def debug(load_tables=False, save_tables=True):
# reset graph
tf.keras.backend.clear_session()
# hyper-parameters settings
batch_size = 1
image_shape = (batch_size, 28, 28, 1)
noise_epsilon = 1.0 # max value of the images is 1.0
exploration_decay = 0.9
exploration_decay_steps = 1000
alpha = 0.5
noise_type = 'gaussian'
#############################################
similarity_threshold = 0.1
reward_thres = 0.85
env = Environment(batch_size)
agent = BlackBoxAgent(image_shape, noise_epsilon, alpha, exploration_decay,
noise_type)
secondAgent = varietyAgent(similarity_threshold, reward_thres)
# set tensorboard
log_dir = "logs_double_" + noise_type + "/" + "nmax_" + str(
noise_epsilon) + "_alpha_" + str(agent.alpha) + "_threshold_" + str(
agent.reward_threshold)
summary_writer = tf.summary.create_file_writer(log_dir)
if load_tables:
agent.LoadTables()
secondAgent.LoadTables()
acc_calculator = []
i = -1
while agent.exploration_rate >= 1e-2:
i += 1
state, state_label = env.State()
action, noise = agent.GenerateAdvSample(state)
reward = env.Reward(action, state_label)
reward = secondAgent.PipeLine(state, noise, reward)
agent.UpdateTable(state, noise, reward)
img_table_size = len(agent.image_table)
noise_table_size = len(agent.noise_table)
acc_calculator.append(reward)
size_of_secondagent = secondAgent.SizeOfTable()
with summary_writer.as_default():
tf.summary.scalar('img_table_size', img_table_size, step=i)
# if secondAgent.Logging(10):
with summary_writer.as_default():
tf.summary.scalar('second_agent_size',
np.sum(size_of_secondagent),
step=i)
tf.summary.scalar('secag_imgt_size',
len(secondAgent.original_img_table.keys()),
step=i)
if (not agent.decay_cmd):
agent.IfDecay()
if agent.decay_cmd and ((i + 1) % exploration_decay_steps == 0):
agent.UpdateExplorationRate()
with summary_writer.as_default():
tf.summary.scalar('exploration_rate',
agent.exploration_rate,
step=i)
if save_tables:
if i % 2000 == 0:
agent.SaveTables()
secondAgent.SaveTables()
if len(acc_calculator) >= 20:
acc_calculator = np.array(acc_calculator)
acc = np.mean(acc_calculator)
with summary_writer.as_default():
tf.summary.scalar('average_confidence_loss', acc, step=i)
acc_calculator = []
if i >= 70000:
break
else:
continue
def debug(noise_epsilon,
alpha,
noise_type,
load_tables=False,
save_tables=True):
# reset graph
tf.keras.backend.clear_session()
# hyper-parameters settings
batch_size = 1
image_shape = (batch_size, 28, 28, 1)
# noise_epsilon = 0.8 # max value of the images is 1.0
exploration_decay = 0.8
exploration_decay_steps = 800
# alpha = 0.5
env = Environment(batch_size)
agent = BlackBoxAgent(image_shape, noise_epsilon, alpha, exploration_decay,
noise_type)
# set tensorboard
log_dir = "logs_" + noise_type + "/" + "nmax_" + str(
noise_epsilon) + "_alpha_" + str(agent.alpha) + "_threshold_" + str(
agent.reward_threshold)
summary_writer = tf.summary.create_file_writer(log_dir)
if load_tables:
agent.LoadTables()
acc_calculator = []
i = -1
while agent.exploration_rate >= 1e-2:
i += 1
state, state_label = env.State()
action, noise = agent.GenerateAdvSample(state)
reward = env.Reward(action, state_label)
agent.UpdateTable(state, noise, reward)
img_table_size = len(agent.image_table)
noise_table_size = len(agent.noise_table)
acc_calculator.append(reward)
with summary_writer.as_default():
tf.summary.scalar('img_table_size', img_table_size, step=i)
tf.summary.scalar('noise_table_size', noise_table_size, step=i)
if (not agent.decay_cmd):
agent.IfDecay()
if agent.decay_cmd and ((i + 1) % exploration_decay_steps == 0):
agent.UpdateExplorationRate()
with summary_writer.as_default():
tf.summary.scalar('exploration_rate',
agent.exploration_rate,
step=i)
if save_tables:
if i % 2000 == 0:
agent.SaveTables()
if len(acc_calculator) >= 20:
acc_calculator = np.array(acc_calculator)
acc = np.mean(acc_calculator)
with summary_writer.as_default():
tf.summary.scalar('average_confidence_loss', acc, step=i)
acc_calculator = []
if i >= 20000:
break
else:
continue