Nan values with log_prob()

1

Hello,

I’m having a lot of NaN values with the log_prob function.

See the code:

def forward(self, observation):
        # Convert tuple into tensor
        observation_as_list = []
        observation_as_list.append(observation[0])
        observation_as_list.append(observation[1])
        observation_as_list = np.asarray(observation_as_list)
        observation_as_list = observation_as_list.reshape(1,2)
        observation = observation_as_list
        
        state = torch.from_numpy(observation).float()
        state = F.tanh(self.affine(state))
        
        state_value = self.value_layer(state)
        weight_prelu = torch.tensor([[5.0]])
        action_parameters = (self.action_layer(state)) # A voir quelle activation function mettre
        action_distribution = Normal(action_parameters[0][0], action_parameters[0][1])
        
        action = action_distribution.sample() # Torch.tensor; action
        self.logprobs.append(action_distribution.log_prob(action+ 1e-6)) # HERE ###
        self.state_values.append(state_value)
        #print(action.item())
        return action.item() # Float element

The whole code is here: https://github.com/nbrosson/Actor-critic-MountainCar/blob/master/agent.py

Any ideas how I’m getting this problem?

Thanks !!

2

Hello, I am also facing the same problem ! have you found the solution to it ?