ValueError: expected sequence of length 361 at dim 1 (got 208)[BUG]

[nomawni]

Hi, I am facing this following error, while optimizing the PPO policy: ValueError: expected sequence of length 361 at dim 1 (got 208) I first simulated a dialogue between the user and system. Which then I gathered and tried to train the policy using the code provided in the following link: https://github.com/ConvLab/ConvLab-3/blob/master/tutorials/Train_RL_Policies/example_train.py. Here is how to reproduce the error: user_nlu = BERTNLU()

not use dst

user_dst = RuleDST()

rule policy

user_policy = RulePolicy(character='usr')

template NLG

user_nlg = TemplateNLG(is_user=True)

assemble

user_agent = PipelineAgent(user_nlu, user_dst, user_policy, user_nlg, name='user')

sys_nlu = BERTNLU() sys_nlg = TemplateNLG(is_user=False) sys_dst = RuleDST() #sys_policy = RulePolicy(character="sys") sys_policy = PPO(is_train=True, seed=0, dataset_name='multiwoz21') sys_agent = PipelineAgent(sys_nlu, sys_dst, sys_policy, sys_nlg, name="sys")

evaluator = MultiWozEvaluator() session = BiSession(sys_agent, user_agent, kb_query=None, evaluator=evaluator) simulator = PipelineAgent(user_nlu, user_dst, user_policy, user_nlg, name='user') env = Environment(sys_nlg, simulator, sys_nlu, sys_dst, evaluator=evaluator)

def set_seed(r_seed): random.seed(r_seed) np.random.seed(r_seed) torch.manual_seed(r_seed)

set_seed(20200131)

num_dialogues = 20 dialog_data = []

def determine_success(final_reward, success_threshold=0): return 1 if final_reward >= success_threshold else 0

sys_response = '' success_threshold = 0 success_labels = []

for i in range(num_dialogues): session.init_session() dialog = [] session_over = False cumulative_reward = 0 print(f"Turn round {i}") print() while not session_over: response = session.next_turn(sys_response) system_response, user_response, session_over, reward = response dialog.append((system_response, user_response)) cumulative_reward += reward print(f"User: {user_response}") print(f"System: {system_response}") print(f" Reward: {reward}") print() success_label = determine_success(cumulative_reward, success_threshold) dialog_data.append((dialog, success_label))

Get the success of the evaluator

success = session.evaluator.task_success() success_labels.append(success)

print(f"Task success: {session.evaluator.task_success()}") print(f"Book rate: {session.evaluator.book_rate()}") print('inform precision/recall/f1:', session.evaluator.inform_F1()) print('-'*50) print('final goal:') pprint(session.evaluator.goal) print('='*100)

And then use the code in the following link to try to train the policy and reproduce the error : https://github.com/ConvLab/ConvLab-3/blob/master/tutorials/Train_RL_Policies/example_train.py (for i in range(epoch): print(f'We are currently in epoch {i} in the training of PPO process') update(env, sys_policy, batchsz, i, process_num))

Here is the full error that I am facing:

ValueError Traceback (most recent call last) Cell In[50], line 3 1 if name == "main": ----> 3 main() 5 policy.save("trained_dialogue_policy")

Cell In[49], line 49 47 for i in range(epoch): 48 print(f'We are currently in epoch {i} in the training of PPO process') ---> 49 update(env, sys_policy, batchsz, i, process_num)

Cell In[10], line 189 186 def update(env, policy, batchsz, epoch, process_num): 187 # sample data asynchronously 188 print(f'In the update method, before the sampling') --> 189 batch = sample(env, policy, batchsz, process_num) 191 print(f'In the update method after the sampling method') 193 # data in batch is : batch.state: ([1, s_dim], [1, s_dim]...) 194 # batch.action: ([1, a_dim], [1, a_dim]...) 195 # batch.reward/ batch.mask: ([1], [1]...)

Cell In[10], line 174 171 evt = Event() # This can be a simple threading event. 173 # Call the sampler function directly instead of starting new processes. ... ---> 68 s_vec = torch.Tensor(policy.vector.state_vectorize(s)) 69 a = policy.predict(s) 71 # interact with env

ValueError: expected sequence of length 361 at dim 1 (got 208)

[ChrisGeishauser]

Hi @nomawni,

this tutorial has not been maintained since years. Nevertheless, your bug should be fixed easily:

https://github.com/ConvLab/ConvLab-3/blob/master/tutorials/Train_RL_Policies/example_train.py#L60

Here in line 60, you currently have s_vec = torch.Tensor(policy.vector.state_vectorize(s))

You should have instead: s_vec, action_mask = self.vector.state_vectorize(state)

You should follow the same pattern in line 70.

