# clear the previous files
controller.remove_files()
isFirst = True
# train for number of trails
for trial in range(MAX_TRIALS):
if isFirst:
actions = state
isFirst = False
else:
with policy_sess.as_default():
K.set_session(policy_sess)
actions = controller.get_action(
state) # get an action for the previous state
# print the action probabilities
state_space.print_actions(actions)
print("Predicted actions : ", state_space.parse_state_space_list(actions))
# build a model, train and get reward and accuracy from the network manager
#与子模型关联较弱，只需返回reward与accuracy即可；reward为accuracy指数滑动平均后的值
# if isFirst:
# reward, previous_acc = manger.main(action=[1,'relu',1,'linear'])
# isFirst = False
# else:
reward, previous_acc = manger.main(
action=state_space.parse_state_space_list(actions))
# reward, previous_acc = (1,1)
print("Rewards : ", reward, "Accuracy : ", previous_acc)
with policy_sess.as_default():
def main(_):
# create a shared session between Keras and Tensorflow
policy_sess = tf.Session()
K.set_session(policy_sess)
NUM_LAYERS = 3 # number of layers of the state space
MAX_TRIALS = 250 # maximum number of models generated
MAX_EPOCHS = 60 # maximum number of epochs to train
BATCHSIZE = 100 # batchsize
EXPLORATION = 0.5 # high exploration for the first 1000 steps
REGULARIZATION = 1e-3 # regularization strength
CONTROLLER_CELLS = 32 # number of cells in RNN controller
CLIP_REWARDS = False # clip rewards in the [-0.05, 0.05] range
RESTORE_CONTROLLER = True # restore controller to continue training
# construct a state space
state_space = StateSpace()
# add states
#state_space.add_state(name='kernel', values=[3])
state_space.add_state(name='filters', values=[30, 60, 100, 144])
#state_space.add_state(name='stride', values=[1])
# print the state space being searched
state_space.print_state_space()
previous_acc = 0.0
total_reward = 0.0
with policy_sess.as_default():
# create the Controller and build the internal policy network
controller = Controller(policy_sess,
NUM_LAYERS,
state_space,
reg_param=REGULARIZATION,
exploration=EXPLORATION,
controller_cells=CONTROLLER_CELLS,
restore_controller=RESTORE_CONTROLLER)
print('done')
# create the Network Manager
manager = NetworkManager(FLAGS, clip_rewards=CLIP_REWARDS)
# get an initial random state space if controller needs to predict an
# action from the initial state
state = state_space.get_random_state_space(NUM_LAYERS)
print("Initial Random State : ", state_space.parse_state_space_list(state))
#print()
# train for number of trails
for trial in range(MAX_TRIALS):
with policy_sess.as_default():
actions = controller.get_action(
state) # get an action for the previous state
# print the action probabilities
state_space.print_actions(actions)
print("Predicted actions : ",
state_space.parse_state_space_list(actions))
# build a model, train and get reward and accuracy from the network manager
reward, previous_acc = manager.get_rewards(
model_fn_cnn, state_space.parse_state_space_list(actions))
print("Rewards : ", reward, "Accuracy : ", previous_acc)
with policy_sess.as_default():
total_reward += reward
print("Total reward : ", total_reward)
# actions and states are equivalent, save the state and reward
state = actions
controller.store_rollout(state, reward)
# train the controller on the saved state and the discounted rewards
loss = controller.train_step()
print("Trial %d: Controller loss : %0.6f" % (trial + 1, loss))
# write the results of this trial into a file
with open('train_history.csv', mode='a+') as f:
data = [previous_acc, reward]
data.extend(state_space.parse_state_space_list(state))
writer = csv.writer(f)
writer.writerow(data)
print()
print("Total Reward : ", total_reward)
manager = NetworkManager(dataset, epochs=MAX_EPOCHS, batchsize=BATCHSIZE, clip_rewards=CLIP_REWARDS)
# get an initial random state space if controller needs to predict an
# action from the initial state
state = state_space.get_random_state_space(NUM_LAYERS)
print("Initial Random State : ", state_space.parse_state_space_list(state))
print()
# train for number of trails
for trial in range(MAX_TRIALS):
with policy_sess.as_default():
K.set_session(policy_sess)
actions = controller.get_action(state) # get an action for the previous state
# print the action probabilities
state_space.print_actions(actions)
print("Predicted actions : ", state_space.parse_state_space_list(actions))
# build a model, train and get reward and accuracy from the network manager
reward, previous_acc = manager.get_rewards(model_fn, state_space.parse_state_space_list(actions))
print("Rewards : ", reward, "Accuracy : ", previous_acc)
with policy_sess.as_default():
K.set_session(policy_sess)
total_reward += reward
print("Total reward : ", total_reward)
# actions and states are equivalent, save the state and reward
state = actions
controller.store_rollout(state, reward)