def __init__(self, sess, model_name, images, is_training=True):
self.sess = sess
self.images = images
self.batch_size = 256
self.input_height = 32
self.input_width = 32
self.num_input = 3 * 32 * 32
self.epoch = 80
self.learning_rate = 0.1
self.num_classes = 10
self.temperature = 5
self.weight_decay = 0.0001
self.checkpoint_dir = "/tmp/training/" + model_name
self.checkpoint_file = "/student.cpkt"
self.checkpoint_path = "/tmp/training/" + model_name + "/student.cpkt"
self.is_training = is_training
self.model_name = model_name
hparams = tf.contrib.training.HParams(
train_size=50000,
validation_size=0,
eval_test=1,
dataset=FLAGS.dataset,
data_path=FLAGS.data_path,
batch_size=128,
gradient_clipping_by_global_norm=5.0)
self.dataset = data_utils.DataSet(hparams)
# Store layer's weight and bias using var_scope
# Using ResNet
self.build_model()
def __init__(self, fooled_images=None, fooled_labels=None):
self.fooled_images = fooled_images
self.fooled_labels = fooled_labels
hparams = tf.contrib.training.HParams(
train_size=50000,
validation_size=0,
eval_test=1,
dataset=FLAGS.dataset,
data_path=FLAGS.data_path,
batch_size=256,
gradient_clipping_by_global_norm=5.0)
self.origin_dataset = data_utils.DataSet(
hparams) # Dataset(cifar-10, preprocessed)
self.checkpoint_path = '/tmp/training/distillation/resnet_8'
self.beta = 1
self.weighted = 1.0
self.adv = tf.placeholder(tf.bool, name='adv')
self.flag = tf.placeholder(tf.bool, name='flag')
self.epochs = 100
self.batch_size = 256
self.height = 32
self.width = 32
self.temperature = 3.0
self.num_classes = 10
self.learning_rate = 0.05
self.curr_train_index = 0
self.display_step = 50
def __init__(self, hparams):
self._session = None
self.hparams = hparams
self.model_dir = os.path.join(FLAGS.checkpoint_dir, 'model')
self.log_dir = os.path.join(FLAGS.checkpoint_dir, 'log')
# Set the random seed to be sure the same validation set
# is used for each model
np.random.seed(0)
self.data_loader = data_utils.DataSet(hparams)
np.random.seed() # Put the random seed back to random
self.data_loader.reset()