I would recommend you to rather look into the training script of PPO since this is maintained and working and it is basically the same code as in the tutorial :) see here:

https://github.com/ConvLab/ConvLab-3/blob/master/convlab/policy/ppo/train.py

[nomawni]

@ChrisGeishauser Thanks. I am going to check the new link.

[nomawni]

Hi @ChrisGeishauser, After using the new training script that you suggested. I am facing the following error:

ValueError                                Traceback (most recent call last)
Cell In[24], line 13
     11 print()
     12 while not session_over:
---> 13   response = session.next_turn(sys_response)
     14   system_response, user_response, session_over, reward = response
     15   dialog.append((system_response, user_response))

File c:\users\convlab-3\convlab\dialog_agent\session.py:122, in BiSession.next_turn(self, last_observation)
    100 def next_turn(self, last_observation):
    101     """Conduct a new turn of dialog, which consists of the system response and user response.
    102 
    103     The variable type of responses can be either 1) str or 2) dialog act, depends on the dialog mode settings of the
   (...)
    120             The reward given by the user.
    121     """
--> 122     user_response = self.next_response(last_observation)
    123     if self.evaluator:
    124         self.evaluator.add_sys_da(self.user_agent.get_in_da_eval(), self.sys_agent.dst.state['belief_state'])

File c:\users\convlab-3\convlab\dialog_agent\session.py:96, in BiSession.next_response(self, observation)
     94 def next_response(self, observation):
     95     next_agent = self.next_agent()
---> 96     response = next_agent.response(observation)
...
     29     new_acts['categorical'].append(
     30         {"intent": intent, "domain": domain, "slot": slot, "value": value})
     31 return new_acts

ValueError: too many values to unpack (expected 4)

To replicate the error, here is the code that I am using: 

user_nlu = BERTNLU()
# not use dst
user_dst = RuleDST()
# rule policy
user_policy = RulePolicy(character='usr')
# template NLG
user_nlg = TemplateNLG(is_user=True)
# assemble
user_agent = PipelineAgent(user_nlu, user_dst, user_policy, user_nlg, name='user')

sys_nlu = BERTNLU()
sys_nlg = TemplateNLG(is_user=False)
sys_dst = RuleDST()
policy_sys = PPO(is_train=True, seed=conf['model']['seed'],
                    vectorizer=conf['vectorizer_sys_activated'])
sys_policy = PPO(is_train=True, seed=0, dataset_name='multiwoz21')
sys_agent = PipelineAgent(sys_nlu, sys_dst, policy_sys, sys_nlg, name="sys")

evaluator = MultiWozEvaluator()
session = BiSession(sys_agent, user_agent, kb_query=None, evaluator=evaluator)
simulator = PipelineAgent(user_nlu, user_dst, user_policy, user_nlg, name='user')
env = Environment(sys_nlg, simulator, sys_nlu, sys_dst, evaluator=evaluator)

def set_seed(r_seed):
  random.seed(r_seed)
  np.random.seed(r_seed)
  torch.manual_seed(r_seed)

set_seed(20200131)

num_dialogues = 30
dialog_data = []

sys_response = ''
success_threshold = 0
success_labels = []

for i in range(num_dialogues):
  session.init_session()
  dialog = []
  session_over = False
  cumulative_reward = 0
  print(f"Turn round {i}")
  print()
  while not session_over:
    response = session.next_turn(sys_response)
    system_response, user_response, session_over, reward = response
    dialog.append((system_response, user_response))
    cumulative_reward += reward
    print(f"User: {user_response}")
    print(f"System: {system_response}")
    print(f" Reward: {reward}")
    print()
  success_label = determine_success(cumulative_reward, success_threshold)
  dialog_data.append((dialog, success_label))
  # Get the success of the evaluator
  success = session.evaluator.task_success()
  success_labels.append(success)

[nomawni]

@ChrisGeishauser could you help me with the error that I mentioned previously?