def evaluate(dataset_name, model_name, load_path, calculate_reward=True):
seed = 20190827
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if dataset_name == 'MultiWOZ':
dst_sys = RuleDST()
if model_name == "PPO":
from convlab2.policy.ppo import PPO
if load_path:
policy_sys = PPO(False)
policy_sys.load(load_path)
else:
policy_sys = PPO.from_pretrained()
elif model_name == "PG":
from convlab2.policy.pg import PG
if load_path:
policy_sys = PG(False)
policy_sys.load(load_path)
else:
policy_sys = PG.from_pretrained()
elif model_name == "MLE":
from convlab2.policy.mle.multiwoz import MLE
if load_path:
policy_sys = MLE()
policy_sys.load(load_path)
else:
policy_sys = MLE.from_pretrained()
elif model_name == "GDPL":
from convlab2.policy.gdpl import GDPL
if load_path:
policy_sys = GDPL(False)
policy_sys.load(load_path)
else:
policy_sys = GDPL.from_pretrained()
elif model_name == "GAIL":
from convlab2.policy.gail import GAIL
if load_path:
policy_sys = GAIL(False)
policy_sys.load(load_path)
else:
policy_sys = GAIL.from_pretrained()
dst_usr = None
policy_usr = RulePolicy(character='usr')
simulator = PipelineAgent(None, None, policy_usr, None, 'user')
env = Environment(None, simulator, None, dst_sys)
agent_sys = PipelineAgent(None, dst_sys, policy_sys, None, 'sys')
evaluator = MultiWozEvaluator()
sess = BiSession(agent_sys, simulator, None, evaluator)
task_success = {'All': []}
for seed in range(100):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
sess.init_session()
sys_response = []
logging.info('-'*50)
logging.info(f'seed {seed}')
for i in range(40):
sys_response, user_response, session_over, reward = sess.next_turn(sys_response)
if session_over is True:
task_succ = sess.evaluator.task_success()
logging.info(f'task success: {task_succ}')
logging.info(f'book rate: {sess.evaluator.book_rate()}')
logging.info(f'inform precision/recall/f1: {sess.evaluator.inform_F1()}')
logging.info('-'*50)
break
else:
task_succ = 0
for key in sess.evaluator.goal:
if key not in task_success:
task_success[key] = []
else:
task_success[key].append(task_succ)
task_success['All'].append(task_succ)
for key in task_success:
logging.info(f'{key} {len(task_success[key])} {np.average(task_success[key]) if len(task_success[key]) > 0 else 0}')
if calculate_reward:
reward_tot = []
for seed in range(100):
s = env.reset()
reward = []
value = []
mask = []
for t in range(40):
s_vec = torch.Tensor(policy_sys.vector.state_vectorize(s))
a = policy_sys.predict(s)
# interact with env
next_s, r, done = env.step(a)
logging.info(r)
reward.append(r)
if done: # one due to counting from 0, the one for the last turn
break
logging.info(f'{seed} reward: {np.mean(reward)}')
reward_tot.append(np.mean(reward))
logging.info(f'total avg reward: {np.mean(reward_tot)}')
else:
raise Exception("currently supported dataset: MultiWOZ")
def evaluate(dataset_name, model_name, load_path):
seed = 20200722
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if dataset_name == 'MultiWOZ':
dst_sys = RuleDST()
if model_name == "PPO":
from convlab2.policy.ppo import PPO
if load_path:
policy_sys = PPO(False)
policy_sys.load(load_path)
else:
policy_sys = PPO.from_pretrained()
elif model_name == "DQN":
from convlab2.policy.dqn.DQN.DQN import DQN
if load_path:
policy_sys = DQN(False)
policy_sys.load(load_path)
else:
print('Please add load path.')
elif model_name == "DQfD_RE":
from convlab2.policy.dqn.RE.DQfD import DQfD
if load_path:
policy_sys = DQfD(False)
policy_sys.load(load_path)
else:
print('Please add load path.')
elif model_name == "DQfD_NLE":
from convlab2.policy.dqn.NLE.DQfD import DQfD
if load_path:
policy_sys = DQfD(False)
policy_sys.load(load_path)
else:
print('Please add load path.')
elif model_name == "MLE":
from convlab2.policy.mle.multiwoz import MLE
if load_path:
policy_sys = MLE()
policy_sys.load(load_path)
else:
policy_sys = MLE.from_pretrained()
policy_usr = RulePolicy(character='usr')
simulator = PipelineAgent(None, None, policy_usr, None, 'user')
agent_sys = PipelineAgent(None, dst_sys, policy_sys, None, 'sys')
evaluator = MultiWozEvaluator()
sess = BiSession(agent_sys, simulator, None, evaluator)
task_success = 0
evaluator_success = 0
for seed in range(100):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
sess.init_session()
sys_response = []
cur_success = 0
for i in range(40):
sys_response, user_response, session_over, reward = sess.next_turn(
sys_response)
if reward == 80:
cur_success = 1
task_success += 1
if session_over is True:
break
# logging.debug('Current task success: {}, the evaluator result: {}.'.format(cur_success, sess.evaluator.task_success()))
evaluator_success += sess.evaluator.task_success()
logging.debug('Task success rate: {} and evaluator result: {}.'.format(
task_success / 100, evaluator_success / 100))
return task_success / 100, evaluator_success / 100