def test():
test_env = UserSimulator(disease_symptom_path,
disease_symptom_mapping_path)
total_rewards = 0
for i_episode in range(25):
state = test_env.reset()
top_10_predictions = 0
top_5_predictions = 0
for t in count():
action = select_action(np.expand_dims(state, axis=0))
next_state, reward, done, _ = test_env.step(action.item())
state = next_state
if done:
total_rewards += reward
with torch.no_grad():
q_values = policy_net(np.expand_dims(state,
axis=0)).squeeze(0)
diagnosis_q_values = q_values[test_env.num_symptom:]
# # print (diagnosis_q_values.shape)
top_10_disease, top_5_disease = test_env.get_top_diseases(
diagnosis_q_values)
if test_env.goal in top_10_disease:
top_10_predictions += 1
if test_env.goal in top_5_disease:
top_5_predictions += 1
break
# print ('Test Rewards : ', total_rewards/25.)
return total_rewards / 25., top_10_predictions / 25., top_5_predictions / 25.
class DialogEnv(gym.Env):
def __init__(
self,
user_goals: List[UserGoal],
emc_params: Dict,
max_round_num: int,
database: Dict,
slot2values: Dict[str, List[Any]],
) -> None:
self.user = UserSimulator(user_goals, max_round_num)
self.emc = ErrorModelController(slot2values, emc_params)
self.state_tracker = StateTracker(database, max_round_num)
self.action_space = gym.spaces.Discrete(len(AGENT_ACTIONS))
self.observation_space = gym.spaces.multi_binary.MultiBinary(
self.state_tracker.get_state_size())
def step(self, agent_action_index: int):
agent_action = map_index_to_action(agent_action_index)
self.state_tracker.update_state_agent(agent_action)
user_action, reward, done, success = self.user.step(agent_action)
if not done:
self.emc.infuse_error(user_action)
self.state_tracker.update_state_user(user_action)
next_state = self.state_tracker.get_state(done)
return next_state, reward, done, success
def reset(self):
self.state_tracker.reset()
init_user_action = self.user.reset()
self.emc.infuse_error(init_user_action)
self.state_tracker.update_state_user(init_user_action)
return self.state_tracker.get_state()
plt.ylabel('Prediction Accuracys')
plt.savefig(default_pred_path)
plt.close()
test_rewards_list = []
test_episode_list = []
test_top_5_pred_list = []
test_top_10_pred_list = []
for i_episode in range(num_episodes):
# Initialize the environment and state
state = env.reset()
for t in count():
# Select and perform an action
action = select_action(np.expand_dims(state, axis=0))
next_state, reward, done, _ = env.step(action.item())
reward = torch.tensor([reward], device=device)
# Store the transition in memory
memory.push(torch.tensor([state], device=device),
torch.tensor([[action]], device=device),
torch.tensor([next_state], device=device), reward)
# Move to the next state
state = next_state
# Perform one step of the optimization (on the target network)
optimize_model()
if done:
episode_durations.append(t + 1)
break
# Update the target network, copying all weights and biases in DQN
if i_episode % TARGET_UPDATE == 0:
target_net.load_state_dict(policy_net.state_dict())