distorted=gin.REQUIRED,
model_dir=gin.REQUIRED,
dim_input=gin.REQUIRED,
num_classes=gin.REQUIRED):
"""Training function."""
# Define the paths
features_cifar100_train = os.path.join(FLAGS.baseroute,
features_cifar100_train)
features_cifar100_test = os.path.join(FLAGS.baseroute,
features_cifar100_test)
data_path_cifar100 = os.path.join(FLAGS.baseroute, data_path_cifar100)
features_cifar100_train_distorted = os.path.join(
FLAGS.baseroute, features_cifar100_train_distorted)
model_dir = os.path.join(FLAGS.baseroute, model_dir)
# Load the features for cifar100
(_, y_train), (_, y_test) = cifar_input.load_data(distorted,
data_path_cifar100,
'cifar100')
if distorted:
with tf.gfile.Open(features_cifar100_train_distorted, 'r') as f:
x_train = np.load(f)
else:
with tf.gfile.Open(features_cifar100_train, 'r') as f:
x_train = np.load(f)
with tf.gfile.Open(features_cifar100_test, 'r') as f:
x_test = np.load(f)
dataset = (x_train, y_train), (x_test, y_test)
model = build_model(dataset, model_dir, dim_input, num_classes)
# Compute output probabilities on x_test
# Warning: take time, pass over all the saved weights.
model.predict(x_test)
# Postprocessing and metrics
if FLAGS.algorithm in ['simple', 'dropout', 'precond']:
postprocess.postprocess_cifar(FLAGS.workdir, 'cifar100')
path_postprocess = os.path.join(FLAGS.workdir, 'cifar100')
metrics.Metrics(y_test, path_postprocess)
# Write the gin config in the working directory
util.save_gin(os.path.join(FLAGS.workdir, 'gin_configuration.txt'))
util.write_gin(FLAGS.workdir)
