def __init__(self, movie_dict=None, act_set=None, slot_set=None, params=None):
## parameters associated with dialogue action and slot filling
self.movie_dict = movie_dict
self.act_set = act_set
self.slot_set = slot_set
self.act_cardinality = len(act_set.keys())
self.slot_cardinality = len(slot_set.keys())
self.feasible_actions = dialog_config.feasible_actions
self.num_actions = len(self.feasible_actions)
# rl specific parameters
# epsilon:
self.params = params
self.epsilon = params['epsilon']
#
self.agent_run_mode = params['agent_run_mode']
#
self.agent_act_level = params['agent_act_level']
# experience replay
# self.experience_replay_pool_size = params.get('experience_replay_pool_size', 1000)
# self.experience_replay_pool = [] #Replay_Memory(self.experience_replay_pool_size)
self.hidden_size = params.get('dqn_hidden_size', 60)
# gamma : discount factor
self.gamma = params.get('gamma', 0.99)
self.predict_mode = params.get('predict_mode', False)
self.actor_lr = params.get('actor_lr', 0.0005)
self.critic_lr = params.get('critic_lr', 0.001)
self.gan_critic_lr = params.get('gan_critic_lr', 0.001)
self.discriminator_lr = params.get('discriminator_lr', 0.0005)
self.discriminator_batch_size = params.get('discriminator_batch_size', 1)
self.expert_path = params["expert_path"]
self.reg_cost = self.params.get('reg_cost', 1e-3)
## warm start:
## there is no warm start since there are is no experience replay
# self.warm_start = params.get('warm_start', 0)
self.max_turn = params['max_turn'] + 4
self.state_dimension = 2 * self.act_cardinality + 7 * self.slot_cardinality + 3 + self.max_turn
self.expert_weights = params['expert_weights']
# Build models
self.build_actor_model(self.actor_lr)
self.build_critic_model(self.critic_lr)
self.build_critic_model(self.gan_critic_lr, True)
self.build_discriminator(self.gan_critic_lr)
self.n = params.get('n', 50)
## load a model if present
if params['trained_model_path'] != None:
self.load(params['trained_actor_model_path'], "actor")
self.load(params['trained_critic_model_path'], "critic")
self.load(params['trained_adversarial_critic_model_path'], "advesarial_critic")
self.load(params['trained_discriminator_model_path'], "discriminator")
self.predict_mode = True
self.warm_start = 2
#self.expert = DQN(self.state_dimension, self.hidden_size, self.hidden_size, self.num_actions)
self.expert = self.build_expert_model()
# self.clone_dqn = copy.deepcopy(self.expert)
# self.clone_dqn = keras.models.clone_model(self.expert)
self.cur_bellman_err = 0
# Prediction Mode: load trained DQN model
if params['expert_path'] != None:
# self.dqn.model = model_from_json(params['expert_path'])
# copy.deepcopy(self.load_trained_DQN(params['expert_path']))
# self.dqn.model.load_weights(params['expert_weights'])
self.predict_mode = True
self.warm_start = 2
user_sim = RuleSimulator(params['movie_dictionary'],
params['act_set'],
params['slot_set'],
params['goal_set'],
params['usersim_params'])
self.dialog_manager = DialogManager(self.expert,
user_sim,
params['act_set'],
params['slot_set'],
params['movie_kb'])
user_sim.set_nlg_model(params['nlg'])
user_sim.set_nlu_model(params['nlu'])
nlg_model.load_nlg_model(nlg_model_path)
nlg_model.load_predefine_act_nl_pairs(diaact_nl_pairs)
agent.set_nlg_model(nlg_model)
user_sim.set_nlg_model(nlg_model)
world_model.set_nlg_model(nlg_model)
################################################################################
# load trained NLU model
################################################################################
nlu_model_path = params['nlu_model_path']
nlu_model = nlu()
nlu_model.load_nlu_model(nlu_model_path)
agent.set_nlu_model(nlu_model)
user_sim.set_nlu_model(nlu_model)
world_model.set_nlu_model(nlu_model)
################################################################################
# Dialog Manager
################################################################################
dialog_manager = DialogManager(agent, user_sim, world_model, act_set, slot_set, movie_kb)
################################################################################
# Run num_episodes Conversation Simulations
################################################################################
status = {'successes': 0, 'count': 0, 'cumulative_reward': 0}
simulation_epoch_size = params['simulation_epoch_size']
batch_size = params['batch_size'] # default = 16
warm_start = params['warm_start']